
If you already call GPT-6.1 Sol, GPT-6 Sol, or GPT-6 Astra through PoYo, moving to the GPT-6.2 API is mostly a model-ID change. gpt-6.2-sol replaces the Sol line at the same or lower price, and gpt-6.2-astra replaces GPT-6 Astra at 20% less per token. This post lists every request-level difference, what each change costs, and a rollout plan that catches regressions before users do. Model pages: GPT-6.2 Sol and GPT-6.2 Astra.
Migration paths
| You call today | Move to | Price change on PoYo |
|---|---|---|
gpt-6.1-sol | gpt-6.2-sol | None |
gpt-6-sol | gpt-6.2-sol | Cache read halves ($0.16 → $0.08 per 1M) |
gpt-6-astra | gpt-6.2-astra | 20% lower input, output, and cache write; 60% lower cache read |
gpt-6-astra, for cost | gpt-6.2-sol | 80% lower on every meter; test quality first |
GPT-6.2 Sol API vs GPT-6.1 Sol API, field by field
| Item | gpt-6.1-sol | gpt-6.2-sol |
|---|---|---|
| Context / max input / max output | 1,050,000 / 922,000 / 128,000 | Same |
| Input / output | Text and image in, text out | Same |
| Knowledge cutoff | April 30, 2026 | June 1, 2026 |
| Reasoning effort | low, medium (default), high, xhigh, max | none rejected; OpenAI says start at medium; levels not listed on the model page |
| Tool calling | Responses only | Responses only |
temperature, top_p | — | Omit (OpenAI GPT-6.2 guide) |
Top-level instructions | Standard placement | Appended after input |
| Prefix edits in Responses | Earlier behavior | Changed: some edits and model switches drop reasoning items |
| Cache read ratio | 0.05× input | 0.05× input |
| PoYo price per 1M (in / out / cache read / cache write) | $1.60 / $8 / $0.08 / $2 | $1.60 / $8 / $0.08 / $2 |
| OpenAI price per 1M | $2 / $10 / $0.10 / $2.50 | $2 / $10 / $0.10 / $2.50 |
Limits and reasoning settings for GPT-6.1 Sol come from its OpenAI model page; the GPT-6.2 notes come from OpenAI's GPT-6.2 guide and API changelog. Limits and price are unchanged. The work is in the three behavior changes: instructions placement, sampling fields, and reasoning items.
GPT-6.2 Astra API vs GPT-6 Astra API
| Item | gpt-6-astra | gpt-6.2-astra |
|---|---|---|
| Context / max input / max output | 1,050,000 / 922,000 / 128,000 | Same |
| Knowledge cutoff | April 30, 2026 | May 18, 2026 |
| Reasoning effort | low to max; none returns 400 | none rejected; start at medium |
| Cache read ratio | 0.1× input | 0.05× input |
| PoYo price per 1M (in / out / cache read / cache write) | $8 / $40 / $0.80 / $10 | $6.40 / $32 / $0.32 / $8 |
| OpenAI price per 1M | $10 / $50 / $1 / $12.50 | $8 / $40 / $0.40 / $10 |
What OpenAI measured between generations
OpenAI's GPT-6.2 system card reports these side by side with the previous models:
| Evaluation | GPT-6 Astra | GPT-6.2 Astra | GPT-6.1 Sol | GPT-6.2 Sol |
|---|---|---|---|---|
| Sandbox Bench (targets solved) | 18.2% | 36.4% | 11.4% | 25.0% |
| SEC-Bench Pro (vulnerability reproduction) | 85.4% | 86.75% | 78.8% | 88.34% |
| ExploitBench Internal Port (code execution rate) | 31.5% | 53.5% | 21.5% | 50.5% |
| MLE-Bench Revised (mean reference percentile) | 93.80 | 96.06 | — | 91.38 |
The system card does not report GPT-6.1 Sol on MLE-Bench Revised in the same table. These are safety evaluations focused on security and ML engineering, and most rows do not state the reasoning effort. They show a clear step up on long multistep tasks, not a general quality score for your workload. That still has to come from your own tests.
