GPT-6.2 Sol is the first model to test for routine repository work. GPT-6.2 Astra is the escalation candidate for ambiguous changes or work that keeps coming back from review. OpenAI released both on November 17, 2026. Astra scored higher in Artificial Analysis’s general evaluation, while Sol runs at a quarter of Astra’s standard API token rates.
The choice depends on how often each model produces a patch you can merge and how much time it takes to review. The benchmark results below come from the named evaluators. The cost example uses hypothetical token counts and review time.
Key takeaways
- Standard API input and output rates are $2 and $10 per million tokens for Sol, and $8 and $40 for Astra.
- Artificial Analysis scored Astra 62 and Sol 57 on its Intelligence Index v5.0.1 at top effort. Its Codex evaluation found weaker results for Sol on some coding work than for its predecessor.
- Measure accepted patches, missed requirements and reviewer time before changing the default model.
GPT-6.2 model IDs, access and API pricing
OpenAI’s release lists both models for ChatGPT Work, Codex and the API. Availability and usage limits inside the apps depend on the plan and workspace. Custom integrations should confirm billing and access in the API project that will serve the workload.
The table lists standard short-context rates. OpenAI’s model documentation says that requests above 272,000 input tokens use higher rates across the full request. Cache writes, processing tiers and tool calls can also change the bill. Verify the actual model ID and capture billed usage on every attempt.
| Model ID | Uncached input | Cached input reads | Output |
|---|---|---|---|
| gpt-6.2-astra | $8.00 | $0.40 | $40.00 |
| gpt-6.2-sol | $2.00 | $0.10 | $10.00 |
OpenAI launch announcement and access
What the GPT-6.2 benchmarks say about coding
Artificial Analysis measured a five-point gap between the models on its Intelligence Index at their top effort settings. That index combines several evaluations. On its separate Coding Agent Index, run through Codex at max effort, Astra stayed level with GPT-6 Astra and Sol dropped three points from GPT-6.1 Sol. The evaluator attributed much of Sol’s drop to omitted facts in a repository question-answering test.
OpenAI’s announcement quotes a different result from Cognition’s FrontierCode 2.0. It reports that Sol finished within half a point of Astra at roughly one-fifth of the cost per task, both at max effort. Different test sets and harnesses can reward different capabilities. Test repository explanations and patch implementation as separate task types so a strong patch score can’t hide missing details in a handoff.
| Evaluation | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
| Artificial Analysis Intelligence Index v5.0.1, top effort | 62 | 57 |
| Artificial Analysis Coding Agent Index, Codex at max | Level with GPT-6 Astra | 3 points below GPT-6.1 Sol |
| Cognition FrontierCode 2.0, max, as reported by OpenAI | Reference score for this comparison | Within 0.5 points of Astra at about 20% of the task cost |
How much review does each model need?
Bring a few representative tickets, your current review process and the checks a patch must pass. We can help scope a paired evaluation.
Worked example for a repository review budget
Suppose a single patch attempt uses 20,000 uncached input tokens and 5,000 billed output tokens. At standard short-context rates, Sol costs $0.09 and Astra $0.36. That is $0.27 in API savings per Sol attempt. At an illustrative reviewer cost of $60 per hour, about 16 seconds of extra review costs the same amount. The example excludes tools, caching and repeated attempts.
For a real pilot, add token and tool charges across all attempts and the cost of human review. Divide by the number of accepted patches, including the spend on issues that remained unresolved in the numerator. Record time to an accepted patch separately from time to the first generated patch.
Run a paired test before switching models
Start with historical tickets that your team can grade without knowing which model produced the result. Use the same base commit, instructions, tool permissions and acceptance tests for each candidate. Record the model and effort in every run. If you change settings or the harness in a second round, keep those results in their own group.
- Test several task types separately, such as a contained bug fix, a change across several modules and a repository question with a known answer.
- Score the patch against acceptance tests and the review against the issue brief. Log out-of-scope edits, unsupported claims and work the model silently skipped.
- Agree on correctness and review-time limits before reading results. Start with Sol on task types that meet both limits. Send failed or ambiguous cases to Astra for a bounded second attempt, with a human deciding what merges.
Frequently asked questions
Is GPT-6.2 Astra always better for coding?
Use the workload to decide. The general index favors Astra, while the reported gap on FrontierCode 2.0 was small at max effort. Run both models on your own accepted patches and grade explanations separately from code changes.
How much cheaper is GPT-6.2 Sol?
Its standard short-context API input and output rates are each one quarter of Astra’s. Spend per accepted change also depends on billed token counts, retries, tool charges and review.
Let’s work through your next step.
Tell us what you’re building, what you’ve tried, and where you need a hand. We’ll work with you to define a practical way forward.

