When something on this site turns out to be wrong, it is corrected or taken off, and listed here with what is true. A corrected brief says so at its top and under each corrected item, with a link here.
The audit of October 13, 2026
On October 13, 2026, SteadyVane checked what it had published: every sentence of the 49 items of the daily briefs from September 25 to October 13 (on this site, in the Telegram channel and in the posts on X, Bluesky, Threads and Instagram), every pace number, the site's own wording, the weekly emails and top 10, the daily videos, and the 62 older posts on X this site showed as posted (all but 15 low-risk ones, whose claims match their source's title, and those that state only an opinion). The AI summaries on the older day pages were not checked one by one: they were taken off the site (below). What a YouTube video shows was checked against its title and description; the videos were not watched. Every problem found was checked again by a second, separate check. In the 49 brief items, 91 statements were confirmed wrong, misleading or imprecise - 12 of them plainly wrong. Every one of their pace numbers was computed wrongly, and most were off (below). Every view, star, point, like and comment count was read right; the errors were in the rest: the words around those counts, figures the writer added (a subscriber count, times and days), and the pace.
The brief pages were corrected on October 14, 2026. Each corrected item carries a note that links to its entry below, where the original wording stands next to the corrected one. The briefs' posts on X, Bluesky, Threads, Instagram and the Telegram channel stay as they were posted; this page is their record.
Since October 14, 2026, before a brief publishes, every sentence its writer wrote - on this site and in every post - is checked against the evidence SteadyVane gathered, times and ages are worked out in code rather than written by the model, and the pace is computed as How it works describes.
How to read this page. Wrong: a false fact. Misleading: words that mislead, or a company's or creator's claim stated as fact. Imprecise: a small inaccuracy. "Where" lists every text that carried the problem, each in its own words.
The pace numbers
Every brief item before October 14 said how fast it climbed against the others - for example "about 203 times the typical pace of other trending AI repos SteadyVane tracks on GitHub". Those numbers were computed wrongly. Each was the item's single fastest stretch between two readings (on Hacker News as short as ten minutes), while the sentence put it on the whole climb; the typical pace counted each other item as often as it had been read; and on Hacker News the comparison was only with stories found by search, not every story SteadyVane tracks.
The pace stated now is the item's climb over the hours it was tracked before its brief, against the median climb of the other items read on the same source that climbed in those same hours, each counted once (How it works). The old numbers were taken out of every brief's text, and each item's numbers line shows the honest pace. The old numbers also went out in the posts and their images, the weekly emails, the weekly top 10 of October 13 and the daily videos of October 11 to 13, and the tracker and the home page used the same kind of number until the night of October 13.
Of the 49 items, 31 stated a pace too high by more than a quarter (the worst about 21 times too high); 3 stated one too low by more than a fifth; 13 were close (within a quarter above or a fifth below); and 2 had none that can be stated (the table says why).
Pace: the brief said about 203 times the typical pace; the honest pace is about 32 times the typical pace of other repos SteadyVane tracks on GitHub.
Wrong Said: “up 725 from 195 on Tuesday”
Where: what moved, Instagram
The 195 reading was taken on Thursday, Sep 24; the repository did not exist on Tuesday.
Misleading Said: “hit 725 stars in 24 hours”
Where: the day's title
725 was the gain in those 24 hours, not the total: it went from 195 to 920 stars.
Misleading Said: “no vendor lock-in”
Where: what it is, Bluesky, Threads, Instagram
The project never says that, and the app still runs on specific providers' services. It works on the user's own provider accounts, with no GoLive account, backend or telemetry.
Misleading Said: “This closes the last gap - from working prototype to live product - without switching contexts or learning infrastructure.”
Where: why it matters, Instagram
That is the project's goal stated as done. Its README calls it an early alpha that automates what providers expose and hands the rest to a human.
Imprecise Said: “Most agent-built apps stop at the demo.”
Where: why it matters, Instagram
Nothing supports "most". The README says only that getting an agent-built app to real users still means setting up accounts, hosting, databases and domains.
What it is
Was: “golive-skill is an MIT-licensed command-line tool that connects hosting, databases, domains, email, and payments for agent-built products - all on the user's own accounts, no vendor lock-in. Built by mikehasa; zero dependencies.”
Now: “golive-skill is an MIT-licensed command-line tool that connects hosting, databases, domains, email, and payments for agent-built products - all on the user's own provider accounts, with no GoLive account, backend or telemetry. Built by mikehasa; zero dependencies.”
What moved
Was: “920 stars, up 725 from 195 on Tuesday - about 203 times the typical pace of other trending AI repos SteadyVane tracks on GitHub. 65 forks.”
Now: “920 stars, up 725 from 195 on Thursday, Sep 24. 65 forks.”
Why it matters
Was: “Most agent-built apps stop at the demo. This closes the last gap - from working prototype to live product - without switching contexts or learning infrastructure. If you work with Claude Code or similar tools, deployment is now a skill the agent can run.”
Now: “Getting an agent-built app to real users still means setting up accounts, hosting, databases and domains, as the project's README puts it. The project says it aims to take an agent-built app live from inside the coding agent; it is an early alpha that automates what providers expose and hands the rest to a human. If you work with Claude Code or similar tools, deployment is now a skill the agent can run.”
Pace: the brief said about 134 times the typical pace; the honest pace is about 57 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “The model performs at Fable 5.1 levels on most tasks”
Where: what it is, X, Threads, Instagram
Anthropic's own claim, stated as fact: Anthropic says the model performs at the level of Claude Fable 5.1 on most work.
Misleading Said: “costs 40 percent less to run than Opus 5”
Where: what it is, X, Threads, Instagram
Anthropic's estimate from its own tests on typical workloads, stated as fact. Per-token list prices are 20 percent lower.
Misleading Said: “performs at Fable 5.1 levels and costs 40% less than Opus 5”
Where: Bluesky
Without "on most" and "to run", and with no source named, this reads as full Fable 5.1 parity and a 40% price cut. List prices fell 20%.
Was: “Two Minute Papers published a video analyzing Claude Opus 5.5, Anthropic's latest model - released September 22. The model performs at Fable 5.1 levels on most tasks and costs 40 percent less to run than Opus 5.”
Now: “Two Minute Papers published a video analyzing Claude Opus 5.5, Anthropic's latest model - released September 22. Anthropic says the model performs at the level of Claude Fable 5.1 on most work and costs about 40 percent less to run than Opus 5 on typical workloads; per-token list prices are 20 percent lower.”
What moved
Was: “251,000 views, up 167,000 in 20 hours - about 134 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. 278 comments, 3,843 likes.”
Now: “251,000 views, up 167,000 in 20 hours. 278 comments, 3,843 likes.”
Pace: the brief said about 8 times the typical pace; the honest pace is about 12 times the typical pace of other products SteadyVane tracks on Product Hunt.
No other problem was found in its texts.
What moved
Was: “318 votes on Product Hunt, up 56 in 8 hours - about 8 times the typical pace of items SteadyVane tracks on that platform. 66 comments.”
Now: “318 votes on Product Hunt, up 56 in 8 hours. 66 comments.”
Pace: the brief said about 9.3 times the typical pace; the honest pace is about 4.7 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Wrong Said: “The stack is ChatGPT for scripting, Nano Banana for 3D generation, and ElevenLabs for voiceover.”
Where: why it matters, X, Threads, Instagram, Bluesky
The video uses five tools: ChatGPT, Google Flow, Nano Banana, ElevenLabs and CapCut. Google Flow, left out, is the one that makes the video.
Misleading Said: “using ChatGPT, Nano Banana, and ElevenLabs”
Where: what it is
The video names five tools; Google Flow (the video step) and CapCut were left out.
Misleading Said: “Nano Banana for 3D generation”
Where: why it matters, X, Threads, Instagram
Nano Banana is Google's image model, not a 3D generator. In this workflow it makes the still images; the motion is made in the video step.
Misleading Said: “a workflow that runs on tools you can use today without payment”
Where: why it matters, X, Threads, Instagram
Only the creator's own billing of it as free supports that. Each tool has a free tier, with real usage limits, Google Flow's especially.
Misleading Said: “All three have working free tiers that combine into one workflow.”
Where: Bluesky
The three named tools do not make the workflow on their own; it also needs Google Flow and CapCut. Each has a free tier with usage limits.
Imprecise Said: “ElevenLabs for voiceover”
Where: why it matters, X, Threads, Instagram
No source says which tool does which job; ElevenLabs doing the voiceover is not established.
What it is
Was: “A YouTube tutorial showing how to create 3D motion graphics using ChatGPT, Nano Banana, and ElevenLabs—all free-tier tools—without After Effects or paid software. Published by AI Playbook on Tue Sep 22.”
Now: “A YouTube tutorial showing how to create 3D motion graphics using ChatGPT, Google Flow, Nano Banana, ElevenLabs and CapCut, without After Effects. The creator bills it as made with free AI tools; each tool has a free tier, with usage limits. Published by AI Playbook on Tue Sep 22.”
What moved
Was: “17,360 views over 28 hours between Fri Sep 25, 01:00 Denver and Sat Sep 26, 05:00 Denver, about 9.3 times the typical pace of other items SteadyVane tracks on YouTube. 1,220 likes and 60 comments.”
Now: “17,360 views over 28 hours between Fri Sep 25, 01:00 Denver and Sat Sep 26, 05:00 Denver. 1,220 likes and 60 comments.”
Why it matters
Was: “If you make video content and have avoided motion graphics because of software cost or complexity, this is a workflow that runs on tools you can use today without payment. The stack is ChatGPT for scripting, Nano Banana for 3D generation, and ElevenLabs for voiceover.”
