Proceedings · Session S-896 · filed November 18, 2026

AI & Emerging Tech in R&DSession paper

GPT-6.2 Astra scores 62 on independent AI index at $4 per task

GPT-6.2 Astra scored 62 on the Artificial Analysis Intelligence Index at $4 per task, trailing Claude Fable 5.2's 65 at $26.54, with Epoch AI finding little change on research tasks.

By Sophie Lindqvist3 min read652 words

Summary

  • GPT-6.2 Astra scored 62 on the Artificial Analysis Intelligence Index at about $4 per task; Claude Fable 5.2 scored 65 at $26.54.
  • OpenAI announced GPT-6.2 Astra on Nov. 17, claiming a 9-point gain on Terminal-Bench-Science at less than a third of the predecessor's cost.
  • Epoch AI found little overall change from the previous Astra on research assignments; reports improved but fell short of its standards.
  • API pricing: $8 per million input and $40 per million output tokens, 20% below GPT-6 Astra and four times GPT-6.2 Sol's rates.
  • In OpenAI's Bio-Probe forecasting test, Astra's best Brier score was 0.232, slightly worse than GPT-6 Astra's 0.226.
OpenAI’s GPT-6.2 Astra trails Claude Fable 5.2 by three points on an independent AI index at a fraction of the cost per
FigureOpenAI’s GPT-6.2 Astra trails Claude Fable 5.2 by three points on an independent AI index at a fraction of the cost per — geralt / pixabay

OpenAI's GPT-6.2 Astra, announced Nov. 17, scored 62 on Artificial Analysis's Intelligence Index at roughly $4 per task at maximum reasoning effort — three points behind Anthropic's Claude Fable 5.2, which achieved a score of 65 at $26.54 per task, according to independent testing from the analytics firm.

The gap frames the launch for research organizations weighing model budgets. Astra costs $8 per million input tokens and $40 per million output tokens in the API, 20% below GPT-6 Astra and four times the pricing of GPT-6.2 Sol, the everyday model OpenAI released alongside it, at $2 and $10 per million tokens respectively.

OpenAI positioned the new model squarely at scientists. In the announcement, the company called it "the best model in the world for scientists," citing a new high on Terminal-Bench-Science, a benchmark testing 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 the company's own figures.

What does independent testing show?

Artificial Analysis's results suggest the headline gain sits mostly in cost per task rather than raw capability. On its Intelligence Index, Astra's 62 at maximum reasoning effort trails Claude Fable 5.2's 65 while undercutting the Anthropic model's $26.54 cost by more than a factor of six.

The comparison table from Artificial Analysis:

  • Claude Fable 5.2 — Max effort — Intelligence Index 65 — $26.54 per task
  • GPT-6.2 Astra — Max effort — Intelligence Index 62 — $4.00 per task
  • Gemini 4 Carbon — High effort — Intelligence Index 60 — $5.48 per task
  • Claude Opus 5.5 — Max effort — Intelligence Index 59 — $10.04 per task
  • GPT-6.2 Sol — Max effort — Intelligence Index 57 — $1.01 per task

Does the model hold up on real research work?

An outside evaluation from Epoch AI, published in OpenAI's system card, found little overall change from the previous Astra on research-grade assignments. Epoch gave the model tasks drawn from its own researchers' work, including designing experiments and assessing AI data centers. The new model wrote somewhat better reports, but they still fell short of Epoch's standards — a direct signal that experimental design remains a weak point for labs considering the model for autonomous research workflows.

OpenAI's own forecasting test complicated the picture further. In Bio-Probe, an exploratory evaluation in the system card, models read the setups of 21 unpublished experiments from seven academics and estimated how likely each result was to support a stated hypothesis. Measured by Brier score, where lower is better, GPT-6.2 Astra's best result was 0.232 — worse than GPT-6 Astra's 0.226 and better than GPT-6.2 Sol's 0.261.

The baselines matter for interpretation. Assigning even odds to every hypothesis scores 0.250, and expert forecasts scored 0.266. That places Astra between a coin flip and a domain expert, but behind its own predecessor. OpenAI called the sample small and noted many outcomes were hard to anticipate from the existing literature.

Where does the cheaper sibling fit?

GPT-6.2 Sol, the least expensive model in the Artificial Analysis table at $1.01 per task, may cover routine lab coding work. On Cognition's FrontierCode 2.0, which tests whether changes to real codebases meet maintainers' standards for merging, OpenAI reported Sol came within half a point of Astra at about one-fifth the cost per task, with both running at maximum reasoning effort.

For R&D managers, the launch presents a split decision. Astra offers near-frontier scores at a mid-tier price, but Epoch's evaluation and the Bio-Probe results show measured limits on the exact tasks — experiment design and research forecasting — that the marketing targets. OpenAI has not yet released follow-up evaluations addressing those gaps.

via openai.com (Original)

Filed under

  • gpt-6-2-astra
  • ai-model-benchmarks
  • research-forecasting
  • experiment-design
  • r-d-costs
Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Correspondent covering money, funding and market shifts around research and development at Hypothesis Wire.

242 articles

References

  1. Moonshot's Kimi K3.1 Pro Nears Frontier at Half Opus Pricing
  2. OpenAI's BioBasicsBench: top model hits 66.7% on wet-lab data reading
  3. GPT-6 Astra scores 62.7% on interactive reasoning benchmark
  4. Anthropic's Fable 5.2 nearly doubles hardware-design score at double task cost
  5. DOE Releases 35 Agent Skills for National-Laboratory Science

« Previous articleNext article »