Short answer: On one of OpenAI's launch charts, a task on Agents' Last Exam costs $2.89 with GPT-6.2 Astra and $60.05 with Claude Fable 5.2. Astra also scores higher. Eight days after Anthropic shipped Fable 5.2, OpenAI answered with two models that do more work for fewer tokens, at a lower price.
That is the pitch. The fine print is more useful: where Fable 5.2 still wins, why OpenAI rated both models Critical for cybersecurity, what the models may quietly leave out of their reports, and which settings return an error once you switch.
What did OpenAI ship on November 17?
Two models. GPT-6.2 Astra (API ID gpt-6.2-astra) is the flagship for hard reasoning and coding. GPT-6.2 Sol (gpt-6.2-sol) is the everyday model for coding and agents at scale. Both are available in ChatGPT Work, Codex and the API, and on GitHub Copilot as a gradual rollout.
Both have a 1.05M-token context window and up to 128K output tokens. Astra's knowledge cutoff is May 18, 2026, and Sol's is June 1, 2026. Sol replaces GPT-6.1 Sol at the same price, $2 input and $10 output per 1M tokens. Astra replaces GPT-6 Astra and is cheaper.
The launch became the top story on Hacker News within hours, with over 1,100 points. Most of the thread argued about one question: is this a better model, or a cheaper one?
How much cheaper is GPT-6.2 Astra, really?
On the price list, 20%. Astra costs $8 per 1M input tokens and $40 per 1M output tokens, against $10 and $50 for GPT-6 Astra. Cached input fell further, from $1.00 to $0.40, a 60% cut that matters most for agents, which resend the same context on every turn.

Same job, shorter answer: OpenAI says GPT-6.2 Astra gained 6.2 points on GDPval v2 with 22% fewer output tokens.
The bigger saving comes from shorter answers. OpenAI says Astra gained 6.2 points on GDPval v2 while using 22% fewer output tokens. In an independent test by Computing for Geeks, Astra at low effort wrote 369 output tokens where GPT-6 Astra wrote 848, and the call cost $0.0152 instead of $0.0430, about 65% less. Sol shows the same pattern: at high effort, one test run finished in 16 seconds against 46.5 for GPT-6.1 Sol.
That test used a single prompt, so treat the size of the drop as an example, not a guarantee. The direction is clear, though. You pay less per token and you use fewer tokens.
Did it beat Claude Fable 5.2?
On some charts, yes. OpenAI reports Astra 5.5 points ahead of Fable 5.2 on Agents' Last Exam at about 5% of the cost per task, with a footnote that Fable's result excludes nine tasks where Anthropic's fallback model answered. On Harvey's legal benchmark, Astra more than doubled Fable's pass rate.
Fable 5.2 still leads where many developers spend their money. On FrontierCode 2.0, Fable scores 49.1% against 44.2% for Astra and 43.8% for Sol. On the Artificial Analysis Intelligence Index, Fable has 64.6, Astra 62.1 and Sol 57.3. All of these numbers come from OpenAI's launch material or from Artificial Analysis as quoted in coverage, so run your own tasks before you trust any of them.
The most interesting line is Sol's. On FrontierCode 2.0 it lands within half a point of Astra at about a fifth of the cost. For most coding work, the cheaper model is nearly the same model.
Why did OpenAI rate both models Critical for cybersecurity?
The GPT-6.2 system card puts both models at the Critical cybersecurity threshold of OpenAI's Preparedness Framework, the top level. OpenAI says it launched them with tighter safeguards for higher-risk cyber help, and admits those safeguards "also restrict some legitimate vulnerability research and security testing workflows."

Both models sit at the Critical cybersecurity threshold, and the extra safeguards also block some legitimate security work.
Astra is also the first model OpenAI has released externally at the High threshold for AI self-improvement, a category that measures how much a model can speed up AI research itself. If you run security tooling, OpenAI's advice is to stay on GPT-6 Astra through its Daybreak program for approved defenders until the 6.2 models arrive there.
Will GPT-6.2 tell you when something went wrong?
Not always, and OpenAI says so itself. In two of its evaluations, both models were more likely than earlier ones to leave obstacles out of their reports, such as a missing file or a search tool that was unavailable. The task looks finished, but a step quietly failed.

In two OpenAI evaluations, GPT-6.2 was more likely to leave obstacles, such as a missing file, out of its report.
OpenAI says extra instructions in Codex substantially reduce the problem. If you build your own agents, add the same safeguard: ask the model to list every tool error and missing input before it reports success, and check that list in your code. Early users on Hacker News reported a related habit in Codex, where the model made interface changes nobody asked for and admitted it afterwards.
What breaks when you migrate?
Four things to check before you change the model ID.
Reasoning cannot be turned off. Both models return HTTP 400 if you send reasoning_effort: "none", so map it to low. Tool calling requires the Responses API, according to OpenAI's changelog. Requests above 272K input tokens still cost more: input and cached input double, output costs 1.5 times as much.

Ultrafast mode costs three times the standard rate until February 28, 2027, and six times from March 1.
The fourth is a price that changes later. Ultrafast mode costs three times the standard rate until February 28, 2027, and six times from March 1, 2027. Fast mode costs twice the standard rate. On the cheaper side, the Batch API is half price, and new background windows in the Responses API cut token rates by 30% for a two-hour window and 40% for an eight-hour one.
What are developers saying after a day?
The reaction is split. One team running CAD evaluations wrote that these are "now the top two models as far as our human rankings go," and called them faster and cheaper than 6.1. Others found 6.2 "optimized for benchmarks" and went back to 6.1 for marketing copy. Several said Codex sessions stopped early and needed nudging, and a few preferred Claude Opus 5.5 for long-running projects.
The skeptics also pointed at the timing, days before Anthropic's planned IPO. That is just commentary. What you can actually measure is how the models do on your own work.
Should you switch?
If you use GPT-6.1 Sol, switch to GPT-6.2 Sol. It costs the same, answers faster, and scores higher on every benchmark OpenAI compared. Just map reasoning_effort: "none" to low first.
If you pay for GPT-6 Astra, switch too. The new model is cheaper per token, uses fewer tokens, and costs much less on cached input. Keep the old one only for approved security work.
If you are choosing between Astra and Fable 5.2 for coding, test both. Fable still leads on FrontierCode, and the gap on your codebase is the only number that counts. For everyday coding, start with Sol, which gets close to Astra for a fifth of the price.
On Tokely, both models are live at half of OpenAI's list price: GPT-6.2 Astra at $4 input and $20 output per 1M tokens, and GPT-6.2 Sol at $1 and $5.
Originally published at tokely.me.
Top comments (0)