GPT-6 Astra through the API: setup and cost

Update, October 2026. GPT-6 Astra is currently unavailable in Zukko. The current model list is on the Models page↗ or at GET https://api.zukko.pro/v1/models. Closest replacement: gpt-5.6-sol.

Reviews of new models appear quickly, and nearly all of them cover the same set of public benchmarks. Something else is more useful: how to call the model in your own setup, and how to tell whether it suits your work.

Connecting

A model is connected through the same key as the rest of the catalogue. One field changes — the identifier (the example already uses Astra's replacement, gpt-5.6-sol):

curl https://api.zukko.pro/v1/chat/completions \
  -H "Authorization: Bearer fetch_your_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "messages": [{"role": "user", "content": "hello"}]
  }'

Clients work the same way: in Cursor the model is added to the list by identifier, in your own code it goes into the model field, in Claude Code it is assigned to a slot in settings.

The base URL for OpenAI-compatible clients is https://api.zukko.pro/v1; Anthropic-compatible ones use the same host without the suffix.

Cost

While Astra was in the catalogue it cost $0.8 per million tokens on input and the same on output. The context window is one million tokens.

The rate is flat: input and output cost the same. That noticeably changes the arithmetic for agent workloads, where input tokens outnumber output several times over — there the input price decides.

For reference, in the current catalogue claude-opus-5-5 and gpt-5.6-sol cost $0.3 per million and glm-5.3-flash $0.04. Astra was clearly more expensive, and that is the first thing to weigh for any model: the price difference has to pay for itself on your task.

Evaluate rather than read an evaluation

Public benchmarks measure an average over someone else's work. Checking on your own takes half an hour and answers about you.

Take five tasks from the past week — ones you have already solved and know the right answer to. Run each on the candidate model (for example gpt-5.6-sol) and on the model you use now.

Compare three things. How many attempts it took to reach a usable result. How many tokens went in total — the number comes back in the usage field of every response. And where exactly the answers diverged: if the divergence clusters on one type of task, that is already a usable area of application.

The outcome often does not favour the more expensive model: it may win on two tasks out of five while the other three show no difference — in which case it makes more sense to keep it for that specific class of work rather than for everything.

Switching is cheap

Every model in the catalogue is available through one key and one balance, so the experiment needs no new accounts or separate invoices. The current model list is at GET /v1/models; prices and context are at GET /v1/models/info (or the Models page in the dashboard):

curl https://api.zukko.pro/v1/models
curl https://api.zukko.pro/v1/models/info

If the model does not work out, going back is a one-field edit. That is the main argument for testing rather than reading reviews: the cost of being wrong is close to zero.