Migrating from GPT-5.6 to Astra
four things that trip you up
OpenAI's official migration guide lists four hard constraints for gpt-6-astra: remove temperature, top_p and top_logprobs (plus logprobs on Chat Completions); the none reasoning tier is unsupported, so anything on none or minimal starts at low; tool calling requires the Responses endpoint; and the 128,000 output ceiling includes reasoning tokens. The third one is the one people mistake for a bug in their own code.
Updated 2026-09-09
The four hard constraints
Must be removed
temperature, top_p, top_logprobs — and logprobs as well if you are on Chat Completions. Straight from the official migration guide.
Five reasoning levels only
low, medium, high, xhigh, max. The guide says anything previously on none or minimal should start at low and be compared.
The only endpoint for tool calling
Official wording: Astra supports Chat Completions, but tool calling requires Responses. It will converse; it will not call tools.
Output ceiling includes reasoning
Reasoning tokens count toward output, so sizing max_tokens by the visible reply truncates early.
The one that looks like your bug
Most migration problems present as a fault in your own code: the same function-calling path works on GPT-5.6, and on gpt-6-astra the tools simply are not invoked, with no obvious error. The official migration guide is direct about this — Astra supports Chat Completions, but tool calling requires the Responses endpoint. On Chat Completions it holds a normal conversation and just never calls a tool. Check the endpoint before you debug your schema.
Why the sampling parameters have to go
The migration guide requires removing temperature, top_p and top_logprobs, plus logprobs if you use Chat Completions. Those are usually set by default in a team's wrapper layer, so the cheapest fix is to filter by model name on the way out rather than editing every call site. Related: if your wrapper defaults reasoning effort to none or minimal, handle that in the same pass — Astra does not have that tier.
Timeline
GPT-6 Astra ships with API model id gpt-6-astra: 1,050,000 context, 922,000 max input, 128,000 max output, knowledge cutoff 2026-04-30.
The official migration guide lands alongside it: drop the sampling parameters, no none tier, tool calling on Responses. Supported endpoints are Chat Completions, Responses and Batch.
This page was checked line by line against the official model documentation and migration guide; no second-hand summary of the parameters was used.
What is written and what is not
✅ Straight from the official guide
Remove temperature, top_p and top_logprobs (plus logprobs on Chat Completions); none is unsupported, so start at low and compare; Astra supports Chat Completions but tool calling requires Responses; reasoning levels are low, medium, high, xhigh, max; supported endpoints are Chat Completions, Responses and Batch; context 1,050,000, max input 922,000, max output 128,000.
⚠️ Not in the official docs
The parameter count, a SWE-bench Verified score, and whether long context (input above 272,000) reprices the whole request or only the overage are all absent from the official documentation. The third one materially affects long-context bills, so measure with a small request before scaling. Numbers circulating online for these three are not official and are not used on this page.
Two migration paths
Filter in the wrapper layer
Strip unsupported parameters by model name on the way out and map none/minimal to low. One change, every call site benefits, and rolling back is one change too. Best when call sites are scattered.
Edit each call site
Precise but slow, and it tends to miss wrapper defaults. Only worth it when there are few call sites, or when different sites genuinely need different handling.
A sensible order
Move the tool-calling path to the Responses endpoint first and get it working, because it is the only one that fails silently. Then strip the four sampling parameters in the wrapper layer and map none/minimal to low. Next re-check max_tokens against the new accounting — 128,000 includes reasoning tokens. Finally run the same real requests against gpt-5.6-sol and gpt-6-astra and let the quality and cost difference decide how much you move.
On QCode
gpt-6-astra is in QCode's model table and has real traffic over the last 30 days. It shares one key and one quota with the three GPT-5.6 tiers, Claude and Gemini, which makes the migration comparison easy: same endpoint, same key, change the model field and run the same batch of requests against both. The cost and quality difference shows up the same day.
Frequently asked questions
Why doesn't my function calling work on Astra?
Usually the endpoint. The official migration guide states that Astra supports Chat Completions, but tool calling requires Responses. On Chat Completions it converses normally and simply never calls a tool. Move that path to Responses before debugging your schema.
Which parameters have to be removed?
temperature, top_p and top_logprobs — plus logprobs if you are on Chat Completions. Those four are what the official guide lists.
I was using reasoning effort = none. Now what?
Astra does not support none. The guide advises starting at low and comparing. Note that reasoning tokens count toward output, so a higher level costs more on the output side.
How should I set max_tokens?
Remember the 128,000 output ceiling includes reasoning tokens. If you size it by the visible reply, the model can spend the allowance during reasoning and truncate early. Re-check that ceiling against the new accounting when you migrate.
How much context can I use?
The official model documentation gives a 1,050,000 context window with 922,000 max input and 128,000 max output. Note that input above 272,000 moves the API to a more expensive tier, and the docs do not state whether that reprices the whole request or only the overage.
What is the low-risk way to switch?
On QCode, send the same batch of real requests to gpt-5.6-sol and gpt-6-astra — one key, one quota, only the model field changes. Decide whether the quality difference is worth the cost difference before choosing between routing by task and moving everything.
Sources
OpenAI's official migration guide (sampling parameters, reasoning levels, tool-calling endpoint) and the official model documentation (context, output ceiling, supported endpoints, knowledge cutoff), both captured 2026-09-09. Items the docs do not state are marked unpublished here rather than filled in from third parties. QCode call volumes come from this site's own 30-day usage statistics.
One key, run both models on the same batch
gpt-6-astra shares one QCode key and one quota with the three GPT-5.6 tiers, Claude and Gemini; a migration comparison is a model-field change. Plans start at ¥60/month and the key is live the moment you pay.
Related reading
GPT-6 Astra complete guide
Specs, pricing, and the four things OpenAI still has not published.
Astra quota consumption
The official credits table: exactly 2.5× GPT-5.6 Sol.
Enabling GPT-5.6 in Codex
How the previous generation was enabled and switched — useful as a migration reference.
Every constraint on this page comes from OpenAI's official migration guide and model documentation, captured 2026-09-09; OpenAI may update them, so defer to the official pages. Items the docs do not publish are not filled in from third-party numbers.