GPT-6 Astra API: first request and migration from GPT-5.6 Sol

For a first BetterToken smoke test, change the Model ID to gpt-6-astra and send a short request to the OpenAI-compatible Chat Completions endpoint. Migrating a working application needs more than a model-name swap: GPT-6 Astra does not support reasoning effort none, several sampling and logprobs parameters must be removed, and tool calling must use the Responses API.

The sequence below starts with a simple text request and then applies the smallest controlled diff from gpt-5.6-sol. It is a documented path; we did not make a paid request with a real API key while preparing this article.

Prerequisites

As of September 6, 2026, the public BetterToken catalog lists both exact models, gpt-5.6-sol and gpt-6-astra, in the GPT group with an OpenAI-compatible endpoint. You need your own BetterToken account, an API key created for the current group, Base URL https://www.bettertoken.ai/v1, and an HTTP client that can send a Bearer token and JSON.

Create a BetterToken account and API key for GPT-6 Astra

Check the current Model ID and access conditions in the catalog and your key’s Setup panel. Never paste a real key into repository code, an issue, a screenshot, or a support message.

Send the first request

Export the key only in your local environment:

export BETTERTOKEN_API_KEY="YOUR_API_KEY"

Use the endpoint documented in the public BetterToken API reference:

curl "https://www.bettertoken.ai/v1/chat/completions" \
  -H "Authorization: Bearer $BETTERTOKEN_API_KEY" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "gpt-6-astra",
    "messages": [
      {
        "role": "user",
        "content": "Reply with exactly: ASTRA_OK"
      }
    ]
  }'

A successful response from this endpoint contains text in choices[0].message.content; the fixture expects ASTRA_OK. Verify three observable signals:

1. The request finishes without 401, 403, or a model-not-found error.
2. choices[0].message.content is non-empty and follows the short instruction.
3. The BetterToken Dashboard records gpt-6-astra, request status, and input/output/cache token usage.

This confirms the key, endpoint, Model ID, and basic response shape. It does not confirm tool calling, your SDK’s streaming adapter, production quality, or behavioral parity with Sol.

Minimum diff from GPT-5.6 Sol

A text request without tools changes as follows:

 {
-  "model": "gpt-5.6-sol",
+  "model": "gpt-6-astra",
   "messages": [
     {"role": "user", "content": "Summarize this incident report."}
   ]
 }

If the old request explicitly used reasoning effort none or minimal, start Astra at low. In Chat Completions the field is reasoning_effort:

{
  "model": "gpt-6-astra",
  "reasoning_effort": "low",
  "messages": [
    {
      "role": "user",
      "content": "Summarize this incident report. Preserve dates and owners."
    }
  ]
}

If Sol already used another effective reasoning level, OpenAI recommends preserving it for the first comparison. Do not raise effort across every request automatically: it changes latency and usage as well as complex-task quality.

Remove unsupported parameters from the Astra request:

 {
   "model": "gpt-6-astra",
   "reasoning_effort": "low",
-  "temperature": 0.2,
-  "top_p": 0.95,
-  "top_logprobs": 5,
-  "logprobs": true,
   "messages": [
     {"role": "user", "content": "Summarize this incident report."}
   ]
 }

OpenAI lists temperature, top_p, and top_logprobs as unsupported for Astra; in Chat Completions, remove logprobs as well. If your SDK supplies them by default, inspect the serialized body rather than only the configuration object.

Chat Completions for smoke tests, Responses for tools

GPT-6 Astra supports both endpoints, but OpenAI requires Responses for tool calling. The public general BetterToken API reference currently documents Chat Completions only; Responses is documented separately for the BetterToken Codex custom-provider setup.

Goal	Path
Verify key, Model ID, and an ordinary text response	POST https://www.bettertoken.ai/v1/chat/completions using the public API reference
Use Astra with tools in Codex	custom provider with wire_api = "responses" from the BetterToken Codex guide
Call raw Responses from your own app	Confirm the current contract for your channel in Docs or support first; the public general reference does not publish it yet

Do not move a Chat Completions tool schema to Responses by changing only the URL. Input items, tool-call formats, state continuation, and text extraction differ. Transport compatibility also does not imply identical style or tool choice between Astra and Sol.

Controlled migration with rollback

Build a small fixture set from the real application:

• one short request with an exact expected answer;
• one structured output validated by JSON Schema;
• one long prompt typical of production;
• one tool workflow, if the application uses tools;
• one intentionally invalid request to compare error handling.

For Sol, record the Model ID, endpoint, effective reasoning effort, serialized request body, and acceptance check. Run the same set on Astra with the minimum diff. Compare observable conditions: valid schema, preserved mandatory facts, correct tool call, no timeout, and existing tests passing.

Move traffic only when every required fixture passes, monitoring distinguishes gpt-5.6-sol from gpt-6-astra, and rollback can restore the old Model ID and parameters without a new code release. Start with one internal workflow or a small controlled share of traffic; there is no universal percentage. Keep the Sol configuration until observation is complete.

Fast troubleshooting

• 401 Unauthorized: verify Authorization: Bearer ..., the environment inherited by the process, and copied whitespace. Do not log the full key.
• 403 or model not found: confirm the current group and copy exactly gpt-6-astra from Setup or the catalog.
• 400 after the swap: inspect the serialized body; remove temperature, top_p, top_logprobs, and logprobs, then replace none or minimal with low.
• Text works but tools fail: check the endpoint. A Chat Completions smoke test does not prove a Responses tool workflow. Keep wire_api = "responses" in Codex and use a confirmed Responses contract in your app.
• The request succeeds but the UI is empty: save a secret-free raw response and inspect choices[0].message.content, status, finish reason, and the adapter after HTTP.

The migration is complete when Astra passes your fixtures, can be observed separately, and can be rolled back through configuration. The first 200 is a useful connectivity result, but it is not the whole migration.