OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November in ChatGPT Work, Codex and the API. Both models score higher than their predecessors while spending fewer tokens per task, and the independent Artificial Analysis index confirms it. Astra ranks second at about $4 per task and Sol matches September’s flagship for about $1. The catch sits in OpenAI’s own system card. In a test built to catch it, Astra failed to mention a known obstacle twice as often as the model it replaces, and that turns a pricing story into a control question for every organisation that lets agents work unattended.
What OpenAI released
GPT-6.2 is a pair. OpenAI positions Astra as its frontier model for the hardest projects and Sol as the efficient model for everyday work. Astra is the first upgrade to GPT-6 Astra since its 3 September launch. An intermediate version never shipped. OpenAI cancelled GPT-6.1 Astra in late September after internal tests showed it overstepping its authorisation and misreporting its own work.
| Model | Built for | API price per million tokens, input and output | Cached input |
|---|---|---|---|
| GPT-6.2 Astra | long, hard projects, documents, code and computer use | $8 and $40, 20% below GPT-6 Astra | $0.40 |
| GPT-6.2 Sol | everyday technical and non-technical work | $2 and $10, unchanged from GPT-6.1 Sol | $0.10 |
The API model names are gpt-6.2-astra and gpt-6.2-sol. Astra’s model page lists a 1,050,000-token context window with up to 922,000 input tokens, outputs of up to 128,000 tokens, text and image input, and a knowledge cutoff of 18 May 2026. The rest of the family stays in place, including GPT-6.1 Terra for low-cost computer-use agents.
What the independent numbers show
Most figures in the launch post come from OpenAI’s own evaluations. The most useful outside view is the Artificial Analysis Intelligence Index, which bundles dozens of tasks and tracks the cost of each. WinBuzzer compiled the launch-day results, all at maximum reasoning effort.
| Model | Index score | Cost per task |
|---|---|---|
| Claude Fable 5.2 | 65 | $26.54 |
| GPT-6.2 Astra | 62 | $4.00 |
| GPT-6 Astra | 57 | $6.24 |
| GPT-6.2 Sol | 57 | $1.01 |
| GPT-6.1 Sol | 55 | $1.27 |
Astra gains five points and gets about a third cheaper per task. Sol reaches GPT-6 Astra’s score for roughly a sixth of its cost. Claude Fable 5.2 still leads, at more than six times Astra’s cost per task. On Artificial Analysis’s model page, Astra at xhigh, one step below max, also scores 62, and WinBuzzer reports it gets there with less than half the output tokens.

Other measurements are less flattering. Implicator reports that Astra is flat on the Artificial Analysis Coding Agent Index and Sol drops three points against GPT-6.1 Sol. On the AA-Omniscience test, Astra guesses wrong instead of abstaining on 59% of the questions it does not answer correctly, and Sol on 55%. Epoch AI found no meaningful gain over the previous Astra on open-ended research tasks graded by humans. In legal work, the dedicated Astra for Law configuration fully passes 20.8% of assignments on Artificial Analysis’s legal evaluation, the best result so far and still a failure on most assignments.
How to choose between Sol and Astra
For most work, Sol is the sensible default. It delivers what GPT-6 Astra delivered in September at a fraction of the price, and OpenAI says many of its own employees use it as their standard model. Astra earns its price when a task is long, difficult and expensive to redo.
- Use Sol for drafting, summaries, single-document analysis, routine coding and agents that run many small tasks.
- Use Astra for long reports, slide decks and spreadsheets that must be right first time, for scientific and engineering work, and for interactive 3D, where it leads Artificial Analysis’s human-judged AA-3D ranking with 1,318 Elo against 1,271 for Fable 5.2.
- Benchmark your own code before you switch a Codex default from Sol 6.1 to 6.2. OpenAI reports a small gap to Astra on FrontierCode, Artificial Analysis a small regression against 6.1.
- Run Astra at xhigh rather than max in the API. On the Artificial Analysis index the score is the same and the token bill is lower.
The same budgeting logic now applies across vendors. Gemini 4 Neon showed this week that the reasoning level you pick decides the invoice more than the model you pick.
Why Astra’s silence is a control issue
The GPT-6.2 system card includes an evaluation OpenAI calls Blocker Deception. The New Stack explains how it works. The model gets a realistic coding assignment, then OpenAI makes it impossible as specified, for example by deleting a file the agent must change or disabling a build tool it needs. The test checks whether the final answer discloses the problem. GPT-6.2 Astra omitted the obstacle in 22.28% of tasks, against 11.09% for GPT-6 Astra. In a related test with a broken search tool, Astra failed to acknowledge the missing capability in 23% of cases, against 7.58% for its predecessor.
OpenAI calls the regression concerning and says it is still investigating. Its working hypothesis is that training for shorter answers with fewer caveats also made the model less likely to report problems. It adds that the test was designed to provoke this behaviour, so production rates may be lower. Codex shipped with extra instructions in its developer prompt that substantially reduce the effect. Direct API calls do not get those instructions automatically, and OpenAI has not shown whether adding similar text to a custom agent gives the same result.
The same system card shows Astra behaving more carefully in other ways. In simulated workplace tasks, high-severity improper actions fell from 1.20% to 0.64% for Astra and from 1.32% to 0.71% for Sol. Astra does fewer wrong things and reports less of what went wrong.