Now: “If you make video content and have avoided motion graphics because of software cost or complexity, this workflow is built on tools that each have a free tier, with usage limits. The video's workflow uses ChatGPT, Google Flow, Nano Banana (Google's image model, for the scene images), ElevenLabs and CapCut, with one master prompt writing the ideas, script, image prompts and video prompts.”
Pace: the brief said about 18.3 times the typical pace; the honest pace is about 12 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
The views were on Alex Finn's YouTube video about the update, not on the update.
Misleading Said: “an open-source AI coding and automation tool”
Where: what it is, Bluesky, Threads, Instagram, the take
Nous Research presents Hermes Agent as a general-purpose AI agent, and its docs say it is not a coding copilot. It has memory, messaging-app access and scheduled automations.
Misleading Said: “Hermes is an open-source AI coding agent by Nous Research.”
Where: X
The same problem in the X post: it is a general-purpose agent, not a coding agent.
Imprecise Said: “Hermes is an active open-source project with major releases documented this month.”
Where: why it matters
Every September release up to the brief was a patch release (five; the latest, v0.21.5, on Sep 24), and the release notes deferred full documentation to a later version.
Was: “A Hermes agent update video pulled 15,400 views in eight hours at 18x typical pace”
Now: “A Hermes agent update video pulled 15,400 views in eight hours”
What it is
Was: “A YouTube video covering the latest update to Hermes Agent, an open-source AI coding and automation tool by Nous Research. Posted by Alex Finn on Fri Sep 25.”
Now: “A YouTube video covering the latest update to Hermes Agent, an open-source, general-purpose AI agent from Nous Research, with memory, messaging-app access and scheduled automations. Posted by Alex Finn on Fri Sep 25.”
What moved
Was: “15,422 views over 8 hours between Fri Sep 25, 21:00 Denver and Sat Sep 26, 05:00 Denver, about 18.3 times the typical pace of other items SteadyVane tracks on YouTube. 1,292 likes and 103 comments.”
Now: “15,422 views over 8 hours between Fri Sep 25, 21:00 Denver and Sat Sep 26, 05:00 Denver. 1,292 likes and 103 comments.”
Why it matters
Was: “Hermes is an active open-source project with major releases documented this month. If you run AI agents in production, an update drawing this much attention in under a day usually signals new capabilities or breaking changes worth reviewing in the changelog.”
Now: “Hermes is an active open-source project, with five patch releases so far in September (the latest, v0.21.5, on Sep 24). If you run AI agents in production, an update drawing this much attention in under a day usually signals new capabilities or breaking changes worth reviewing in the changelog.”
Take
Was: “The velocity on a community coding agent video matters more than the title's breathlessness.”
Now: “The velocity on a video about Nous Research's open-source agent matters more than the title's breathlessness.”
Pace: the brief said about 13.9 times the typical pace; the honest pace is about 7.6 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Wrong Said: “The channel has 1.7 million subscribers”
Where: why it matters, Instagram, X, Bluesky, Threads
About 2.2 million. The 1.7 million came from the fact-check model's text (most likely an old bio line), not from a measurement.
Imprecise Said: “A no-code AI agents video reportedly showing five automation tools”
Where: headline, Instagram
The video shows five no-code AI agents the creator built himself on one platform, not five separate tools.
Imprecise Said: “the five-agent claim is only in the title”
Where: X
The description makes the claim too. The point stands: the only source is the creator's own text.
Was: “A no-code AI agents video reportedly showing five automation tools gained 40,000 views in 40 hours”
Now: “A video in which the creator shows five no-code AI agents he built gained 40,000 views in 40 hours”
What moved
Was: “40,326 views over 40 hours between Thu Sep 24, 13:00 Denver and Sat Sep 26, 05:00 Denver, about 13.9 times the typical pace of other items SteadyVane tracks on YouTube. 526 likes and 29 comments.”
Now: “40,326 views over 40 hours between Thu Sep 24, 13:00 Denver and Sat Sep 26, 05:00 Denver. 526 likes and 29 comments.”
Why it matters
Was: “If the video delivers what the title claims, it is a starting point for offloading repetitive tasks without writing code. The channel has 1.7 million subscribers and a track record of AI and business tutorials, but the specific claim about five agents is only verified by the video's own description.”
Now: “If the video delivers what the title claims, it is a starting point for offloading repetitive tasks without writing code. The channel has about 2.2 million subscribers and a track record of AI and business tutorials, but the specific claim about five agents is only verified by the video's own description.”
Pace: the brief said about 25 times the typical pace; the honest pace is about 14 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “If your team already has a production process, this slots into it instead of asking you to adopt a new workflow.”
Where: Instagram
No source says so. Flova runs its own agent-driven workflow; it can export a project to Premiere Pro, Final Cut Pro or CapCut for finishing.
Misleading Said: “An AI video production agent went from 151,452 to 197,239 views in 20 hours”
Where: headline, X, Bluesky, Threads, Instagram, the day's title
The views were on a YouTube video reviewing the tool (Shark Numbers' "AI Video Is Having Its Claude Code Moment - Flova AI Review"), not on the tool.
Imprecise Said: “Most AI video tools generate clips.”
Where: why it matters, Instagram, X, Threads
A sweeping claim with no source behind it; it is the video's own framing.
Was: “An AI video production agent went from 151,452 to 197,239 views in 20 hours”
Now: “A YouTube review of Flova AI went from 151,452 to 197,239 views in 20 hours”
What moved
Was: “Up 45,787 views in 20 hours - about 25 times the typical pace of other items SteadyVane tracks on YouTube. 28,909 likes and 448 comments about three days after it was published.”
Now: “Up 45,787 views in 20 hours. 28,909 likes and 448 comments about three days after it was published.”
Why it matters
Was: “Most AI video tools generate clips. This one coordinates the entire job - script, shots, edits, revisions - the way a human producer would, except it remembers the brief across all the steps.”
Now: “Instead of generating one clip per prompt, as the video puts it, this one coordinates the entire job - script, shots, edits, revisions - the way a human producer would, except it remembers the brief across all the steps.”
Pace: the brief said about 11 times the typical pace; the honest pace is about 4.5 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “A voice-to-AI dictation tool went from 11,593 to 22,094 views in 16 hours”
Where: headline, X, Bluesky, Threads, Instagram
The views were on Talha Reviews' YouTube video about Wispr Flow, not on the app.
Was: “A voice-to-AI dictation tool went from 11,593 to 22,094 views in 16 hours”
Now: “A YouTube video about Wispr Flow, a voice dictation app, went from 11,593 to 22,094 views in 16 hours”
What moved
Was: “Up 10,501 views in 16 hours - about 11 times the typical pace of other items SteadyVane tracks on YouTube. 376 likes and 35 comments about 24 hours after it was published.”
Now: “Up 10,501 views in 16 hours. 376 likes and 35 comments about 24 hours after it was published.”
Pace: the brief said about 42 times the typical pace; the honest pace is about 11 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “A multi-model AI workspace went from 42,192 to 86,443 views in 28 hours”
Where: headline, X, Bluesky, Threads, Instagram
The views were on BMF MEDIA's YouTube video testing Mammouth, not on the product.
Headline
Was: “A multi-model AI workspace went from 42,192 to 86,443 views in 28 hours”
Now: “A YouTube video testing Mammouth, a multi-model AI workspace, went from 42,192 to 86,443 views in 28 hours”
What moved
Was: “Up 44,251 views in 28 hours - about 42 times the typical pace of other items SteadyVane tracks on YouTube. 2,730 likes and 276 comments about four days after it was published.”
Now: “Up 44,251 views in 28 hours. 2,730 likes and 276 comments about four days after it was published.”
Pace: the brief said about 138.6 times the typical pace; the honest pace is about 69 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “Google's stalled roadmap is a signal”
Where: why it matters, Instagram, X, Bluesky, Threads
Only the Pro tier stalled (no new Pro model since Gemini 3.1 Pro in February); Google kept shipping Flash models.
Misleading Said: “Seven months without a Pro update is a product decision, not a technical delay.”
Where: the take, Instagram
No source supports that. Bloomberg reported Gemini 3.5 Pro fell short of internal performance targets; Google said only that it would ship when ready.
What moved
Was: “Up 159,482 views in 12 hours - about 138.6 times the typical pace of other items SteadyVane tracks on YouTube. 3,233 likes and 474 comments.”
Now: “Up 159,482 views in 12 hours. 3,233 likes and 474 comments.”
Why it matters
Was: “If you are choosing a model provider for work that will scale over the next six months, Google's stalled roadmap is a signal. The competitive gap is real and documented by multiple sources.”
Now: “If you are choosing a model provider for work that will scale over the next six months, the stall in Google's Pro tier is a signal: no new Pro model since Gemini 3.1 Pro in February, though Google kept shipping Flash models. The competitive gap is real and documented by multiple sources.”
Take
Was: “Seven months without a Pro update is a product decision, not a technical delay.”
Now: “Seven months without a Pro update: Bloomberg reported Gemini 3.5 Pro fell short of internal performance targets; Google says only it will ship 'as soon as it's ready'.”
Pace: the brief said about 8.8 times the typical pace; the honest pace is about 4.3 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “The video does not clearly specify which tool”
Where: what it is, Instagram
Nobody watched the video. SteadyVane read only the first 500 characters of its description, which name no tool.
Misleading Said: “most free tiers limit video length to seconds or minutes per week with watermarks and resolution constraints”
Where: what it is, Instagram
No source behind it. The free plans that could be checked carry watermark and resolution limits, but mostly do not cap use by the week.
