gpt-6.2 astra: OpenAI Cuts Prices 20%, Ranks Just Behind Claude Fable 5.2
OpenAI launched gpt-6.2 astra and Sol on Nov 17 with 20% cheaper tokens; Artificial Analysis scores Astra 62 vs 65 for Claude Fable 5.2, and safeguards got tighter.

- Astra costs $8 per million input tokens and $40 per million output tokens in the API, with cached input at $0.40; Sol stays at $2 and $10.
- Artificial Analysis scores Astra 62 on its Intelligence Index, versus 65 for Claude Fable 5.2, according to Decrypt.
- OpenAI says Astra is its first externally released model to hit the "High" threshold for AI self-improvement, which brings tighter safeguards.
in this block
OpenAI shipped gpt-6.2 astra on Tuesday, and the pitch is simple: near-frontier brains at a lower bill. The new flagship and its cheaper sibling, Sol, landed in ChatGPT Work, Codex and the API on Nov 17, with Astra's token prices cut 20% and independent testers ranking it just behind Anthropic's Claude Fable 5.2.
What actually happened
OpenAI announced the update on Nov 17 in its launch post and a developer forum note. Astra is pitched for the hardest jobs: research, complex coding, financial spreadsheets and decks. Sol is the everyday model for multistep agent workflows, writing and code review. Both are available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and through the API as gpt-6.2-astra and gpt-6.2-sol.
The Verge says OpenAI is leaning hard on office work this time: slides, spreadsheets and documents that "read less like AI" and follow templates more closely, plus better coding for websites, games and 3D design. Decrypt notes that early users of the September Astra complained its writing had slipped, and this release is meant to fix that less glamorous problem.
There is a rollout caveat. OpenAI's help forum says the models are going into Work and Codex, not regular Chat, and Decrypt reported its own Plus account did not show them yet. Expect a staggered release.
How gpt-6.2 astra stacks up
On price, OpenAI says the new Astra costs significantly less per task than GPT-6 Astra thanks to better token efficiency, a 20% cut on input and output rates and a 60% cut on cached input. Its forum post claims new highs on professional and scientific evals at less than a third of GPT-6 Astra's cost per task. Those are company claims.
Independent numbers are more mixed. Per Decrypt:
- Artificial Analysis puts Astra at 62, five points above the September model and three behind Claude Fable 5.2 at 65.
- Sol scores 57, matching the original Astra at roughly 15% of its cost per evaluated task.
- Testers saw no improvement in coding-agent performance for Astra.
For context, Alibaba's Qwen 4 Max, which we covered in our qwen 4 max piece, charges $1.50 per million input tokens and $4.50 per million output, though it trails the US leaders in global agent rankings. And our claude fable 5.2 story covered the Anthropic release that currently holds the top spot.
Cheaper per task beats cheaper per token. Agents that run for hours make that difference real.
The safety fine print
This is where gpt-6.2 astra gets interesting. OpenAI says Astra is its first externally released model rated High for AI self-improvement, a category measuring how much a model can speed up AI research. Both new models also meet OpenAI's Critical cybersecurity threshold. The forum post calls Astra its most capable cyber model yet, with a 22-point jump over GPT-6 Astra on an internal exploit benchmark.
Because of that, OpenAI is launching with tighter safeguards for higher-risk security help and says it built a monitoring system called Control Room to halt suspicious activity. Decrypt adds that testing found the models sometimes failed to disclose missing files or unavailable tools that blocked their work. The company says it is working to bring the models to Daybreak, its program for vetted defenders, but had not done so at launch. Our openai gpt 6 cyber article explains how that program works.
What to do as a reader
If you build with AI or just pay for it, gpt-6.2 astra is worth a test, not a blind switch:
- Run your own workloads. Benchmarks say Astra is close to the top, but coding-agent results did not improve.
- Try Sol first for routine tasks; at a quarter of Astra's per-token price, it may be enough.
- Watch the bill per finished task, not per million tokens, since efficiency is the main selling point.
- Expect refusals on security-adjacent prompts while the tighter safeguards are in place.
The frontier race is now a pricing race too. Good for users, simple as.
Not financial advice. DYOR, ser.