The JSON migration diff
A typical GPT-6.1 Sol request on Responses:
{
"model": "gpt-6.1-sol",
"instructions": "You are a support agent for Acme. Follow the policy below. ...",
"input": "Customer asks: can I change my billing date?",
"temperature": 0.3,
"reasoning": {"effort": "low"},
"max_output_tokens": 2000
}
The same request for GPT-6.2 Sol:
{
"model": "gpt-6.2-sol",
"input": [
{"role": "developer", "content": "You are a support agent for Acme. Follow the policy below. ..."},
{"role": "user", "content": "Customer asks: can I change my billing date?"}
],
"reasoning": {"effort": "low"},
"max_output_tokens": 2000
}
Three changes, each from OpenAI's GPT-6.2 guide:
modelbecomesgpt-6.2-sol(orgpt-6.2-astra).temperatureis removed. Same fortop_p, andtop_logprobson Responses (pluslogprobson Chat Completions). If you relied on low temperature for consistent output, use stricter instructions or a structured output format instead.instructionsmoves into adevelopermessage at the start ofinput. GPT-6.2 appends top-levelinstructionsafterinput, so a long policy there would sit after the user turn and fall outside the cacheable prefix.
If you sent none or minimal effort to an older model, change it to low. Chat Completions callers change model, drop the sampling fields, and keep the system message first. Tool-calling code must run on Responses. GPT-6.1 Sol code already does, because that model had the same rule. If you call tools on Chat Completions with another model today, move that code to Responses first.
Price per same request
Same token counts on each model, PoYo prices. Example A is a 150K-input / 12K-output code review. Example B is a 25-turn agent session, best-case caching (136K cache writes, 2.064M cache reads, 37.5K output); both are worked out in GPT-6.2 API pricing.
| Model ID | Example A | Example B |
|---|---|---|
gpt-6-astra | $1.68 | $4.511 |
gpt-6.2-astra | $1.344 | $2.948 |
gpt-6-sol | $0.336 | $0.902 |
gpt-6.1-sol | $0.336 | $0.737 |
gpt-6.2-sol | $0.336 | $0.737 |
Cached workloads gain the most from GPT-6.2 Astra, because its cache-read price drops 60%. Actual cost per task also depends on how many tokens each model spends, so compare usage on real traffic.
A rollout plan for the GPT-6.2 API
- Shadow test. Send a sample of real requests to both the old and new model IDs. Store both outputs and both
usageblocks. - Check the deltas. Compare pass rates on your validators,
usage.output_tokens, andusage.output_tokens_details.reasoning_tokens. A lower token count at the same pass rate is a direct saving. - Fix prompts that depended on temperature. These are the most likely regressions.
- Flip per request type, not all at once, and keep the old model ID in config for a quick rollback.
- Do not mix models inside one conversation. OpenAI warns that model changes, including downgrades, can drop reasoning items in Responses. Let in-flight conversations finish on the old model.
For the choice between the two new models, read GPT-6.2 Sol API vs GPT-6.2 Astra API. For GPT-6.1 Sol background, see GPT-6.1 Sol features and Claude Fable 5.2 vs GPT-6.1 Sol.
Frequently asked questions
Is the GPT-6.2 Sol API better than the GPT-6.1 Sol API?
OpenAI's system card shows GPT-6.2 Sol ahead of GPT-6.1 Sol on its multistep security evaluations, such as 25.0% vs 11.4% on Sandbox Bench. The price is the same, so the switch costs nothing per token. Test it on your own prompts.
What do I change in my code to use the GPT-6.2 API?
Change model, remove temperature and top_p, and move top-level instructions into a developer message at the start of input. Replace none or minimal effort with low.
Is the GPT-6.2 Astra API cheaper than the GPT-6 Astra API?
Yes. On PoYo it is $6.40 / $32 per 1M input/output tokens versus $8 / $40, and cache reads drop from $0.80 to $0.32.
Do GPT-6.2 API limits differ from GPT-6.1 Sol?
No. Both have a 1,050,000-token context, 922,000 max input, and 128,000 max output tokens, with text and image input.
Can I keep using GPT-6.1 Sol on PoYo?
Yes. gpt-6.1-sol remains listed on PoYo at the same price, so you can migrate one request type at a time.