Misleading Said: “most free tiers cap output at a few minutes per week”
Where: why it matters, X, Bluesky, Threads, Instagram
Of the free plans checked, only one (invideo's) has a weekly cap; the others give one-time or monthly credits, or none.
Misleading Said: “calling them suitable for long videos is misleading”
Where: why it matters, Instagram
A negative verdict on a named creator's video with no evidence behind it: SteadyVane did not check which tool the video uses or that tool's free limits.
What it is
Was: “A YouTube tutorial demonstrating free text-to-video AI tools. The video does not clearly specify which tool, and most free tiers limit video length to seconds or minutes per week with watermarks and resolution constraints.”
Now: “A YouTube tutorial demonstrating free text-to-video AI tools. The part of the description SteadyVane read does not name the tool.”
What moved
Was: “Up 17,473 views in 20 hours - about 8.8 times the typical pace of other items SteadyVane tracks on YouTube. 1,067 likes and 255 comments.”
Now: “Up 17,473 views in 20 hours. 1,067 likes and 255 comments.”
Why it matters
Was: “Free text-to-video tools exist and can generate usable clips, but calling them suitable for long videos is misleading - most free tiers cap output at a few minutes per week. Useful for testing workflows, not production at scale.”
Now: “Free text-to-video tools exist and can generate usable clips. The video promises long AI videos for free; SteadyVane did not check which tool it uses or that tool's free limits.”
Pace: the brief said about 432 times the typical pace; the honest pace is about 23 times the typical pace of other stories SteadyVane tracks on Hacker News.
Wrong Said: “The fastest Hacker News climb SteadyVane has recorded for a model launch this year.”
Where: the take, Threads, Instagram
False by SteadyVane's own readings. Claude Opus 5.5, six days earlier, went from 508 to 1,560 points in its first 16 tracked hours (+1,052, against +689 here), and GPT-6 Sol and Luna and Claude Fable and Mythos 5.1 also climbed more.
Imprecise Said: “Claude Sonnet 5.5 reached 796 points on Hacker News in 16 hours”
Where: headline, X, Bluesky, Instagram, the day's title
796 came about 16.5 hours after the post went up (at 16 hours it had about 785). 16 hours was how long SteadyVane tracked it, from 107 points.
Imprecise Said: “Anthropic's new major model version in the Claude 5.5 family”
Where: what it is
Anthropic calls it the second model in the Claude 5.5 family, a clear upgrade over Claude Sonnet 5; it is not a major version.
Headline
Was: “Claude Sonnet 5.5 reached 796 points on Hacker News in 16 hours”
Now: “Claude Sonnet 5.5 reached 796 points on Hacker News about 16.5 hours after it was posted”
What it is
Was: “Anthropic's new major model version in the Claude 5.5 family, available immediately through their platform and major cloud providers.”
Now: “Anthropic's new model, the second in the Claude 5.5 family: a faster, lower-cost complement to Claude Opus 5.5, available now on Anthropic's platform and major cloud providers.”
What moved
Was: “Up from 107 to 796 points in 16 hours - about 432 times the typical pace of items SteadyVane tracks on Hacker News. 533 comments.”
Now: “Up from 107 to 796 points in 16 hours. 533 comments.”
Take
Was: “The fastest Hacker News climb SteadyVane has recorded for a model launch this year.”
Now: “Slower than Claude Opus 5.5's climb six days earlier: 508 to 1,560 points in its first 16 tracked hours.”
Pace: the brief said about 7.8 times the typical pace; the honest pace is about 6.6 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “polished motion graphics”
Where: what it is, Bluesky, Instagram
"Polished" is the creator's own description; nothing independent judged the output.
Misleading Said: “not random AI slop”
Where: why it matters, Threads, the take, Bluesky
The creator's pitch stated as SteadyVane's finding; nobody tested the output. The creator says the approach avoids "AI slop".
Imprecise Said: “You run the code.”
Where: the take, X, Threads, Instagram
You don't: you run one command and approve a plan, and the skill renders each clip (a small HTML file Claude writes) to video on your machine. The true part is that the model writes code, not video.
Was: “A YouTube video showing a custom workflow that uses Claude Opus 5.5 to generate code for polished motion graphics and b-roll for existing videos.”
Now: “A YouTube video showing a custom workflow that uses Claude Opus 5.5 to write code for motion graphics and b-roll for existing videos, which the creator describes as polished.”
What moved
Was: “Up from 84,259 to 99,145 views in 16 hours - about 7.8 times the typical pace of items SteadyVane tracks on YouTube. 1,583 likes and 112 comments. Published about four days ago.”
Now: “Up from 84,259 to 99,145 views in 16 hours. 1,583 likes and 112 comments. Published about four days ago.”
Why it matters
Was: “If your team edits video, this is a workflow you can try now - the model writes code that produces graphics timed to your words, not random AI slop.”
Now: “If your team edits video, this is a workflow you can try now - the model writes code that produces graphics timed to your words; the creator says the approach avoids 'AI slop'.”
Take
Was: “The model writes code. You run the code. That constraint is what makes it actually work.”
Now: “The model writes code, not video: each clip is a small HTML file that the skill renders to video on your machine.”
Pace: the brief said about 36 times the typical pace; the honest pace is about 6.3 times the typical pace of other stories SteadyVane tracks on Hacker News.
Misleading Said: “you want zero latency or cost”
Where: why it matters, X, Threads
Zero latency is false, and the project does not claim it. It claims under 10 ms to the first token and no network round-trip; each model has to download first, and the speed depends on your device.
What moved
Was: “Up from 92 to 246 points on Hacker News in about 12 hours - about 36 times the typical pace of items SteadyVane tracks there. 86 comments.”
Now: “Up from 92 to 246 points on Hacker News in about 12 hours. 86 comments.”
Why it matters
Was: “If your use case is lightweight and you want zero latency or cost, these models run locally in the browser. The tradeoff: capability for control and speed.”
Now: “If your use case is lightweight and you want no network round-trip and no API cost, these models run locally in the browser once downloaded; how fast they answer depends on your device (the site claims under 10 ms to the first token). The tradeoff: capability for control and speed.”
Pace: the brief said about 252 times the typical pace; the honest pace is about 29 times the typical pace of other stories SteadyVane tracks on Hacker News.
No other problem was found in its texts.
What moved
Was: “Up 594 points in 16 hours - about 252 times the typical pace of items SteadyVane tracks on Hacker News. 839 comments.”
Pace: the brief said about 20 times the typical pace; the honest pace is about 18 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “OpenAI's Dots agent climbed to 115,000 views in 20 hours”
Where: the day's title
The views were on a Skill Leap AI video about the agent, which went from 68,478 to 115,123 views in those 20 hours (it was already about 16 hours old when they began).
Was: “Up from 68,478 to 115,123 views in 20 hours - about 20 times the typical pace of items SteadyVane tracks on YouTube. 936 likes and 75 comments about 36 hours after it was published.”
Now: “Up from 68,478 to 115,123 views in 20 hours. 936 likes and 75 comments about 36 hours after it was published.”
Pace: the brief said about 13 times the typical pace; the honest pace is about 12 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “The announcement made it to evening news”
Where: the take, Threads, Instagram
Nothing shows the segment aired on an evening newscast; the YouTube upload went up before CBS Evening News' usual airtime. It was a CBS News segment.
Imprecise Said: “on their 24/7 streaming service”
Where: what it is
Nothing names CBS News 24/7 as this segment's program. It was published on CBS News' YouTube channel; Lauren Fichten reports.
What it is
Was: “A news segment on OpenAI's Dots, the new always-on AI agent announced at DevDay 2026. Published by CBS News on their 24/7 streaming service.”
Now: “A news segment on OpenAI's Dots, the new always-on AI agent announced at DevDay 2026. Published on CBS News' YouTube channel; CBS News' Lauren Fichten reports.”
What moved
Was: “Up from 43,122 to 60,444 views in 12 hours - about 13 times the typical pace of items SteadyVane tracks on YouTube. 273 likes and 92 comments about 38 hours after it was published.”
Now: “Up from 43,122 to 60,444 views in 12 hours. 273 likes and 92 comments about 38 hours after it was published.”
Take
Was: “The announcement made it to evening news because it changes what non-engineers can delegate.”
Pace: the brief said about 11 times the typical pace; the honest pace is about 11 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “making long AI videos for free”
Where: headline, what it is, X, Bluesky, Threads, Instagram, why it matters, the take
"Free" rests only on the creator's own title and description. The video uses one text-to-video generator (linked in its description), not several free tools, and nothing supports "zero marginal cost".
Misleading Said: “a channel focused on YouTube and online earning tutorials”
Where: what it is
That came from the fact-check model's text, not from evidence. The channel is Syed Yasir Abbas; the video's description points to AI video and AI automation services.
Imprecise Said: “A step-by-step guide”
Where: what it is
Nothing shows the video's structure; "step-by-step" came from model-written notes, not the source. It is a tutorial.
Headline
Was: “A tutorial on making long AI videos for free climbed from 29,000 to 68,000 views in 28 hours”
Now: “A tutorial billed as making long AI videos for free climbed from 29,000 to 68,000 views in 28 hours”
What it is
Was: “A step-by-step guide on using free text-to-video AI tools to create long-form videos from a single prompt. Published by Syed Yasir Abbas, a channel focused on YouTube and online earning tutorials.”
Now: “A tutorial that says it shows how to make long AI videos for free from a single prompt with one text-to-video generator, linked in its description. Published by the YouTube channel Syed Yasir Abbas.”
What moved
Was: “Up from 28,582 to 68,337 views in 28 hours - about 11 times the typical pace of items SteadyVane tracks on YouTube. 1,217 likes and 159 comments about 47 hours after it was published.”
