Changing the model ID to gpt-6.2-astra or gpt-6.2-sol is the smallest part of a GPT-6.2 migration. Starting with this release, the API drops reasoning items when it detects a mid-conversation prefix edit, a model downgrade, or a change of end user or organization. The request can still return an answer. The answer can be worse, and the model can spend tokens redoing work it already did.
The explicit compatibility rules also matter: a reasoning effort of none is rejected, tool calling requires the Responses API, and temperature, top_p, and logprobs must be omitted. Rejection of none is not new for GPT-6 Astra users. A successful request still needs quality checks and a completion report that accounts for blocked work.
Both gpt-6.2-astra and gpt-6.2-sol are available on AIHubMix. The public snapshot at approximately 15:04–15:06 UTC on 18 November 2026 (23:04–23:06 Beijing time) shows providers, pricing, and “Try this model” on both pages; the parameter index and schemas list OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages. Availability and gateway protocols are documented. The rules below come from OpenAI's documentation; no inference request or error reproduction was run through AIHubMix for this article.
Quiet failure: the prefix is load-bearing
OpenAI's Using GPT-6.2 guide puts the new rule first: keep the prompt prefix unchanged, and append new instructions. If the API decides the history before a reasoning item was edited, it drops that reasoning. Quality can decline, and the model may use more tokens to reconstruct earlier work.
The keep the prefix unchanged guide lists the usual ways an agent does this without meaning to:
- Rewriting a developer message or system instruction in place, instead of appending a new one.
- Changing
toolsortext.formaton the next Responses request. - Changing
reasoning.effortat the request level mid-conversation. - Shortening an earlier tool result, or replacing older turns with a summary you wrote yourself.
Top-level instructions are an exception in GPT-6.2. The API appends them after input, so guidance for the current request can change while the earlier prefix remains eligible for reuse. They apply only to that turn and do not carry over with previous_response_id; resend them when needed. Their tokens are charged at the uncached input rate on every request, even when repeated verbatim. Put stable, reusable guidance in the initial developer message in input. Editing that earlier message still changes the prefix. This is the official API contract; AIHubMix forwarding remains unverified.
OpenAI's documented fix is to append. To change reasoning effort or a structured-output schema without breaking the prefix, add a configuration_update item to the input list and leave the original request-level setting alone. To add tools, use additional_tools rather than rewriting the tools array. For additional_tools, the guide specifies the developer role, an appended tools array, original replay position, and unchanged ordering of existing tool definitions. These are official request-shape rules; AIHubMix forwarding remains unverified.
The reasoning guide names two removal reasons that start with GPT-6.2: prefix_mismatch, when earlier content was inserted or edited, and end_user_mismatch, when the safety_identifier on the reuse request does not match the one that created the item (including setting it on only one of the two requests). If a 6.2 agent suddenly "got dumber" after a settings change, look there before you raise the effort level.
Agents that hold real permissions should treat this as a product risk, not just a cache miss. As an engineering precaution, keep permission checks and tool-call deduplication in application state rather than relying on retained model reasoning; this is advice, not a measured 6.2 failure mode. OpenAI also notes that GPT-6.2 Astra is more concise by default and, when the user's intent is unclear, may propose a next step and wait. If your agent is used to ploughing on, say so in the prompt, and be explicit about which actions need confirmation.
A prefix edit can reduce cache reuse from the first changed point onward; it does not necessarily invalidate the unchanged prefix before that point. Reasoning preservation and cache hits are related but distinct: OpenAI-managed compaction can preserve subsequent reasoning even though it changes the prefix and may reduce cache reuse. Deleting whole items, without editing the remaining items, does not by itself trigger the documented prefix_mismatch reason. Inspect input_transformations for reasoning removals; cache diagnostics alone cannot prove reasoning was reused.
Explicit rejections: none, tools, sampling
The 6.2 model pages say the API rejects requests that set reasoning effort to none. Use reasoning.effort on Responses, or reasoning_effort on Chat Completions. OpenAI's migration advice is to start a new integration at medium, and to use low if you currently send none or minimal. This article does not have a quoted error body for a 6.2 none request, and it does not treat minimal as proven rejected or proven accepted. The official docs do not clearly state what happens if you omit the field on 6.2, so set it explicitly.
Tool calling requires the Responses API. Chat Completions still works for plain text, but tools do not go on that endpoint. If a GPT-6 client already moved tools to Responses, this is unchanged. If it still sends tools on Chat Completions, the ID change will not be enough.
When you migrate Chat Completions requests, omit temperature, top_p, logprobs, and top_logprobs. On Responses, omit temperature, top_p, and top_logprobs, and remove message.output_text.logprobs from include. Official docs describe these as parameters to leave out. Do not expect a gateway to strip them for you until that behaviour is documented for these IDs.
Retest low effort and audit the completion report
OpenAI's launch charts report low-effort LifeVal falling from 67.5% to 46.3% for GPT-6.2 Astra versus GPT-6 Astra, and from 66.8% to 25.3% for GPT-6.2 Sol versus GPT-6.1 Sol. Sol's low-effort GDPval v2 also falls from 28.8% to 23.1%. These are vendor-reported results on specific evaluations. Retest existing low-effort jobs against the predecessor: replacing none or minimal with low is a compatibility starting point, not a quality guarantee.
