OpenAI’s Early Tester: GPT-6.2 Astra Ran 11 Days, Zero Fixes

Theo, the developer who posts as @theo (t3.gg) and tested OpenAI’s GPT-6.2 Astra before launch, published the letter he sent OpenAI about the model on Tuesday, 17 November, in an X article titled “Why I hate GPT-6.2 Astra”. He wrote that OpenAI “fixed a small handful of my complaints, but not enough for me to recommend using it”.

Our number: on his own benchmark project, a TypeScript compiler rewrite in Rust he calls ts-rust, Theo wrote that GPT-6.2 Astra ran for 7 days and then 4 more, 11 days in all, and landed zero changes, by our arithmetic. He wrote that earlier OpenAI models used $300k+ in tokens on the project without breaking ground, while Anthropic’s Opus finished it for about $20k, a gap of roughly 15x on his figures, by our arithmetic.

Business Pill · Usage-Based Pricing

An evergreen Business Engineer explainer on usage-based pricing, paying for work rather than seats. It teaches the general idea only and does not discuss OpenAI, Anthropic or the letter.

The key insight: An early-access tester’s letter judges GPT-6.2 Astra on cost per finished task, 11 days of runs and zero changes landed on his project by our arithmetic, while the independent scores so far put it level with Claude Fable 5.2 at the same list price.

What the Tester Wrote

Theo posted at 18:23 UTC that “I think this is the worst model OpenAI has ever released”, and published the full letter at 18:35 UTC. In it he wrote: “I do not like 6.2 Astra. I would not choose this over Opus 5.5 for my work as a developer”.

He listed six strengths first, including that the model is a “Great reviewer, catches things other models miss”, “World class at 3d & game stuff” and that “Computer use is nuts”. He wrote that it helped him build a reverse proxy that “will save me $25k/month”.

Token spend on the tester’s ts-rust compiler project as he reported it: earlier OpenAI models $300k+ wit
Token spend on the tester’s ts-rust compiler project as he reported it: earlier OpenAI models $300k+ without breaking ground, Anthropic’s Opus about $20k, finished in under two weeks; GPT-6.2 Astra ran 7 days and then 4 more and landed no changes (no spend figure given). Source: Theo (@theo, t3.gg), X Article, 17 November 2026; the tester’s own figures, not independently verified. The roughly 15x ratio and 11-day total are our arithmetic.

His Three Numbered Complaints

  • Knowledge gaps. On long tasks, he wrote, the model stops to ask questions the user cannot answer. Two hours into a request to build a game level, it was “blocked” on a question about video file names.
  • Unreadable output. He called its explanations “the most unparsable outputs of any llm I’ve used this year” and wrote that he switches to Opus 5.5 to summarise its runs.
  • Loops. On ts-rust, he wrote, it ran 7 days with “0 improvements” and made the compiler “56% SLOWER”, then ran “4 more days and landed 0 changes”.

In a separate post, he wrote that he and a colleague each asked the model for a new video-game level in different games and “Both were named “Bellwether””.

How Far the Letter Travelled

X showed 203,507 impressions, 2,849 likes and 268 replies on the article when we read it about 50 minutes after posting. The AI commentator @kimmonismus, who shared a screenshot of it, wrote: “This really is a catastrophic release”.

What the Measured Data Shows

The independent scores so far point the other way from the letter. Epoch AI scores GPT-6.2 Astra at 174.17 on its capability index, 0.67 points from Claude Fable 5.2, and both models list at $8 per million input tokens, as we reported earlier on Tuesday.

OpenAI rates the model High for AI self-improvement, the first public model it has rated at that level, and released a legal configuration, Astra for Law, the same day.

The Business Lever

The lever is developer choice in the model picker. Theo’s complaint is about cost per finished task, not list price: he wrote that users will “see how much usage it burns and how much less it got done compared to Opus”. The letter landed the day before Anthropic is expected to price its IPO, reportedly above $80 a share.

The Structural Read

As we read the letter, its complaints are about long-running agent work, blocked runs, unreadable summaries and loops, not about the short coding tasks where Theo wrote that it writes fine code.

As we read the figures, the roughly 15x token gap compares earlier OpenAI models with Opus on one project; the letter gives no token figure for GPT-6.2 Astra itself.

Theo (@theo), letter to OpenAI, 17 November 2026

“I really, really, REALLY hope I’m wrong about all this”

Three Implications

FOR DEVELOPERS Theo wrote that he would not choose GPT-6.2 Astra over Opus 5.5 for his work, and that OpenAI fixed a small handful of his complaints before launch.

FOR OPENAI Epoch AI scores the model at 174.17, 0.67 points from Claude Fable 5.2, at the same $8 input list price, from our earlier reporting.

FOR ANTHROPIC The letter landed the day before Anthropic is expected to price its IPO, reportedly above $80 a share, from our earlier reporting.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework says: “Foundation models become relatively commoditized as orchestration layers capture differentiation.”

As we read it, a developer switching models mid-run to get a readable summary is choosing at the orchestration layer, where the framework places the differentiation. The framework predates this release and does not discuss GPT-6.2 Astra or the letter.

What Is Not Established

The letter is one developer’s experience during early access, and he wrote that “I am likely overreacting (I am admittedly known for that)”. OpenAI fixed some of his complaints before launch, and his letter does not say which. The $300k+ and $20k figures are his and cover earlier OpenAI models, not GPT-6.2 Astra.

He also wrote that OpenAI gave him unlimited tokens for early access. The day counts and the roughly 15x ratio are our arithmetic from his figures. We did not contact OpenAI or Theo.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The Agentic Harness War

A Business Engineer analysis of how value shifts from foundation models to the orchestration layers built on top of them.

Read the Map of AI →

The Bottom Line

An early tester of GPT-6.2 Astra published his letter to OpenAI saying he would not choose it over Opus 5.5, citing 11 days of runs that landed zero changes on his benchmark project, by our arithmetic. Epoch’s pre-launch scores put the model level with Claude Fable 5.2.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. The letter, its quotations and the token, day and speed figures are from Theo’s X article of 17 November 2026 and two of his posts that day, read in full through the X API; they are his own account of early-access testing and were not independently verified. Engagement figures are X’s public counts at the time we read the article. Epoch scores, list prices, OpenAI’s self-improvement rating and the Anthropic IPO report are from our earlier coverage. The 11-day total and the roughly 15x ratio are our arithmetic. We did not contact OpenAI or Theo.

Sources: Theo (@theo), X Article: Why I hate GPT-6.2 Astra, 17 Nov 2026 · @theo on X, 18:23 UTC, 17 Nov 2026 · @theo on X, 18:42 UTC, 17 Nov 2026 · @kimmonismus on X, 18:42 UTC, 17 Nov 2026

Scroll to Top
FourWeekMBA