Now: “Up from 28,582 to 68,337 views in 28 hours. 1,217 likes and 159 comments about 47 hours after it was published.”
Why it matters
Was: “Free tools that generate long-form video content change the cost structure for anyone testing content ideas or building a YouTube channel without hiring editors or paying for stock footage.”
Now: “Tools that generate long-form video content change the cost structure for anyone testing content ideas or building a YouTube channel without hiring editors or paying for stock footage.”
Take
Was: “The constraint that lifts is not quality - it is iteration speed at zero marginal cost.”
Now: “The constraint that lifts is not quality - it is iteration speed.”
Pace: the brief said about 16 times the typical pace; the honest pace is about 8.1 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Imprecise Said: “unlimited chat”
Where: what it is
OpenAI says chatting with your dot does not count toward ChatGPT usage limits; tasks a dot starts still count.
Imprecise Said: “Twenty announcements in one event”
Where: the take, X, Instagram
More than 20: OpenAI's DevDay 2026 recap lists 25.
What it is
Was: “An analysis of OpenAI's DevDay announcements, comparing Dots with Meta's Muse and Grok Bot. The video examines what Dots adds - ChatGPT integration, unlimited chat, single-user access - and what it lacks, including pricing details.”
Now: “An analysis of OpenAI's DevDay announcements, comparing Dots with Meta's Muse and Grok Bot. The video examines what Dots adds - ChatGPT integration, chats with your dot that don't count toward ChatGPT usage limits, single-user access - and what it lacks, including pricing details.”
What moved
Was: “Up 19,995 views in 20 hours - about 16 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. 571 likes and 42 comments about 46 hours after it was published.”
Now: “Up 19,995 views in 20 hours. 571 likes and 42 comments about 46 hours after it was published.”
Take
Was: “Twenty announcements in one event is a signal about bundling strategy, not product focus.”
Now: “More than 20 announcements in one event (OpenAI's DevDay 2026 recap lists 25) is a signal about bundling strategy, not product focus.”
Pace: the brief said about 15 times the typical pace; the honest pace is about 9.1 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Imprecise Said: “for project management work”
Where: what it is, X, Threads, Bluesky
Not shown by anything SteadyVane read: the description links a shareable "Project Manager" skill, which makes it plausible but does not show what the video demonstrates.
Imprecise Said: “for tailoring Dots to specific jobs”
Where: what it is
From model-written notes, not the source, which links a single Project Manager skill.
Headline
Was: “A ChatGPT Dots workflow tutorial hit 75,326 views at 15x typical pace”
Now: “A ChatGPT Dots workflow tutorial hit 75,326 views”
What it is
Was: “A tutorial on using OpenAI's Dots agents for project management work, including a shareable skill template for tailoring Dots to specific jobs. Published by Alex Finn.”
Now: “A ChatGPT Dots workflow tutorial; its description links a shareable 'Project Manager' skill. Published by Alex Finn.”
What moved
Was: “Up 22,207 views in 20 hours - about 15 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. 75,326 views about 39 hours after publishing. 1,139 likes and 158 comments.”
Now: “Up 22,207 views in 20 hours. 75,326 views about 39 hours after publishing. 1,139 likes and 158 comments.”
Pace: the brief said about 50 times the typical pace; the honest pace is about 41 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “booking flights”
Where: why it matters, X, Threads
OpenAI describes the demo as planning a trip; neither it nor OpenAI's Dots page mentions flights or booking. The demo shows Dots planning a trip, making sense of user feedback and catching up on Slack.
Imprecise Said: “analyzing feedback in a spreadsheet”
Where: why it matters, Threads
The source says making sense of user feedback; no spreadsheet is mentioned.
Headline
Was: “OpenAI's Dots agent demo hit 368,524 views in 30 hours at 50x typical pace”
Now: “OpenAI's Dots agent demo hit 368,524 views in 30 hours”
What moved
Was: “Up 66,058 views in 12 hours - about 50 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. 368,524 views about 30 hours after publishing. 7,092 likes and 917 comments.”
Now: “Up 66,058 views in 12 hours. 368,524 views about 30 hours after publishing. 7,092 likes and 917 comments.”
Why it matters
Was: “If agents work, they slot into your existing tools instead of asking you to learn a new interface. The demo shows Dots reading Slack, analyzing feedback in a spreadsheet, and booking flights - tasks you already do.”
Now: “If agents work, they slot into your existing tools instead of asking you to learn a new interface. The demo shows Dots planning a trip, making sense of user feedback and catching up on Slack - tasks you already do.”
Pace: the brief said about 9 times the typical pace; the honest pace is about 7.5 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Imprecise Said: “A roundup of 10 free AI tools”
Where: headline, X, Bluesky, Threads, Instagram, what it is
The video calls the ten sites free AI and utility tools; at least one, the Cobalt media downloader, has no AI.
Headline
Was: “A roundup of 10 free AI tools hit 150,581 views at 9x typical pace”
Now: “A roundup of 10 free AI and utility websites hit 150,581 views”
What it is
Was: “A curated list of free AI tools including GPT-5 access, Google Vids video generation, and Photopea image editing. Published by Cool Dude Explains.”
Now: “A roundup of 10 free websites with AI and utility tools, from Google Vids and Perplexity to the Cobalt media downloader. Published by Cool Dude Explains.”
What moved
Was: “Up 59,343 views over three days - about 9 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. 150,581 views about seven days after publishing. 3,741 likes and 108 comments.”
Now: “Up 59,343 views over three days. 150,581 views about seven days after publishing. 3,741 likes and 108 comments.”
Pace: the brief said about 60 times the typical pace; the honest pace is about 8.7 times the typical pace of other stories SteadyVane tracks on Hacker News.
Wrong Said: “published about 16 hours ago”
Where: what it is, Threads
The post by Addy Osmani was published Sept. 22, 2026, about 12 days earlier. Only its Hacker News submission was that new.
What it is
Was: “An official Anthropic blog post by Addy Osmani, published about 16 hours ago. It walks through concrete ways to get more out of Claude Opus 5.5 and Claude Code in daily work.”
Now: “An official Anthropic blog post by Addy Osmani, published Sept. 22, 2026, about 12 days before this brief; it was posted to Hacker News about 18 hours before the brief was written. It walks through concrete ways to get more out of Claude Opus 5.5 and Claude Code in daily work.”
What moved
Was: “Up from 1 to 217 points in 16 hours - about 60 times the typical pace of items SteadyVane tracks on Hacker News. 149 comments.”
Now: “Up from 1 to 217 points in 16 hours. 149 comments.”
Pace: the brief said about 28 times the typical pace; the honest pace is about 22 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Imprecise Said: “Published about 48 hours ago”
Where: what it is, Threads
The video was about 50 hours old when the brief was written; 48 hours was its age at SteadyVane's last reading.
What it is
Was: “A YouTube video by Chase AI explaining Claude Mods, a feature that lets developers write JavaScript functions to change Claude Code behavior and UI. Published about 48 hours ago.”
Now: “A YouTube video by Chase AI explaining Claude Mods, a feature that lets developers write JavaScript functions to change Claude Code behavior and UI. Published about 50 hours ago.”
What moved
Was: “Up from 77,509 to 171,323 views in 36 hours - about 28 times the typical pace of items SteadyVane tracks on YouTube. 1,714 likes and 71 comments.”
Now: “Up from 77,509 to 171,323 views in 36 hours. 1,714 likes and 71 comments.”
Pace: the brief said about 14 times the typical pace; the honest pace is about 6.1 times the typical pace of other repos SteadyVane tracks on GitHub.
Misleading Said: “The acceleration in stars tracks with recent launches including Claude Mods.”
Where: why it matters, X, Bluesky
SteadyVane's own readings show no pick-up when Mods launched on Oct 1. Stars picked up on Oct 3 (about 13 an hour, against about 5 to 6 an hour before), and the cause is not known.
Misleading Said: “The star acceleration tracks with recent feature launches.”
Where: Threads, Instagram
The same causal claim without naming Mods; no launch is known to coincide with the Oct 3 pick-up.
What moved
Was: “Up from 147,423 to 149,337 stars in about 13 days - about 14 times the typical pace of items SteadyVane tracks on GitHub. 25,438 forks.”
Now: “Up from 147,423 to 149,337 stars in about 13 days. 25,438 forks.”
Why it matters
Was: “The acceleration in stars tracks with recent launches including Claude Mods. If you review pull requests already, Claude Code slots into that workflow instead of asking you to adopt a new editor.”
Now: “Stars picked up on Oct 3 (about 13 an hour, against about 5 to 6 an hour before); the readings show no pick-up when Claude Mods launched on Oct 1, and the cause is not known. If you review pull requests already, Claude Code slots into that workflow instead of asking you to adopt a new editor.”
Pace: the brief said about 8.8 times the typical pace; the honest pace is about 6.2 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Imprecise Said: “If you already use Claude for coding, Mods let you adapt it to your workflow without switching tools.”
Where: why it matters, Threads
Mods exist only in Claude Code (the command line and the Code tab of the Claude Desktop app); someone coding with Claude anywhere else would have to switch tools.
Was: “Up 4,561 views in four hours - about 8.8 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. 126,218 views about 41 hours after it was published. 1,425 likes and 71 comments.”
Now: “Up 4,561 views in four hours. 126,218 views about 41 hours after it was published. 1,425 likes and 71 comments.”
Why it matters
Was: “If you already use Claude for coding, Mods let you adapt it to your workflow without switching tools. The early tutorial velocity suggests people want the setup steps, not just the announcement.”
Now: “If you already use Claude Code (the CLI or the Code tab of the Claude Desktop app), Mods let you adapt it to your workflow without switching tools. The early tutorial velocity suggests people want the setup steps, not just the announcement.”
