Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 September 1st 2026. Both are the same underlying model wrapped in different safeguard layers, one generally available and one restricted to vetted organisations. Both landed on the same date, a Tuesday, with no warning.
That is a completely ordinary week in AI now, and it’s exactly why I’ve spent the last few weekends building an AI model release tracker of my own. Two releases before lunch felt like the right day to put it out.
TrackAi is a website that tracks AI model releases as they happen: what’s confirmed, what’s still rumoured, and what people actually experienced once they got their hands on the model. Every release gets a page that separates the lab’s own claims from real reports from real users.
First, the Intention: I wanted to write about every Ai model release. Then work got busy.
The original plan was simple: write something here every time a significant model shipped. Keep a running record. It lasted about as long as you’d expect, because I have a job, and the job gets busy exactly when the releases cluster.
The problem isn’t the writing. It’s everything before the writing. Noticing a model shipped at all. Finding the announcement. Working out whether the thing announced last month ever actually came out, or quietly slipped. Remembering that the interesting comment about it was in a forum thread you read on your phone nine days ago and can no longer find.
Keeping up with AI releases had become a second job, and I already have a first one. So rather than keep failing at it manually, I built the thing that does it for me, and made it public. If it’s useful to me on a Tuesday morning, it’s probably goint to be useful to someone else too.
What TrackAi tracks: Ai Models shipped, announced, and merely rumoured
The homepage is a timeline. Every model sits on the date it came out. If it hasn’t shipped yet, it sits on the date it’s expected instead. Confirmed releases carry their provider’s real brand colour. Anything unreleased is greyed out and marked EST., so colour on the grid means the model actually exists.
Rumours get tracked too, which turns out to be half the value. Plenty of models are “coming in the next few weeks” for months. The tracker records the predicted date, and if that date passes with nothing shipped, the entry strikes through its own estimate and gets flagged as overdue rather than quietly sitting there looking like history.
Then each model page splits in two. Above the line sits the claim: the lab’s own summary, whichever benchmark numbers it chose to publish, and a link to the announcement. Below sits the reality: short takeaways from people who used it on real work, tagged by task, each linking back to where it came from. The site never blends the two into a single score, because they don’t deserve the same trust.
The Automation Layer: it updates itself, or it emails me
This is the part that makes an AI model release tracker survivable as a weekend project. Almost nothing about the site requires me to sit down at it. The work runs in n8n on a schedule, outside the website entirely, and splits by how much trust the data has earned.
The claim layer publishes itself
When a lab announces a model, a nightly workflow finds the announcement, pulls out the summary and whatever figures are printed on that page, and publishes without asking me. That’s only safe because it’s tightly constrained: it reads the provider’s own page and nothing else, never a news write-up, never an aggregator. It can’t convert, average or infer a number. If it isn’t confident, it returns nothing rather than guessing, and the page says the summary was drafted automatically with the source one click away.
The reality layer waits for me
Reports about how a model actually performs are lower-confidence by nature, so none of them go live on their own. Overnight, collectors read Hacker News, developer forums and YouTube comments looking for posts where somebody describes using a specific tracked model, summarise each into a short takeaway, and drop it in a queue.
Then at 8am I get one email with everything waiting, and I approve or reject from my phone. No email is sent when the queue is empty. A daily email that’s usually empty is how you teach yourself to ignore it.
One detail I’m oddly pleased with: the email doesn’t contain an “approve” link you click directly. Mail providers and security scanners fetch every URL in a message before you ever see it, so a one-click approve link would mean Google’s scanner silently publishing my entire review queue on delivery, every morning. Instead the link opens a read-only page, and approving is a deliberate second action. One extra tap is the difference between a review queue and an auto-publisher.
User Test Sources: some of it automates, some of it I just do myself
Not every platform lets you read it programmatically at a price a weekend project can justify, so the honest answer is that trackai is part pipeline and part me.
| Source | How |
|---|---|
| Hacker News | Automated. Open API, no key, good signal on launch days. |
| Developer forums | Automated. Cursor, OpenAI and Hugging Face: the best signal of anything I tested, since people post there while actively using the model on real work. |
| YouTube | Automated, comments only. Noisier than forums, still useful. |
| X / Twitter | Manual. The API starts at $200 a month, which is not a weekend-project budget, so I file the good ones by hand. They embed as the original post rather than being paraphrased. |
| Written, paused. Reddit closed self-service API access in late 2025, so the collector is finished and waiting on a credential that isn’t available to a project this size. |
Which is the real reason the submit form exists. The automated sources have blind spots, the manual ones have my blind spots, and neither covers the person who spent yesterday afternoon finding out how a model handles something specific.
Submitting a report: about a minute, and a human reads it
- Pick the models. Search by name or provider. Up to five on one report, because people run the same prompt across several models to compare them. That’s one test, not five.
- Say what happened. A task category and a short takeaway describing what you asked it to do and how it went, not a review.
- Link the source. Your own post, a thread, a video: anywhere the claim can be checked.
- It joins the queue. Same 8am email as everything else. Nothing publishes automatically, and rejected reports never appear.
There’s no voting or ranking here, and no algorithm deciding what surfaces. One editorial pass and a link back to the source so you can disagree with it.
The Tech Stack.
One constraint shaped the architecture. The site runs on Vercel’s free plan, which only gives any single task ten seconds to finish before it gets cut off. That’s plenty for loading a page, but nowhere near enough to scrape forums, summarise a post with Claude, or check whether a rumoured release actually shipped. So the website itself does almost nothing. It reads a database and renders. Everything slow happens elsewhere, on a schedule, and writes back through a token-authenticated ingest API.
| Next.js 16 | App Router, React 19. Ships with a file warning that literally says this is not the Next.js you know: middleware.ts became proxy.ts, params are all async, and Turbopack is the default bundler. |
| Neon | Postgres with the schema as plain TypeScript. Direct connection for migrations, pooled for serverless traffic. |
| Tailwind v4 | A small token set: near-black ground, one gold accent, and real provider brand colours carrying the rest. |
| n8n | Every collector, summariser and digest. Off Vercel entirely, so the ten-second limit never applies to the slow work. |
| Claude | Haiku 4.5 classifies candidate posts, at roughly $1.50 a month. Opus 5 handles the claim layer, because that’s the one that publishes without me checking it. |
| Vercel + GitHub | Push to main, live in about thirty seconds. |
Join The August Dispatch Newsletter
For curated news, practical prompts, and how‑to guides—plus first dibs on workshops.
FAQ
What is trackai? trackai is a free AI model release tracker built by August Dispatch. It shows confirmed model launches, rumoured ones still waiting to ship, and short first-person reports from people who have actually used each model on real work.
Is trackai free to use? Yes. The grid is free to browse with no account required. Submitting a usage report is also free and takes about a minute.
How does trackai tell a confirmed release from a rumour? Confirmed models carry their provider’s real brand colour and link directly to the lab’s own announcement. Anything not yet shipped is greyed out and marked EST., and if the predicted date passes without a launch, the entry is flagged as overdue rather than left looking like settled history.
It’s live, and it’s better with reports in it
The grid is public and free, there’s no account to make, and you can subscribe to an email when new models land if you’d rather not check.
The claims side will fill itself in. The reality side only works if people write things down, and the most useful report is almost always the unglamorous one: the thing you asked a model to do this week, and whether it actually did it.
Built something with one of these models? File a report. What you asked for, what happened, and a link. About a minute, up to five models at once, and I read every one before it goes up.
Leave a Reply