GPT-6 Astra API Pricing: How to Calculate Input, Output and Cache Costs

GPT-6 Astra standard API pricing is $10 per million input tokens and $50 per million output tokens. ChatGPT Plus and Pro subscription fees are not API balances. Developers budgeting a task need to distinguish ordinary input, cache usage, output and service mode before multiplying a headline rate by token count.

Checked September 16, 2026. All prices are in US dollars. The examples are budget calculations; no paid generation was performed for this article.

Standard and cached token rates

The OpenAI model card lists these base rates per million tokens:

Do not count the same input fully as ordinary input and again as a cache read. Repeating a prompt does not establish that all input qualifies for the cached rate.

For input above 272K tokens, the model card applies 2× input and cache rates and 1.5× output rates to the entire request, not only the excess. Fast is 2× the applicable rate; Batch and Flex are 50% of Standard. These are OpenAI API rules. Confirm which modes a third party offers and how it bills them; do not apply the API Fast multiplier to Codex credits.

The context window is 1,050,000 tokens, with a maximum output of 128,000 tokens. Capacity and pricing thresholds are separate: fitting into the window does not guarantee short-context pricing.

Three reproducible request budgets

Multiply each token category by its applicable rate, then divide by one million. These examples include text tokens only, excluding tools such as search, additional requests, taxes and provider surcharges.

text cost = (ordinary input × input rate
           + cache reads × read rate
           + cache writes × write rate
           + output × output rate) / 1,000,000

The cache example covers the current request only; initial cache preparation is not included in $0.128. The long-context calculation uses the higher rates for the whole request. Keeping the first 272K at $10/M and increasing only the remaining 28K would be incorrect.

Under the first assumption, 1,000 requests cost $200 in text tokens. A user task requiring three such requests costs $0.60. That difference comes from request count, not a changed rate. Record model, mode, usage and billed cost per request, including retries.

Include the first cache write in a task budget

A cache-hit request is not the whole session bill. This illustrative three-request task costs $0.50, assuming Standard pricing below the long-context threshold and the exact usage categories shown below. It is arithmetic, not measured cache behavior or a promise of a cache hit.

The first request is (2,000 × $10 + 20,000 × $12.50 + 1,000 × $50) / 1,000,000 = $0.32. Each follow-up is (2,000 × $10 + 20,000 × $1 + 1,000 × $50) / 1,000,000 = $0.09. These input categories are disjoint: do not also charge the cached tokens as ordinary input. At the same token volumes without caching, three requests would cost 3 × (22,000 × $10 + 1,000 × $50) / 1,000,000 = $0.81.

Actual usage may differ. Reconcile cache categories against the provider’s usage record and invoice; include retries, tools and any applicable service-mode charges. The example excludes those extras and tax. For the client side, follow the Astra setup and migration checks.

How does Astra compare with GPT-5.6 Sol?

The September 16 Ofox public catalog lists these base rates:

Astra input is 2× Sol and output about 1.67×, so the entire bill does not have a fixed 2× ratio. For 10,000 input and 2,000 output tokens, Sol is $0.11 and Astra $0.20, roughly 1.82×. Output length and request count can change the task-level comparison. Read the Astra versus Sol comparison for model selection, then budget with your own tasks.

Separate subscription usage from API billing

The official usage guide distinguishes ChatGPT sign-in using plan allowance from an API key using API billing.

Plus Astra availability should not be described as access in every ChatGPT interface. Official guidance distinguishes Plus in Work/Codex from GPT-6 Pro in Chat for eligible plans. For a missing model, use the Plus access checklist. For exhausted usage, read the limits and resets guide.

Connect through OpenAI or Ofox

Keep the key, endpoint and model ID with the same provider. The Ofox Astra model page and September 16 catalog list openai/gpt-6-astra, with /v1/responses and /v1/chat/completions. A catalog listing does not guarantee every request from every account will succeed.

Set your OFOX_API_KEY before running this example. Executing it incurs API usage charges; it does not promise a fixed output length or request price.

curl https://api.ofox.run/v1/responses \
  -H "Authorization: Bearer $OFOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-6-astra",
    "input": "Explain three things to check in a code review."
  }'

For coding tools, continue with Astra agent setup. Before increasing volume, inspect usage and the bill for a small task, then increase context, output requirements and concurrency gradually.