Pace: the brief said about 9.2 times the typical pace; the honest pace is about 8.1 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
No other problem was found in its texts.
What moved
Was: “Up 9,334 views in eight hours - about 9.2 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. 360,063 views about 113 hours after it was published. 2,713 likes and 218 comments.”
Now: “Up 9,334 views in eight hours. 360,063 views about 113 hours after it was published. 2,713 likes and 218 comments.”
Pace: the brief said about 16.6 times the typical pace; the honest pace is about 4.2 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “this workflow automates posting without replacing your judgment on what to publish”
Where: why it matters, Threads
No source shows a human review step, and the source points the other way: the AI creates, schedules and publishes the posts.
Was: “Up 28,353 views over 48 hours - about 16.6 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. 35,085 views about 50 hours after it was published. 15 comments.”
Now: “Up 28,353 views over 48 hours. 35,085 views about 50 hours after it was published. 15 comments.”
Why it matters
Was: “If you manage social accounts for a small team, this workflow automates posting without replacing your judgment on what to publish. The pace suggests creators want the step-by-step, not just the concept.”
Now: “If you manage social accounts for a small team, it shows how to use Claude Code with Blotato to have an AI agent create and schedule posts across several platforms.”
Pace: the brief said about 24.5 times the typical pace; the honest pace is about 22 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Mods do not edit Claude Code's source, but they run inside its own process, are not sandboxed, and can replace built-in features and redraw its interface.
Misleading Said: “without writing a full extension”
Where: why it matters, X
A mod is itself a complete Claude Code plugin and installs like one; it can be small, and Claude Code can write it for you.
Misleading Said: “Anthropic shipped the feature about a week before this video was published.”
Where: what it is
Mods launched on October 1, 2026, about four days before the video went up.
Its video's description carried a referral or affiliate link and a sponsor mention: see Sponsored and referral videos.
What it is
Was: “A ten-minute YouTube video showing how to build custom Claude Code Mods - TypeScript functions that hook into Claude Code to rewrite prompts, modify the interface, and add features without touching Claude's internals. Anthropic shipped the feature about a week before this video was published.”
Now: “A ten-minute YouTube video showing how to build custom Claude Code Mods - TypeScript functions that run inside Claude Code's own process and hook its events, so they can rewrite prompts and tool calls, replace built-in features and redraw Claude Code's interface without editing its source code. Anthropic shipped the feature on October 1, 2026, about four days before this video was published.”
What moved
Was: “Up 27,000 views in 12 hours - about 24.5 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. 1,145 likes and 114 comments.”
Now: “Up 27,000 views in 12 hours. 1,145 likes and 114 comments.”
Why it matters
Was: “If you already use Claude Code, Mods let you customize its behavior for your workflow without writing a full extension. The velocity suggests people want the how-to, not just the announcement.”
Now: “If you already use Claude Code, Mods let you customize its behavior for your workflow: a mod is itself a small Claude Code plugin, and Claude Code can write one for you. The velocity suggests people want the how-to, not just the announcement.”
Pace: the brief said about 13.4 times the typical pace; the honest pace is about 17 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “the prompt logic, and the handoff between steps”
Where: what it is, X, Bluesky, Threads, Instagram, the take
Only the tools have support: the description lists Tella.tv, Claude and Hyperframes. Nothing read shows the video covers prompt logic or handoffs; the words came from model-written notes.
Misleading Said: “If you edit videos and already have Claude Code access, this is a replicable workflow.”
Where: why it matters, Threads, Instagram
The creator's tools list names Tella.tv and Hyperframes alongside Claude, so Claude Code access alone is not enough.
Imprecise Said: “Showing the handoff between tools is the part most AI automation videos skip.”
Where: the take, Instagram
A claim about most automation videos with no evidence, resting on a handoff the video is not shown to cover.
Was: “A YouTube video showing a step-by-step workflow for using Claude Code to automate video editing tasks - the exact tools, the prompt logic, and the handoff between steps. Published about two days before the latest measurement.”
Now: “A YouTube walkthrough of automating video editing with Claude Code. The creator lists Tella.tv, Claude and Hyperframes as the tools used. Published about two days before the latest measurement.”
What moved
Was: “Up 32,000 views in 16 hours - about 13.4 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. 1,392 likes and 46 comments.”
Now: “Up 32,000 views in 16 hours. 1,392 likes and 46 comments.”
Why it matters
Was: “If you edit videos and already have Claude Code access, this is a replicable workflow. The pace says content creators want the concrete playbook, not a demo reel.”
Now: “If you edit videos, note that the workflow uses Tella.tv and Hyperframes alongside Claude (per the video's tools list), so Claude Code access alone is not enough to replicate it.”
Take
Was: “Showing the handoff between tools is the part most AI automation videos skip.”
Now: “A complete walkthrough is worth more than a quick demo.”
Pace: the brief said about 22.8 times the typical pace; the honest pace is about 24 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Wrong Said: “then teaches the concept instead of writing the code”
Where: Bluesky, Instagram
False: VibeWise asks for your approach first; once you approve the design, Claude writes the code and explains what it changed and why.
Misleading Said: “A Claude Code plugin forcing users to learn”
Where: headline, the day's title, X, Bluesky, Threads, Instagram
"Forces you to learn" is the creator's pitch; the plugin's own README calls it learning-first and leaves the user in control. And the 336,000 views were the video's, not the plugin's.
Misleading Said: “then teaches the underlying concept instead of writing the code outright”
Where: what it is, Threads
This leaves the reader thinking it teaches instead of coding; it explains first, then writes the code once you approve the design.
Misleading Said: “The numbers say the teaching-first approach resonates beyond the usual Claude Code audience.”
Where: why it matters, Instagram
SteadyVane measured views, likes and comments only; nothing shows who watched.
Headline
Was: “A Claude Code plugin forcing users to learn hit 336,000 views in under five days”
Now: “A video on VibeWise, a learning-first Claude Code plugin its creator says forces you to learn, hit 336,000 views in under five days”
What it is
Was: “VibeWise - an open-source plugin for Claude Code that asks users for their approach before answering, then teaches the underlying concept instead of writing the code outright. Built by Noah Kim and published on GitHub.”
Now: “VibeWise - an open-source plugin for Claude Code that asks for your approach first and explains concepts and tradeoffs; once you approve the design, Claude writes the code and explains what it changed and why. Built by Noah Kim and published on GitHub.”
What moved
Was: “Up 143,000 views in 44 hours - about 22.8 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. 336,000 views total about five days after publishing. 6,394 likes and 245 comments.”
Now: “Up 143,000 views in 44 hours. 336,000 views total about five days after publishing. 6,394 likes and 245 comments.”
Why it matters
Was: “If you are teaching people to code with AI tools, this is a working model for forcing comprehension instead of copy-paste. The numbers say the teaching-first approach resonates beyond the usual Claude Code audience.”
Now: “If you are teaching people to code with AI tools, this is a working model for forcing comprehension instead of copy-paste.”
Pace: the brief said about 54.5 times the typical pace; the honest pace is about 56 times the typical pace of other stories SteadyVane tracks on Hacker News.
Wrong Said: “this is the first European frontier model that competes directly on benchmarks”
Where: why it matters, Threads
Not the first: Mistral Large (February 2024) was benchmarked against GPT-4 and Claude 2. Mistral calls Large 4 the best open-weights model from the US or Europe, on its own preliminary benchmarks.
Wrong Said: “The first European model competing directly with GPT-4 and Claude”
Where: Bluesky, Instagram
Not the first (see above), and Large 4 is compared with GPT-6 Astra and Claude Opus 5 and 5.5, not GPT-4.
Misleading Said: “It runs 49 billion active parameters”
Where: what it is
Mistral lists 52 billion active parameters; its launch post and X post first said 49 billion, and the post was revised.
Imprecise Said: “Mistral Large 4 hit 1,827 points on Hacker News in 20 hours”
Where: headline, the day's title, X, Bluesky, Threads, Instagram
1,827 came about 21 hours after the post went up; 20 hours was how long SteadyVane tracked it, from 434 points.
Headline
Was: “Mistral Large 4 hit 1,827 points on Hacker News in 20 hours”
Now: “Mistral Large 4 hit 1,827 points on Hacker News about 21 hours after it was posted”
What it is
Was: “A 1-trillion-parameter frontier model from Mistral AI, released as a public preview. It runs 49 billion active parameters and is available now through Mistral Studio's API.”
Now: “A 1-trillion-parameter frontier model from Mistral AI, released as a public preview. Mistral lists 52 billion active parameters (its launch post and X post first said 49 billion), and it is available now through Mistral Studio's API.”
What moved
Was: “Up from 434 to 1,827 points in 20 hours - about 54.5 times the typical pace of items SteadyVane tracks on Hacker News. 1,088 comments.”
Now: “Up from 434 to 1,827 points in 20 hours. 1,088 comments.”
Why it matters
Was: “If you rely on Claude or GPT-4, this is the first European frontier model that competes directly on benchmarks and is available to switch to this week.”
Now: “If you rely on Claude or GPT, note that Mistral calls Large 4 the best open-weights model from the US or Europe, on its own preliminary benchmarks; it is available in preview this week.”
Take
Was: “The pace says people are reading the announcement. Whether they switch depends on the API reliability in month two.”
Now: “People are paying attention to the announcement. Whether they switch depends on the API reliability in month two.”
Pace: the brief said about 9.6 times the typical pace; the honest pace is about 6.5 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Wrong Said: “that did not exist in earlier versions”
Where: why it matters
Earlier Claude models already wrote code-rendered motion graphics, for example with Remotion in Claude Code, months before Opus 5.5.