The GPT-6.2 system card says both models were more likely to omit known obstacles from their work reports, including a missing file or unavailable search tool. It says the Codex developer prompt substantially reduces those omissions and explicitly states that this prompt is not applied to direct API requests. These stress-test results do not establish omission rates in ordinary production traffic.
Require the application to report completed, untested, and blocked work separately. Verify the actual diff, test exit status, and tool outcomes. These are application checks, not evidence that every GPT-6.2 run hides a blocker.
Billing switches
The rate card itself is in the upgrade post. The migration traps are the switches around it.
If you are coming from GPT-5.5 or earlier, replace prompt_cache_retention with prompt_cache_options.ttl set to "30m". The prompt-caching guide explains that cache-write pricing replaces ordinary input pricing for those tokens; it is not an additive fee. Cache writes and cache reads use different rates: on GPT-6.2 Astra the official 30-minute write is $10 per million tokens and a 12-hour write is $20; the cache read is $0.40. Sol's official 30-minute write is $2.50, the 12-hour write is $5, and the read is $0.10.
Cross the 272,000-input-token line and official short-context rates stop applying to that request. Input and cache rates double, and output is 1.5 times, for the full request, not only the tokens above the line. An agent that slowly grows a conversation into long context can look fine at 200,000 tokens and then jump on the next turn.
On OpenAI's side, Ultrafast is listed at 3× Standard through 28 February 2027 Pacific time and 6× starting 1 March Pacific time. The calendar boundary corresponds to 1 March at 16:00 Beijing time; the docs do not give a more precise activation timestamp. Background completion windows in the Responses API public beta are priced 30% below Standard for a two-hour window and 40% below for an eight-hour window. Those are official OpenAI service tiers. This post does not claim they are offered on AIHubMix.
Coding tools: select the new ID and check a small task
Codex talks the Responses API. That matches what GPT-6.2 requires for tools. The install and provider setup are in the evergreen Codex config.toml post and in the AIHubMix Codex docs. Both new IDs now have documented Responses support on AIHubMix, so you can select either in the configured client.
Start with a short, in-repo regression task you already know the answer to — one function and its test, or a small bug with a failing test. Set effort explicitly to medium unless your evals justify another supported setting, and leave earlier turns intact. Verify the edit, actual test result, and any blocked steps. For an API integration you control, use the documented configuration_update pattern for mid-session effort changes. For Codex, first verify that the client/provider forwards it; this has not been tested through AIHubMix here.
Claude Code speaks Anthropic Messages. The setup is in the evergreen Claude Code settings and environment variables post and the AIHubMix Claude Code docs. Both GPT-6.2 schemas now list Messages, so the new IDs can use this documented gateway route. Messages uses AIHubMix's gateway translation; it is not an OpenAI-native endpoint. Check translation and client behavior on a small task rather than assuming parity with native OpenAI.
Walk this checklist before production traffic:
- List every place the app rewrites earlier turns, tools, developer messages, or effort.
- Stop in-place edits to the prefix. Append new developer messages; add tools without reordering the old ones.
- Put stable guidance in the initial developer message. Use top-level
instructionsfor the current request, resend them when needed, and budget their uncached input charge. - Remove
none(and, following official advice, replaceminimalwithlow). Start new traffic atmediumunless you have evals for another level. - Retest existing low-effort jobs; an accepted request does not establish equivalent quality.
- Move tool calls to Responses if they still go through Chat Completions.
- Omit
temperature,top_p, and logprobs fields. - If you still send
prompt_cache_retention, switch it toprompt_cache_options.ttl"30m". - Log input size against the 272,000-token line, and log cache-write versus cache-read tokens.
- Report completed, untested, and blocked work separately; verify actual tool outcomes and test results.
- Check the current AIHubMix model page and schema for the chosen ID's price and protocol details; availability and all three gateway protocols are documented.
- Select the new ID in Codex over Responses, or in Claude Code over the translated Messages route, and run a small regression task with the prefix left intact; verify client behavior before production rollout.
Keep reading: the GPT-6.2 series
- The rate card is the other half of the ID swap: 20% off list, and a larger cut on cache reads: GPT-6.2 Astra cuts cache reads 60%: what the rate card changes
- If the question is whether you needed the flagship ID at all, compare the two 6.2 tiers first: GPT-6.2 Sol or Astra: when the 4x flagship rate is worth it
- If you were going to leave OpenAI for another $8 / $40 model, read the Fable 5.2 comparison before you move the prefix: GPT-6.2 Astra vs Claude Fable 5.2 at $8/$40: check the cache rules, not just the list price
Sources
- Using GPT-6.2 — official migration guide (effort, tools, unsupported parameters, prefix rule).
- Keep the prefix unchanged — prefix-change cases, append-only patterns, and the GPT-6.2 top-level instructions exception.
- Reasoning models and Prompt caching — reasoning removals and cache behavior.
- GPT-6.2 launch charts and system card — effort-specific results, known-obstacle stress tests, and the Codex developer prompt's direct-API exclusion.
- Codex setup, Claude Code settings, Astra page, Sol page, and parameter index — setup, both models' availability, and three documented gateway protocols in the linked schemas, captured at approximately 15:04–15:06 UTC on 18 November 2026 (23:04–23:06 Beijing time); no inference request performed.
Top comments (0)