GPT 6 Astra Stable: 1,050,000 tokens, on an Anthropic-shaped route

The largest window here, and what it costs

stable-gpt-6-astra takes 1,050,000 tokens of context. That is the largest figure in the catalog — the several Claude ids at 1,000,000 are the closest, and the rest of the OpenAI text lineup tops out at 372,000. Output is capped at 128,000 tokens, matching the GPT 5.6 family. It supports text, tool calling and thinking, and was released 2026-09-05.

It lists at $10 per million input tokens and $50 output, billed here at a 0.2× multiplier — $2.00 and $10.00 per million in practice. That is 4× the effective input rate of gpt-5.6-sol and 3.33× its output rate, so the extra 678,000 tokens of room are not free.

The other figure worth knowing before you plan around the window: cached input reads list at $1 per million against $10 for fresh input. Where a long prefix genuinely repeats across turns, that is the lever that keeps a large window from being re-billed at full rate on every request.

It is an OpenAI model on an Anthropic-shaped route

Every other OpenAI id in the catalog answers on /v1/responses. This one does not: its recorded apiRoute is /v1/messages, which means the Anthropic-shaped request — an x-api-key header rather than Authorization: Bearer, an anthropic-version header, a messages array, and a required max_tokens.

The practical consequence is that swapping stable-gpt-6-astra into code written against gpt-5.6-sol is not a one-field change. You are changing endpoint, auth header and body shape at once, which is a different kind of edit from the drop-in substitutions the rest of the OpenAI family allows among themselves.

If your client already speaks the Anthropic message format for Claude ids, the reverse is true and this id slots in with only the model field changed. Check the route column before you plan the integration rather than assuming it from the provider name.

curl https://token-share.app/v1/messages \
  -H "x-api-key: $TOKEN_SHARE_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "stable-gpt-6-astra",
    "max_tokens": 4096,
    "messages": [
      {"role": "user", "content": "Trace how the request id flows through these files."}
    ]
  }'

What a stable lane is, and what a million tokens is not

The stable- prefix marks a dedicated lane for the same upstream model, served on our own key rather than the shared account pool. It does not participate in pool failover. That is a routing property, not a quality claim: the weights are the same GPT 6 Astra either way, and the premium is on the path, not the model.

On the window itself: 1,050,000 tokens is room, not recall. The per-token rate is identical whether the window is nearly empty or nearly full, so pasting a whole repository into every request multiplies the bill without improving the answer. A full 1,000,000-token input at $2.00 per million is $2.00 of input on a single call — worth pricing out before it becomes a per-request habit.

The window earns its cost when the material genuinely has to be resident at once: a codebase whose call graph spans files you cannot pre-select, a long transcript where any earlier turn may matter. Where you can retrieve the relevant slice instead, a 372,000-token id at a quarter of the input rate does the same job for less.

Common questions