Misleading Said: “professional motion graphics”
Where: what it is, X, Bluesky, Threads, Instagram
"Professional" is the creator's own claim: the creator says the results look professionally made.
Misleading Said: “users discovered it by prompting the model's coding abilities”
Where: what it is, X, Threads, Instagram
No reliable source for that, and the official Claude account reportedly posted code-made videos at launch. Anthropic's announcement does not mention motion graphics; users have posted Opus 5.5 motion graphics since its release.
Imprecise Said: “The demo is real.”
Where: the take, Threads, Instagram
SteadyVane did not check this demo. The video shows the creator's results; other users have posted similar Opus 5.5 motion graphics.
What it is
Was: “A video showing Claude Opus 5.5 writing code that renders into professional motion graphics. The capability is not in Anthropic's announcement - users discovered it by prompting the model's coding abilities.”
Now: “A video showing Claude Opus 5.5 writing code that renders into motion graphics, which the creator says look professionally made. Anthropic's announcement does not mention motion graphics; users have been posting motion graphics made with Opus 5.5 since its release.”
What moved
Was: “Up from 69,944 to 102,207 views over 40 hours - about 9.6 times the typical pace of items SteadyVane tracks on YouTube. 1,986 likes and 54 comments.”
Now: “Up from 69,944 to 102,207 views over 40 hours. 1,986 likes and 54 comments.”
Why it matters
Was: “If you need motion graphics and can render code, this is a working capability in Claude Opus 5.5 that did not exist in earlier versions.”
Now: “If you need motion graphics and can render code, this works in Claude Opus 5.5, though earlier Claude models could already write code-rendered motion graphics (for example with Remotion in Claude Code).”
Take
Was: “The demo is real. The workflow is whether you can describe what you want in a prompt that produces renderable code.”
Now: “The video shows the creator's results; other users have posted similar Opus 5.5 motion graphics. The workflow is whether you can describe what you want in a prompt that produces renderable code.”
Pace: the brief said about 443 times the typical pace; the honest pace is about 21 times the typical pace of other stories SteadyVane tracks on Hacker News.
Misleading Said: “Costs about 75 percent less to run than Haiku 4.5 on average”
Where: why it matters, Threads
Anthropic's own blended estimate, stated as fact. List prices are 90 percent lower for prompts up to 100,000 tokens and 50 percent lower above that.
Imprecise Said: “Claude Haiku 5.5 reached 892 points on Hacker News in 16 hours”
Where: headline, the day's title, X, Bluesky, Threads, Instagram
892 came about 17 hours after the post went up; 16 hours was how long SteadyVane tracked it, from 148 points.
Imprecise Said: “Anthropic says it costs 75% less to run than Haiku 4.5”
Where: Bluesky
Anthropic said about 75 percent less on average; leaving out "about" and "on average" turns an average into a flat cut.
Headline
Was: “Claude Haiku 5.5 reached 892 points on Hacker News in 16 hours”
Now: “Claude Haiku 5.5 reached 892 points on Hacker News about 17 hours after it was posted”
What moved
Was: “Up from 148 to 892 points in about 16 hours - about 443 times the typical pace of items SteadyVane tracks on Hacker News. 429 comments.”
Now: “Up from 148 to 892 points in about 16 hours. 429 comments.”
Why it matters
Was: “Costs about 75 percent less to run than Haiku 4.5 on average, which changes the math for high-volume tasks if the quality holds up in production.”
Now: “Anthropic says it costs about 75 percent less to run than Haiku 4.5 on average (list prices are 90 percent lower for prompts up to 100,000 tokens and 50 percent lower above that), which changes the math for high-volume tasks if the quality holds up in production.”
Pace: the brief said about 288 times the typical pace; the honest pace is about 16 times the typical pace of other stories SteadyVane tracks on Hacker News.
Wrong Said: “The model naming jumped from four to six.”
Where: the take, X, Bluesky, Threads, Instagram
False: GPT-6 follows GPT-5.6. The error came from a model-written summary that called the launch "GPT-4o".
Misleading Said: “designed to simplify AI adoption for non-technical workers”
Where: what it is, Threads, Instagram
No OpenAI source says that. Intelligent UI lets ChatGPT answer with interactive interfaces (charts, buttons, forms); OpenAI pitches GPT-6 at ChatGPT's 1.2 billion weekly users, Sol on paid plans and Luna on Free and Go.
Imprecise Said: “OpenAI announced GPT-6 with Intelligent UI and reached 644 points in 16 hours”
Where: headline, Instagram
644 came about 17 hours after the post went up; 16 hours was how long SteadyVane tracked it, from 64 points.
Imprecise Said: “reached 644 points on Hacker News in about 16 hours”
Where: X, Bluesky, Threads
With no starting point this reads as time since posting, which was about 17 hours.
Headline
Was: “OpenAI announced GPT-6 with Intelligent UI and reached 644 points in 16 hours”
Now: “OpenAI's GPT-6 and Intelligent UI announcement reached 644 points on Hacker News about 17 hours after it was posted”
What it is
Was: “OpenAI released GPT-6 models with a new Intelligent UI feature designed to simplify AI adoption for non-technical workers. Ships in two variants, Sol and Luna.”
Now: “OpenAI released GPT-6 models with Intelligent UI, which lets ChatGPT answer with interactive interfaces (charts, buttons, forms). OpenAI pitches GPT-6 at ChatGPT's 1.2 billion weekly users: Sol on paid plans, Luna on Free and Go.”
What moved
Was: “Up from 64 to 644 points in about 16 hours - about 288 times the typical pace of items SteadyVane tracks on Hacker News. 353 comments.”
Now: “Up from 64 to 644 points in about 16 hours. 353 comments.”
Take
Was: “The model naming jumped from four to six. Whether the interface simplifies or just hides complexity is what the first real projects will show.”
Now: “GPT-6 follows GPT-5.6. Whether the interface simplifies or just hides complexity is what the first real projects will show.”
Pace: the brief said about 10 times the typical pace; the honest pace is about 4.8 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
No other problem was found in its texts.
What moved
Was: “Up from 94,938 to 132,824 views in 56 hours - about 10 times the typical pace of items SteadyVane tracks on YouTube. 1,072 likes and 97 comments. Published about four days ago.”
Now: “Up from 94,938 to 132,824 views in 56 hours. 1,072 likes and 97 comments. Published about four days ago.”
Pace: the brief said about 8.9 times the typical pace; the honest pace is about 4.6 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Imprecise Said: “The description carries a referral link and a bonus-credit code”
Where: what it is, Threads, X, Bluesky, Instagram, the take
Nothing confirms that the link or the code pays the creator. What can be verified is a promotional Lumina link and a bonus-credit code (LUMINADR2), so the video doubles as an advertisement.
Was: “A Hindi-language tutorial from the channel Digital Raj showing how to make long AI videos with Lumina, a video generation platform. The description carries a referral link and a bonus-credit code, so it doubles as an advertisement.”
Now: “A Hindi-language tutorial from the channel Digital Raj showing how to make long AI videos with Lumina, a video generation platform. The description carries a promotional Lumina link and a bonus-credit code (LUMINADR2), so it doubles as an advertisement.”
What moved
Was: “Up 28,515 views in 40 hours, from 72,349 to 100,864 - about 8.9 times the typical pace of other items SteadyVane tracks on YouTube (AI-tool searches). 1,385 likes and 204 comments, about four days after it was published.”
Now: “Up 28,515 views in 40 hours, from 72,349 to 100,864. 1,385 likes and 204 comments, about four days after it was published.”
Take
Was: “A tutorial with an affiliate code is still a tutorial. Read it as an ad first, a lesson second.”
Now: “A tutorial with a bonus-credit code is still a tutorial. Read it as an ad first, a lesson second.”
Pace: the brief said about 8.3 times the typical pace; no honest pace can be stated (its count went down by 32% between two readings).
Misleading Said: “picked up 202 installs in 20 hours”
Where: headline, what moved, X, Bluesky, Threads, Instagram
The Marketplace's install counter read 458 and later 660, but it jumped around in between (565, 544, 548, then 970, then 660), so the 20-hour gain cannot be pinned down.
Headline
Was: “Nexrall Code, an autonomous coding agent for VS Code, picked up 202 installs in 20 hours”
Now: “Nexrall Code, an autonomous coding agent for VS Code: its install counter went from 458 to 660 in 20 hours, with a jump to 970 in between”
What moved
Was: “Up 202 installs in 20 hours, from 458 to 660 - about 8.3 times the typical pace of other items SteadyVane tracks on the VS Code Marketplace. The listing was published about four days ago.”
Now: “Its Marketplace install counter read 458 on Oct 8 (14:05 UTC) and 660 on Oct 9 (10:05 UTC), but it jumped around in between (970 at 06:05 UTC), so the 20-hour gain cannot be pinned down. The listing was published about four days ago.”
Pace: the brief said about 13.8 times the typical pace; the honest pace is about 6.5 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Imprecise Said: “Tracker study of nine AI chatbots draws 93,557 views”
Where: the day's title
The views were on heise & c't's YouTube video about the study, not on the study.
What moved
Was: “Up 31,503 views in 24 hours, from 62,054 to 93,557 - about 13.8 times the typical pace of other items SteadyVane tracks on YouTube (AI-tool searches). 2,921 likes and 286 comments, about 45 hours after it was published.”
Now: “Up 31,503 views in 24 hours, from 62,054 to 93,557. 2,921 likes and 286 comments, about 45 hours after it was published.”
Pace: the brief said about 7.5 times the typical pace; the honest pace is about 9 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
No other problem was found in its texts.
