Critics Pan OpenAI’s GPT-6.2 Astra; Anthropic Leads 65-62

OpenAI’s GPT-6.2 Astra scores 62 on the Artificial Analysis Intelligence Index, 3 points behind the 65 of Claude Fable 5.2 at its highest setting, yet several developers who tested it early said on X that they would not recommend it. Deedy Das, a partner at Menlo Ventures, wrote that the model “has been a massive fail for every critic that has reviewed it”.

Our number: on that index, Astra reaches 95% of Fable 5.2’s top score at 15% of its cost per task, $4.00 against $26.54, by our arithmetic. Of five named testers whose posts we read, four placed it below Anthropic’s models and the fifth said he does not recommend it (our count).

Business Pill · The Model and the Harness

An evergreen Business Engineer explainer on the difference between a model and the harness around it. It teaches the general idea only and does not discuss this launch.

The key insight: On the Artificial Analysis index, GPT-6.2 Astra scores 62 to Claude Fable 5.2’s 65 at 15% of its cost per task, by our arithmetic, while four of five named early testers we read placed it below Anthropic’s models.

What the Testers Said

Theo, who posts on X as Theo – t3.gg, wrote on Tuesday: “I think this is the worst model OpenAI has ever released”. In an X article 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 also listed strengths: “Great reviewer, catches things other models miss”, “Computer use is nuts” and “Code quality has improved meaningfully”. In another post he said “This model regresses to the mean like nothing I’ve ever seen”.

Matt Shumer wrote “I had a terrible experience testing them. They were essentially unusable” and “I do not recommend these models”. Matthew Berman called the models “good, although I’ve seen others really dislike them”, adding “They are not as good as Opus 5.5 and Fable. OpenAI came up short”.

Ben Davis wrote “It’s a capable model” and “The problem is, working with it sucks”. He estimated that “The two big claude models are probably 10-20% more capable, while being about 20x more pleasant to work with”. Dan McAteer wrote “To my taste, Fable 5.2 is clearly superior”, after 45 minutes of use.

Artificial Analysis Intelligence Index at each model's highest-scoring setting, with cost per index task: Clau
Artificial Analysis Intelligence Index at each model’s highest-scoring setting, with cost per index task: Claude Fable 5.2 (max) 65, $26.54; GPT-6.2 Astra (max) 62, $4.00; Gemini 4 Carbon (high) 60, $7.76; Claude Opus 5.5 (max) 59, $10.04. Source: Artificial Analysis LLM leaderboard, read 18 November 2026. 62/65 = 95% and $4.00/$26.54 = 15% are our arithmetic.

What Was Posted on X

Deedy Das, a partner at Menlo Ventures who lists Anthropic among his investments, posted this on 17 November with a collage of four testers’ posts. The piece reads each of those posts at source.

Matt Shumer’s post on GPT-6.2 Astra and Sol, quoted in the piece.

Ben Davis’s post, which calls Astra a capable model and criticises how it communicates; quoted in the piece.

Polymarket Money posted on 17 November 2026 that Google was still projected to have the best AI model at the end of November, at an 88% chance, after OpenAI launched GPT-6.2 Astra.

What the Benchmarks Show

On the Artificial Analysis leaderboard read on Wednesday, Claude Fable 5.2 at its max setting ranks first of 115 models with 65. GPT-6.2 Astra at max scores 62, level with Fable 5.2 at its xhigh setting. Gemini 4 Carbon scores 60 and Claude Opus 5.5 at max 59.

The same table lists a cost per task of $4.00 for Astra at max against $26.54 for Fable 5.2 at max, and 12.9 minutes per task against 41.1. Earlier this week we reported Astra as the first model to reach 100% on ARC-AGI-3, and that OpenAI cut its price 20% to $8, matching Claude Fable.

Update, 18 November 2026, 13:27 UTC: Epoch’s own work tasks put GPT-6.2 Astra fourth. On Epoch AI’s updated Epoch Automation Reports benchmark, where “A score of 100% means the output meets the internal standard we’d expect from an Epoch employee”, GPT-6.2 Astra averaged 65% and Claude Fable 5.2 82%. Epoch said on X that “All models failed to meet the Epoch employee standard”. More in our piece on Epoch’s results.