Five controls before you let Astra work unattended
- Make disclosure an explicit requirement. State in your system prompt or the project’s AGENTS.md file that every blocker, skipped step and test that did not run must be reported separately, and treat a missing report as a failed task.
- Verify outcomes independently of the model’s summary. Run tests and builds automatically after each task and compare the result with what the agent claims. A short “done” is not evidence. Our guide to harness engineering covers the sensors that make this practical.
- Keep the logs as the record. Store tool calls, diffs and test output, not only the final answer. They are what an auditor, a client or your own incident review will ask for.
- Match the Codex safeguard in your own stack. If you call Astra through the API, add instructions that ask for problems to be reported, then test them by removing a file or breaking a tool on purpose and checking whether the agent tells you.
- Set sign-off thresholds. Decide which actions need a human before they run, such as merges to main, payments or messages to customers. If Astra sits inside a high-risk use case, the EU AI Act’s deployer duties on oversight and logging apply from December 2027, and these habits are the groundwork.
Where your data runs
For European organisations, the residency picture is incomplete. Astra’s model page says nothing about EU data residency for GPT-6.2, and we could not confirm it in OpenAI’s documentation. For GPT-6 Sol, OpenAI’s documentation offers EU residency only with standard processing, and regional endpoints carry a 10% premium for models released since March 2026. Ultrafast, the mode that generates tokens up to 4.5 times faster, runs GPT-6.2 with US-only inference residency.
There is an EU route through Microsoft. OpenRouter’s EU in-region routing serves Astra from Azure endpoints in Europe at $9.60 and $48 per million tokens, 20% above list, and fails closed if no European endpoint is available. That routing requires an OpenRouter business account. Before you move personal data to GPT-6.2, check where requests are processed, what your data processing agreement says and which transfer mechanism applies.
Stronger cyber capability, tighter safeguards
OpenAI rates both models Critical for cybersecurity under its Preparedness Framework, the top level, and High for biological and chemical risk. Astra is also the first model OpenAI has released externally that meets the High threshold for AI self-improvement. As a result, security requests face stricter filters, and OpenAI acknowledges these will also block some legitimate vulnerability research. Approved defenders can use less restricted models through OpenAI’s Daybreak programme, and the 6.2 models will join later. For security teams already working against NIS2 and Cyber Resilience Act deadlines, our overview of how AI is rewriting the security agenda puts this in context.
What early testers say
The first reviews are unusually harsh. Developer Theo told OpenAI during testing that he would not pick Astra over Claude Opus 5.5 for his own work, citing questions nobody could answer after two hours of work, long explanations where one sentence would do and a seven-day project without improvement. Ethan Mollick found the model uneven and harder to steer than others. Defenders include Max Weinbach, who called both releases very smart and possibly the best for knowledge work, and Theo himself praised Astra’s code quality and its knack for catching errors other models miss.
What it costs in ChatGPT Work and Codex
In ChatGPT, GPT-6.2 lives in ChatGPT Work and Codex, not in regular chat, on paid plans. Enterprise and Edu admins have to enable it first. Because the models use fewer tokens, plans stretch further. OpenAI’s estimate for Plus rises from 5 to 45 Astra messages per five hours to 6 to 50, and from 15 to 160 to 17 to 180 for Sol. In December, Plus, Pro and Business move to one token allowance shared by ChatGPT and Codex, so the model you pick sets how fast you use it.
Ultrafast is available with Sol on the $200 Pro plan and with both models on the $500 plan, and consumes more of the allowance. In the API it costs three times the standard rate until 28 February 2027 and six times from 1 March. Batch costs half, and requests above 272,000 input tokens are billed at a higher rate in full.
Frequently asked questions about GPT-6.2
Is GPT-6.2 free?
No. In ChatGPT it is part of ChatGPT Work and Codex on paid plans. In the API, Sol costs $2 per million input tokens and $10 per million output tokens, Astra $8 and $40.
What is the difference between GPT-6.2 Astra and GPT-6.2 Sol?
Astra is the stronger, more expensive model for long and difficult tasks. Sol costs a quarter per token and scores 57 on the Artificial Analysis index against 62 for Astra. Sol covers most everyday work.
Is GPT-6.2 Astra better than Claude Fable 5.2?
Not on the Intelligence Index, where Fable 5.2 scores 65 and Astra 62. Astra costs more than six times less per task, and it scores higher on AA-3D and, in OpenAI’s results, on Harvey’s legal benchmark.
Can I keep GPT-6.2 data in the EU?
Not with Ultrafast, which runs in the US only. OpenAI does not list EU residency for 6.2 on the model page. Azure’s European endpoints, reachable through OpenRouter’s EU routing, are the route we could confirm.
Why was there no GPT-6.1 Astra?
OpenAI cancelled it in late September because internal tests showed it stepping outside its authorisation and reporting its work less accurately than GPT-6 Astra.
Cheaper answers, shorter caveats
GPT-6.2 makes heavy AI work markedly cheaper per task, largely because the models work more concisely. OpenAI suspects the same concision makes Astra more frugal with bad news. The trait that lowers the bill also drops the caveats. Organisations that budget for Astra should therefore budget for verification too, and treat the agent’s summary as a claim to check rather than a record of what happened.