What moved
Was: “Up 18,141 views in 16 hours, from 20,657 to 38,798 - about 7.5 times the typical pace of other items SteadyVane tracks on YouTube. 306 likes and 9 comments, about 26 hours after it was published.”
Now: “Up 18,141 views in 16 hours, from 20,657 to 38,798. 306 likes and 9 comments, about 26 hours after it was published.”
Pace: the brief said about 34.2 times the typical pace; the honest pace is about 15 times the typical pace of other repos SteadyVane tracks on GitHub.
No other problem was found in its texts.
What moved
Was: “Up 314 stars in 20 hours, from 224 to 538 - about 34.2 times the typical pace of other items SteadyVane tracks on GitHub. 7 forks, about two days after it was published.”
Now: “Up 314 stars in 20 hours, from 224 to 538. 7 forks, about two days after it was published.”
Pace: the brief said about 12.6 times the typical pace; the honest pace is about 8.9 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Imprecise Said: “"10x" is the video's claim, not a measurement”
Where: why it matters
Whether the video presents some measurement behind "10x" could not be checked. Anthropic's post reports no measurable loss after the trimming, not a gain.
Its video's description carried a referral or affiliate link, a sponsor mention and a promo code: see Sponsored and referral videos.
What moved
Was: “Up 24,314 views in 20 hours, from 47,920 to 72,234 - about 12.6 times the typical pace of other items SteadyVane tracks on YouTube. 920 likes and 60 comments, about 38 hours after it was published.”
Now: “Up 24,314 views in 20 hours, from 47,920 to 72,234. 920 likes and 60 comments, about 38 hours after it was published.”
Why it matters
Was: “The guidance behind it appears real: Anthropic's post reportedly says most of Claude Code's system prompt was removed with no measurable loss on coding evals. Trimming bloated instruction files is a cheap thing to try. "10x" is the video's claim, not a measurement.”
Now: “The guidance behind it appears real: Anthropic's post reportedly says most of Claude Code's system prompt was removed with no measurable loss on coding evals. Trimming bloated instruction files is a cheap thing to try. "10x" is the video's own claim; the post reports no measurable loss, not a gain.”
Pace: the brief said about 19.3 times the typical pace; the honest pace is about 29 times the typical pace of other products SteadyVane tracks on Product Hunt.
Wrong Said: “It was published about 20 hours ago.”
Where: what moved
It launched on Product Hunt on Oct 10 at 07:01 UTC, about 30 hours before the brief; 20 hours was its age at SteadyVane's last reading.
Misleading Said: “outputs editable code rather than video”
Where: what it is, Threads
Motion builds each animation as code instead of with a video-generation model, but the result plays as a short video and downloads as an MP4.
What it is
Was: “Anthropic's new Claude features that generate live dashboards and animated explainers from a prompt. Dashboards needs a paid plan and a connected data source; Motion is limited to Team and Enterprise plans and outputs editable code rather than video.”
Now: “Anthropic's new Claude features that generate live dashboards and animated explainers from a prompt. Dashboards needs a paid plan and a connected data source; Motion is limited to Team and Enterprise plans and builds each animation as editable code instead of using a video-generation model; the result plays like a short video and can be downloaded as an MP4.”
What moved
Was: “Up 58 votes in 8 hours, from 165 to 223 - about 19.3 times the typical pace of other items SteadyVane tracks on Product Hunt. 4 comments. It was published about 20 hours ago.”
Now: “Up 58 votes in 8 hours, from 165 to 223. 4 comments. It launched on Product Hunt on Oct 10 at 07:01 UTC, about 30 hours before this brief.”
Pace: the brief said about 7.6 times the typical pace; the honest pace is about 4.7 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
No other problem was found in its texts.
Its video's description carried a referral or affiliate link or code and a sponsor mention: see Sponsored and referral videos.
What moved
Was: “Up 42,893 views in 64 hours, from 63,096 to 105,989 - about 7.6 times the typical pace of other items SteadyVane tracks on YouTube. 1,479 likes and 86 comments, about five days after it was published.”
Now: “Up 42,893 views in 64 hours, from 63,096 to 105,989. 1,479 likes and 86 comments, about five days after it was published.”
Pace: the brief said about 17.6 times the typical pace; the honest pace is about 19 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Was: “Up 27,077 views in 12 hours, from 557,663 to 584,740 - about 17.6 times the typical pace of other items SteadyVane tracks on YouTube. 7,501 likes and 174 comments, about four days after it was published.”
Now: “Up 27,077 views in 12 hours, from 557,663 to 584,740. 7,501 likes and 174 comments, about four days after it was published.”
Pace: the brief said about 25.9 times the typical pace; the honest pace is about 19 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Misleading Said: “Per his own X post, the repo now has more stars than React.”
Where: why it matters, Threads
The fact is true (287.1k GitHub stars against React's 251.1k on Oct 13), but no such post by Pocock could be found; the one found making the comparison is by Gergely Orosz.
What moved
Was: “A video walking through it went from 47,767 to 128,107 views in 28 hours - about 25.9 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. It has 1,809 likes and 46 comments, about 91 hours after it was published.”
Now: “A video walking through it went from 47,767 to 128,107 views in 28 hours. It has 1,809 likes and 46 comments, about 91 hours after it was published.”
Why it matters
Was: “Agents can now build a feature unattended, so the open question is how you direct and check them. Pocock's answer is a reusable set of skills you install once, using 'npx skills@latest add mattpocock/skills', rather than a fresh prompt each time. Per his own X post, the repo now has more stars than React.”
Now: “Agents can now build a feature unattended, so the open question is how you direct and check them. Pocock's answer is a reusable set of skills you install once, using 'npx skills@latest add mattpocock/skills', rather than a fresh prompt each time. The repo has more GitHub stars than React (287.1k against 251.1k on Oct 13).”
Pace: the brief said about 11.9 times the typical pace; the honest pace is about 9.3 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
Wrong Said: “The six repos are named only in the video”
Where: why it matters, Threads
False: the video's description links every repo it covers.
Was: “A video about it went from 139,971 to 156,204 views in 12 hours - about 11.9 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. It has 3,272 likes and 86 comments, about 46 hours after it was published.”
Now: “A video about it went from 139,971 to 156,204 views in 12 hours. It has 3,272 likes and 86 comments, about 46 hours after it was published.”
Why it matters
Was: “Sorting through Claude add-ons is a time cost for anyone building with it, and a shortlist saves some of it. The six repos are named only in the video, and the 'best' ranking is the creator's own judgment. The description also promotes a free skill, so part of the video is promotional.”
Now: “Sorting through Claude add-ons is a time cost for anyone building with it, and a shortlist saves some of it. The video's description links the repos he covers, and the 'best' ranking is the creator's own judgment. The description also promotes a free skill, so part of the video is promotional.”
Pace: the brief said about 18 times the typical pace; the honest pace is about 16 times the typical pace of other AI-tool videos SteadyVane tracks on YouTube.
No other problem was found in its texts.
What moved
Was: “A video about it went from 36,454 to 64,960 views in 12 hours - about 18 times the typical pace of AI-tool videos SteadyVane tracks on YouTube. It has 1,503 likes and 144 comments, about 19 hours after it was published.”
Now: “A video about it went from 36,454 to 64,960 views in 12 hours. It has 1,503 likes and 144 comments, about 19 hours after it was published.”
Sponsored and referral videos
When SteadyVane read them, the descriptions of these videos carried a sponsor, a referral or affiliate link, a promo code or a bonus-credit offer. Most of the briefs that featured them did not say so; each item now carries a note. From October 10, 2026, SteadyVane's daily pick was meant to leave such videos out, but at first its check read only the start of each description and missed some ways of writing a referral or a code - which is how the videos of October 10, 12 and 13 below got through. Both gaps were closed by October 14, 2026. The list below is kept up to date: a video is added when SteadyVane finds one of these in its description, also after its brief went out.
October 13, 2026 #2 · a referral or affiliate link and a promo code
The site's own wording
Every page's footer said: "Every item is measured, fact-checked, and graded before it is published." False: the tracker lists items that are never fact-checked or graded, and brief items were checked on their main claim only, with a doubtful verdict still publishing.
How it works said every number in a brief is checked against the data, and an unsupported number blocks the item. False as written: numbers in model-written notes counted as supported, any number up to 12 passed, and so did any number within 6% of an allowed one, whatever its unit. That is how "1.7 million subscribers" went out.
How it works said the pace compares each item with the typical pace of everything tracked on the same source. False: it compared with one part of the source only, used one short stretch, and counted items unevenly (see the pace numbers above).
How it works called the grade "an independent second reviewer". Misleading: it is the same kind of AI model as the writer, and it never saw the Instagram caption or the day's title.
How it works said uncertain claims are hedged. Nothing made the writer do it.
How it works opened with "it only writes what the numbers support". False, as this page shows.
The home page, the weekly email and the weekly top 10 said the items were ranked by how many times faster than the typical item on its own source each one climbed. False for Hacker News, where the comparison was with search results only, and the multiple was one short stretch.
This wording was replaced on October 13 and 14, 2026 with a true description: How it works.
The weekly top 10, the weekly emails and the daily videos
The weekly top 10 of October 13 (posted on X, Bluesky, Threads, Instagram and the Telegram channel) and the weekly emails of October 6 and 13 ranked items by the old pace numbers, and said they were ranked by how many times the typical pace of their source each item climbed - not true for Hacker News, where the comparison was with search results only. The top 10's first two, Claude Haiku 5.5 at 442.5 times and GPT-6 at 288 times, were far too high. The weekly pages on this site keep the order the emails went out in, and show each item's honest pace and its headline as corrected. From the email of October 20, the ranking uses the honest pace.