Why It Matters

The complaints are mostly about working with the model, not about raw scores. Davis wrote that the models show “how absurdly important model behavior and communication is”. Theo wrote that he spent a day writing a letter to OpenAI during early testing and that “They fixed a small handful of my complaints, but not enough for me to recommend using it”.

The launch came the same week OpenAI rated a model its first ‘High’ on AI self-improvement, and one early tester said Astra ran 11 days with zero fixes. The reception on X has so far been more negative than those results.

The Business Lever

The lever is the model picker inside coding tools and the API, where developers choose a model per task and can switch at no cost. A price cut and a benchmark score win the first trial. The posts above describe what decides the second: how the model behaves over a long session.

Who Is Speaking

Das lists Anthropic among his investments in his X profile. Theo wrote on Tuesday that “A big part of why I swung so hard pro-Claude the last few weeks was because I knew the gap wasn’t getting closed any time soon”. The testers are individual developers and commentators, not a sample of users.

The Structural Read

As we read the posts, the complaints are about behaviour over long sessions, such as communication and persistence, not about scores: Davis called it a capable model, and Theo listed code quality and computer use as strengths.

As we read the leaderboard, Astra is close to the top on score and far cheaper per task: 62 against 65, at $4.00 against $26.54.

Ben Davis on X, 17 November 2026

“These models are the ultimate showcase of how absurdly important model behavior and communication is”

Three Implications

FOR DEVELOPERS Matthew Berman wrote that the models are good but not as good as Opus 5.5 and Fable, and Matt Shumer wrote that he does not recommend them.

FOR OPENAI Theo wrote that OpenAI fixed a small handful of his complaints during early testing, but not enough for him to recommend the model.

FOR BENCHMARK READERS Artificial Analysis ranks Claude Fable 5.2 at max first of 115 models with 65, and GPT-6.2 Astra at max at 62.

The Business Engineer Lens

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

The framework says: “The surface, not the model, governs autonomy.”

As we read it, the Astra posts fit the framework: testers judged the model on how it behaves inside their tools over long sessions, not on its index score. The framework predates this launch and does not discuss it.

Update, 18 November 2026, 16:57 UTC: what the betting market says. On Polymarket’s Which company has the best AI model end of November? market, the Google outcome was priced at 92.5% at 16:54 UTC on 18 November, against the 88% chance Polymarket Money posted on 17 November after OpenAI launched GPT-6.2 Astra (post); OpenAI was at 1.6% and Anthropic at 5%. The market resolves on the company with the top-ranked model on the arena.ai Text Arena (Overall) leaderboard, style control off, when checked on 30 November 2026 at 12:00 p.m. ET. Prediction-market prices are a crowd estimate, not a forecast by us.

What Is Not Established

Five posts are not a survey, and we chose them because they were named in Das’s image or widely shared. Index scores are one benchmark suite run by one firm. We did not contact OpenAI or the testers.

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 moves from the model to the harness around it, and who is building which kind.

Read the Map of AI →

The Bottom Line

GPT-6.2 Astra scores 62 to Claude Fable 5.2’s 65 on the Artificial Analysis index, at 15% of its cost per task by our arithmetic. Four of five named early testers we read placed it below Anthropic’s models, and the fifth did not recommend it.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Quotations are from the testers’ own X posts and Theo’s X article, read in full on 18 November 2026; index scores and costs are from the Artificial Analysis leaderboard read the same day. The 95% and 15% figures are our arithmetic and the four-of-five tally is our count. We did not contact OpenAI or the testers.

Sources: Epoch AI, Epoch Automation Reports (17 Nov 2026) · Artificial Analysis LLM leaderboard, read 18 Nov 2026 · @deedydas on X, 22:27 UTC, 17 Nov 2026 · @theo, X article ‘Why I hate GPT-6.2 Astra’, 17 Nov 2026 · @mattshumer_ on X, 17 Nov 2026 · @MatthewBerman on X, 17 Nov 2026 · @davis7 on X, 17 Nov 2026 · @daniel_mac8 on X, 17 Nov 2026

Scroll to Top
FourWeekMBA