OpenAI lowers prices for GPT-6.2 Astra and Sol

OpenAI unveils GPT-6.2 Astra and GPT-6.2 Sol, two models more affordable than their predecessors. Astra targets demanding professional uses, Sol handles everyday tasks.

OpenAI introduced GPT-6.2 Astra and GPT-6.2 Sol, two new language models available since November 17, 2026, for Plus, Pro, Enterprise, and Edu subscribers, as well as via the API. Astra targets the most demanding uses in science, law, and finance, while Sol is aimed at faster, everyday tasks. The challenge is twofold: reduce usage costs without sacrificing too much performance, a balancing act where models are now judged as much by the bill as by the score. And the bill, indeed, matters a lot.

Deux modèles, deux usages, une même logique de compression des coûts

With Astra, OpenAI pushes its model towards specialized work and professional workflows. The system has been trained to better handle conventions in law, finance, and accounting, with outputs designed to replicate the layouts, colors, and typography of the provided templates. Sol, meanwhile, is presented as a versatile and lighter model for everyday needs.

In both cases, OpenAI emphasizes the performance-price ratio. On the API, GPT-6.2 Astra is priced at $8 per million input tokens, $0.40 for cached inputs, and $40 per million output tokens. This is 20% less than its predecessor. GPT-6.2 Sol retains the pricing of GPT-6.1 Sol: $2 for input, $0.10 for cache, and $10 for output per million tokens.

Les benchmarks placent Astra près du sommet, mais pas partout

On evaluations published by OpenAI, Astra makes significant progress on complex tasks. In Agents’ Last Exam, it gains 5.2 points over GPT-6 Astra while costing less than a third of the price, and surpasses Claude Fable 5.2 by 5.5 points at 5% of its cost. On Harvey LAB-AA, the legal benchmark, OpenAI claims it nearly doubles Claude Fable 5.2’s success rate for about one-sixth of the cost. Sol remains close to Astra on code, with only a 0.5 point difference on FrontierCode 2.0, but at one-fifth the cost.

Third-party rankings tell a more nuanced story, however. In the Artificial Analysis Intelligence Index, Astra ranks second, just behind Claude Fable 5.2, and Sol rises to 57 points. According to this same scale, OpenAI holds the frontline on cost-performance balance, but the picture is less flattering on code: the Coding Agent Index shows no leap for Astra, and Sol drops by three points. The models remain performant, without statistical miracles. Benchmarks love podiums; real life, a bit less.

Un virage plus net vers la science, la sécurité et l’auto-amélioration

OpenAI also highlights gains in science and computation. Astra scores 9 points higher than GPT-6 Astra on Terminal-Bench-Science, and leads several tests related to 3D and robotics. Epoch AI assigns it 174 points in its Epoch Capabilities Index, tied with Claude Fable 5.2, but judges its actual performance still uneven on concrete research tasks.

On safety, both models cross the "Critical" threshold of OpenAI’s Preparedness Framework, while Astra becomes the first published model to reach the "High" threshold in AI self-improvement. This means the model could accelerate research on future systems, opening a much broader control issue than a simple benchmark ranking. OpenAI therefore announces strengthened safeguards for risky cyber queries, with restrictions that will also limit some legitimate security research uses.

Key points

  • November 17, 2026: launch of GPT-6.2 Astra and Sol.
  • Astra costs 20% less than its predecessor on the API.
  • OpenAI says Astra outperforms Claude Fable 5.2 on several tests.
  • Sol keeps the prices of GPT-6.1 Sol.
  • Artificial Analysis ranks Astra second in its index.
  • OpenAI announces strengthened safeguards for cyber uses.

By the numbers

  • $8 per million tokens — Astra API input, OpenAI, 2026.
  • $40 per million tokens — Astra API output, OpenAI, 2026.
  • 174 points — Astra’s score in the Epoch Capabilities Index, Epoch AI, 2026.
  • 57 points — Sol’s score in the Artificial Analysis Intelligence Index, Artificial Analysis, 2026.
  • 5.2 points — Astra’s lead over GPT-6 Astra in Agents’ Last Exam, OpenAI, 2026.

To read