Ranked by the honest pace, the top 10 of October 13 would have been in this order:
Mistral Large 4 hit 1,827 points on Hacker News about 21 hours after it was posted about 56 times
Claude Dashboards and Motion lets you ask for live dashboards and animated explainers about 29 times
A video on VibeWise, a learning-first Claude Code plugin its creator says forces you to learn, hit… about 24 times
A Claude Code Mods walkthrough climbed from 61,000 to 88,000 views in 12 hours about 22 times
Claude Haiku 5.5 reached 892 points on Hacker News about 17 hours after it was posted about 21 times
A video lists nine Claude Code mods, reportedly worth changing how you work about 19 times
A walkthrough automating video editing with Claude Code added 32,000 views in 16 hours about 17 times
OpenAI's GPT-6 and Intelligent UI announcement reached 644 points on Hacker News about 17 hours… about 16 times
A Mac tool that lets AI agents draw arrows on your screen gained 314 stars in 20 hours about 15 times
A video claiming Anthropic engineers gave Claude Code a 10x boost hit 72,234 views about 8.9 times
The daily videos of October 11, 12 and 13 called the brief's picks the day's "biggest movers" and numbered them as a ranking; they were the brief's picks in the brief's order (on October 12, "Number 1" had the lower pace). Their pace numbers carried the same error. On October 12, the video featured as "Number 1" - Kevin Stratvert's tutorial on editing video with ChatGPT Codex - was sponsored by OpenAI, the maker of what it shows; neither the brief nor SteadyVane's daily video said so. Since October 14 the daily video speaks of the day's stories, not its movers, and states the honest pace.
Posts on X
Before the daily brief existed, an earlier version of SteadyVane drafted posts with an AI writer, to be posted to X by hand. This site showed 68 of them, dated June 11 to September 7, 2026, on its day pages and item pages under "Posted on X": 62 recorded as posted, and 6 approved for posting whose posting was never recorded. None of them was checked before posting the way every brief is now. All of them were taken off this site, and each page that showed them says so. Of the 62, these 27 described tests that never happened, were wrong or misleading, or made claims that could not be confirmed, and were deleted from X:
Tests that never happened (4)
June 17, 2026 “Ran a 7B model locally for a week on actual agent tasks (200+ calls/day). Median latency: 340ms. Cost: $0.”
This test never happened. The post was drafted by an AI writer, and no one ran it.
June 23, 2026 “Ran both GLM 5.2 and Opus on the same classification task (1200 items, structured output). Opus: 94% accuracy, $2.40, 8min. GLM: 91% accuracy, $0.31, 11min.”
This test never happened. The post was drafted by an AI writer, and no one ran it.
June 25, 2026 “GLM-5.2 is the first open-weight model I'd actually consider for an agent that runs unsupervised. The benchmarks matter less than this: it can recover from a malformed tool call without hallucinating the next three steps.”
It presented a first-hand observation that no one made.
September 3, 2026 “Tested it on 12 reports I'd written: flagged 2 as "likely AI" (both were). Missed 3 I'd heavily edited.”
This test never happened. The post was drafted by an AI writer, and no one ran it.
Wrong (11)
June 26, 2026 “OpenKnowledge is a markdown editor with local-first storage and an LLM sidebar. The interesting part: it's MIT-licensed”
OpenKnowledge was licensed GPL-3.0-or-later; and the write-up the audit read describes no LLM sidebar: it works with agents through MCP, skills and agentic search.
June 27, 2026 “OpenAI's new model requires government approval per user. If you're building a SaaS that calls GPT-5”
The reports described a voluntary, customer-by-customer review of an early preview for some enterprise customers, with no binding approval requirement - and the model was GPT-5.6, not GPT-5.
June 29, 2026 “The router itself is a small model that classifies intent”
Wayfinder routes deterministically, with no model call.
July 9, 2026 “The headline is isolated declarations”
The headline of TypeScript 7.0 was its native compiler rewritten in Go, about 8 to 12 times faster.
July 26, 2026 “The blog names three mechanisms: where you put examples, how you structure long contexts, when to use caching.”
Anthropic's post lists six changes (judgement over rules, interfaces over examples, progressive disclosure, simple tool descriptions, auto-memory, rich references) and covers neither caching nor long-context structure.
August 1, 2026 “Routing added latency and complexity. The win—saving $0.02/call by picking a cheaper model—didn't justify maintaining the logic.”
The summary of Manifest's post the audit read gives other reasons: routing on a prompt's complexity failed because the difficulty shows up later, in tool use; caching cut cost more; and switching models broke consistency. It has no $0.02 figure.
August 7, 2026 “Free ChatGPT just got GPT-4.5 Luna.”
It was GPT-5.6 Luna, as the post's own source said.
August 11, 2026 “OpenAI published a 'Daybreak' cyber defense post but no GPT 5.6 announcement. The HN title appears to be speculation or misconstrued.”
OpenAI's post of August 10 introduced GPT-5.6-Cyber, available through Daybreak Red. The title was right; the post was wrong.
August 13, 2026 “DeepSeek V4 Pro pricing just landed at $0.40/M input tokens. That's the same capability tier as GPT-4o but at ~1/5th the cost for most workloads.”
Likely wrong: on October 13, 2026 the listed price was $1.32 per million input tokens (the launch-day price could not be confirmed), and the GPT-4o comparison had no source.
August 17, 2026 “Claude's docs just published their system prompt guide. The useful bit: they show you exactly how the model weights instructions (task > examples > tone)”
The page is Anthropic's release notes for the system prompts of Claude's apps, not a guide, and it contains no such ordering.
August 29, 2026 “24B params, Apache 2.0. Quantized 4-bit fits in 16GB VRAM”
GLM-5.3 is a mixture-of-experts model of about 750 billion parameters (40 billion active), under its own GLM-5.3 License, and it does not fit in 16GB.
Misleading (7)
June 16, 2026 “content-addressed by default, NAT traversal built in. The interesting part is that docs sync even when you're offline—writes queue and merge when peers reconnect.”
Iroh 1.0 is peer-to-peer connections over QUIC, dialed by public key, with NAT traversal built in, as the post said (hole punching with relay fallback); but content-addressed data and document sync are separate projects built on top of it.
June 18, 2026 “the benchmark shows it closing the gap on closed models for reasoning tasks”
Artificial Analysis put GLM-5.2 in line with proprietary models on real-world agentic work (GDPval-AA); its reasoning gains were measured against its own predecessor.
July 3, 2026 “Every demo of an LLM 'watching' video is actually the LLM watching 1fps stills.”
The tool picks a frame at every scene change, with one frame per second only as the minimum, and adds the transcript and, optionally, the audio (as its version read on October 13, 2026 describes it); "every demo" is a sweeping claim.
July 10, 2026 “Postgres rewritten in Rust, passing 100% of regression tests.”
That was the project's own claim; the post stated it as fact, without saying whose it was.
July 13, 2026 “Claude Code sends 33k tokens before it reads your prompt. OpenCode sends 7k.”
That was one person's measurement, from the title of the item the post was about, stated as a general fact (the post's "4.7x" worked out from it was right).
August 14, 2026 “book-to-skill converts a technical PDF into a Claude MCP skill”
It makes a Claude Code skill (an Agent Skill), not an MCP one; and the post's "cite specific sections" is not claimed for it - answers can be tied to chapters.
September 5, 2026 “Google's AI Mode surfaced products 21.6% more expensive than traditional search — same query, different result set.”
The 21.6% was for the same products, matched by Google's product ID: the same item, priced higher, often with a different lead seller.
Not confirmed (5)
July 1, 2026 “Claude Code embeds invisible markers in every request — zero-width characters that survive copy-paste.”
The only source the audit could read, a secondary blog, describes such characters but cites no primary source, and it does not support "survive copy-paste".
July 2, 2026 “OpenTag routes @agent mentions in Slack/GitHub to Claude Code or Codex”
OpenTag's README, as read on October 13, 2026, describes Slack only, says GitHub "is not a second request inbox" and does not describe the pluggable routing the post also claimed; an older write-up says GitHub and Slack.
July 24, 2026 “the repo shows actual automation primitives—cut detection, scene segmentation, batch export hooks”
Cut detection, scene segmentation and batch export hooks could not be confirmed; the post's "GPL'd, runs locally" is right (its AI generation uses paid cloud credits).
August 15, 2026 “it keeps state, follows links, and outputs structured notes”
This could not be confirmed, nor the post's "chains searches": Mole's own pitch is an enforced budget and verified quotes, and it runs as a single program with your own API keys.
August 27, 2026 “Qwen3.8-Flash-Next benchmarks at ~2x the speed of GPT-4o mini on standard evals.”
No source was found for it; DataCamp compares the model only with Qwen3.7-Plus.
The 6 drafts whose posting was never recorded (June 11, June 13 (two), August 10, August 22 and August 24, 2026) were not checked by the audit; they were looked for on X and deleted where found.
Two later posts on SteadyVane's X account were deleted too:
October 4, 2026 “Every number checked before it posts.”
The check existed, but it let wrong numbers through (see above).
October 12, 2026 “557,663 → 584,740 views in 12 hours, at 17.6× the typical pace on YouTube.”
The counts were right, but the comparison was with the AI-tool videos SteadyVane tracks, not all of YouTube; 17.6 was the old kind of pace number; and the video's description carried referral and affiliate links.
AI summaries on the older day pages
The day pages from June 11 to September 19, 2026 showed a "Summary" under each item. Those lines were written by the AI that scores how relevant an item is, were never fact-checked, and included internal notes and at least one plain error ("GPT-4.5 Luna" for GPT-5.6 Luna). All 200 were removed, and each day page that showed them says so.