
An OpenAI chart showing how GPT 6.2 Astra fares against other models on the Vals Science Index, which gauges how well AI models analyze scientific data, infer mechanisms from experimental evidence and carry out research tasks across disciplines.
OpenAI is pitching its newest frontier model directly at researchers. In announcing GPT-6.2 Astra on Nov. 17, the company called it “the best model in the world for scientists,” citing a new high on Terminal-Bench-Science, a test of scientific work carried out in a terminal. Astra scored 9 percentage points above its predecessor, GPT-6 Astra, at less than a third of the cost, according to OpenAI.
Independent testing from Artificial Analysis shows the gain is mostly in cost per task. At maximum reasoning effort, Astra scored 62 on the firm’s Intelligence Index at about $4 per task. Anthropic’s Claude Fable 5.2 scored 65 at $26.54.
An outside evaluation found that Astra still struggles to design good research experiments. Epoch AI gave the model assignments drawn from its researchers’ work, including designing experiments and assessing AI data centers, and saw little overall change from the previous Astra, according to findings published in OpenAI’s system card. The new model wrote somewhat better reports, though they still fell short of Epoch’s standards.
OpenAI’s forecasting test showed the earlier Astra slightly ahead. In Bio-Probe, an exploratory evaluation in the system card, models read the setup of 21 unpublished experiments from seven academics and estimated how likely each result was to support a stated hypothesis. GPT-6.2 Astra’s best Brier score, where lower is better, was 0.232, against 0.226 for GPT-6 Astra and 0.261 for GPT-6.2 Sol. Assigning even odds to every hypothesis scores 0.250, and expert forecasts scored 0.266. OpenAI called the sample small and said many outcomes were hard to anticipate from the literature.
How GPT-6.2 Astra compares with recent Fable, Gemini and Sol models
| Model | Reasoning effort | Artificial Analysis Intelligence Index | Cost per task |
|---|---|---|---|
| Claude Fable 5.2 | Max | 65 | $26.54 |
| GPT-6.2 Astra | Max | 62 | $4.00 |
| Gemini 4 Carbon | High | 60 | $5.48 |
| Claude Opus 5.5 | Max | 59 | $10.04 |
| GPT-6.2 Sol | Max | 57 | $1.01 |
GPT-6.2 Sol, the least expensive model in the table, is the everyday model OpenAI released alongside Astra. On Cognition’s FrontierCode 2.0, which tests whether changes to real codebases meet maintainers’ standards for merging, OpenAI reported that Sol came within half a point of Astra at about one-fifth the cost per task, with both at maximum reasoning effort. In the API, Astra costs $8 per million input tokens and $40 per million output tokens, 20% less than GPT-6 Astra and four times Sol’s $2 and $10.




Tell Us What You Think!
You must be logged in to post a comment.