Skip to content

OpenAIModelsSafety and security

OpenAI releases GPT-⁠6.2 Astra and Sol

Astra beats Claude Fable 5.2 on two tough tests at a fraction of the cost, while Fable leads on 3D design and the Artificial Analysis index. Pricing and access.

Story date On the site Türkçe

Instagram cover: OpenAI releases GPT-6.2 Astra and Sol

Cover of the Instagram post.

5 sourcesJump to sources

Sources: OpenAI, ChatGPT

In this story
  1. OpenAI’s strongest model for cyber
  2. Ahead of Fable 5.2 on Agents’ Last Exam
  3. Also ahead on Terminal-Bench-Science
  4. Fable 5.2 still leads in 3D design
  5. Fewer tokens, better results
  6. Car dynamics on a mountain road
  7. Pricing and access
  8. Not first on every test
  9. Sources5
  10. About this story
  11. On the same topic3

OpenAI’s strongest model for cyber

OpenAI calls GPT-6.2 Astra its most capable model ever for cyber. On the internal ExploitBench test it scores 22 points above GPT-6 Astra: 53.5% versus 31.5%.

Astra and Sol both meet the Critical cybersecurity threshold under OpenAI’s Preparedness Framework, so they launch with tighter safeguards for higher-risk cyber help. The chart was measured with those safeguards turned off.

Ahead of Fable 5.2 on Agents’ Last Exam

On this tough agent test, Astra scores 78.3% while Claude Fable 5.2 scores 72.8%. OpenAI says Astra costs just 5% as much per task: $2.89 versus $60.05.

OpenAI’s footnote says the Fable 5.2 result excludes nine tasks that used a fallback model.

Also ahead on Terminal-Bench-Science

Astra scores 77.1% on this scientific terminal test, 9 points above GPT-6 Astra at less than a third of the cost per task.

In OpenAI’s run, Claude Fable 5.2 scores 73.3%, but with a different agent harness. Anthropic reports 74.6% for Fable 5.2 in Claude Code.

Fable 5.2 still leads in 3D design

BenchCAD 2.1 Core tests whether models can build 3D parts from engineering drawings and fit them into assemblies. Claude Fable 5.2 leads at 83.5%, with Astra just 1 point behind at 82.5%.

Astra costs about a tenth as much per task: $5.43 versus $50.77. The more efficient Sol scores 73.1%.

Fewer tokens, better results

OpenAI says both models deliver better results on demanding tasks with fewer output tokens. On GDPval v2, Astra scores 6.2 points above GPT-6 Astra while using 22% fewer output tokens. The garden planning site in the video is one of OpenAI’s web design examples.

Car dynamics on a mountain road

OpenAI asked Astra to visualize the chassis dynamics of a car on a mountain road. Astra built the 3D model, wrote a Python solver for the interaction between the tires and suspension, and animated the results in Blender.

Pricing and access

Astra is the frontier model for the hardest projects, while Sol is the more efficient model for everyday work. OpenAI says most of its employees use Sol as their default.

In the API, Astra costs $8 per million input tokens and $40 per million output tokens, 20% less than GPT-6 Astra. Sol costs $2 and $10.

Both models are rolling out in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and are available in the API from day one.

Not first on every test

On the Artificial Analysis Intelligence Index chart on OpenAI’s own page, Claude Fable 5.2 leads with 64.6. Astra is second at 62.1, followed by Gemini 4 Carbon (60.0), Claude Opus 5.5 (59.1) and Sol (57.3).

So Astra beats Fable 5.2 on two tests but trails on the overall index. Its real edge is doing the same work at a much lower cost.

Sources

5 sources
  1. OpenAI“Introducing GPT-6.2 Astra and Sol”Official announcement: benchmarks, pricing, safety. · openai.com
  2. OpenAI (X)“GPT-6.2 Astra and Sol are now available”Announcement post and charts. · x.com
  3. ChatGPT (X)“And here we go…GPT-6.2 Astra and GPT-6.2 Sol”Rollout post. · x.com
  4. OpenAI (X)“ExploitBench Internal Port”Cyber chart. · x.com
  5. OpenAI (X)“Driving simulation”Driving simulation video. · x.com

About this story

Written by
· JARRUS
Story date
On the site
since Nov. 18, 2026
Instagram
@jarrus.tech posted this story on Nov. 18, 2026.
Report an error
Spotted an error in this story? [email protected] · Instagram

On the same topic

  1. 5 sourcesCorrected

    Anthropic’s IPO order book has closed: CNBC says institutional demand is around $400 billion

  2. 4 sources

    Altman and Brockman at DealBook: a gradual handover to AI is a terrible idea, humans must stay in control

  3. 5 sources

    Nvidia’s board does not plan to reinstate Jensen Huang as CEO; H200 sales to China are ending