Short answer: The current GPT-6 Astra Standard API rates are $10 per million uncached input tokens, $1 per million cached input tokens, $12.50 per million cache-write tokens and $50 per million output tokens. When a request contains more than 272,000 input tokens, OpenAI documents 2× input and cache pricing and 1.5× output pricing for the full request. Therefore, a 1,000,000-token uncached-input request in Standard mode costs $20 for input before output; it does not cost $10. Add output separately, and apply a documented processing-mode multiplier only when that mode is actually selected and available.
This page explains Astra token arithmetic. The launch overview covers rollout, the coding guide shows a Responses tool loop, and the GPT-5.6 pricing guide remains the source for Sol, Terra and Luna economics.
Current Standard rates
| Token class | Standard rate per 1M | Use this for |
|---|---|---|
| Input | $10.00 | Tokens not served from the prompt cache |
| Cached input | $1.00 | Input tokens read from an eligible cache hit |
| Cache write | $12.50 | Tokens written to the cache under the documented cache-write rule |
| Output | $50.00 | Tokens generated by the model |
These examples use the OpenAI API model-page rates. The Enterprise Chat, Work and Codex rate card is a separate surface with different Fast, long-context and cache-write treatment; do not substitute it for the API examples. ChatGPT subscription allowances and negotiated Enterprise terms are also separate.
Long-context pricing: the 272K threshold
OpenAI’s model documentation says that if the request has more than 272,000 input tokens, the multiplier applies to the full request: input and cached-input rates are doubled, and output is multiplied by 1.5. Do not apply the multiplier only to the tokens after 272K.
Example: 1M uncached input plus 20K output
- Input: 1,000,000 × ($10 / 1,000,000) × 2 = $20.00.
- Output: 20,000 × ($50 / 1,000,000) × 1.5 = $1.50.
- Illustrative Standard total: $21.50, before taxes, retries, tool charges or other services.
The key correction is the input line: the million-token uncached input costs $20 under the documented long-context rule. A one-million-token request is not priced as $10 merely because the base rate is $10.
Example: 1M cached input plus 20K output
- Cached input: 1,000,000 × ($1 / 1,000,000) × 2 = $2.00.
- Output: 20,000 × ($50 / 1,000,000) × 1.5 = $1.50.
- Illustrative Standard total: $3.50 when the full input is an eligible cache hit.
A cache hit is not the same as a cache write. A first request may incur cache-write pricing under the documented caching behavior; model the cache lifecycle and expiration rather than assuming every repeated prompt is cached.
Processing modes and their multipliers
| Mode | Documented relationship | Budget implication |
|---|---|---|
| Standard | Base rates above, with the long-context rule when applicable. | Use for the baseline estimate. |
| Batch | 50% of Standard rates where the request qualifies. | Useful for asynchronous work; do not assume interactive latency. |
| Flex | 50% of Standard rates where the request qualifies. | Confirm queue and availability behavior before promising a deadline. |
| Fast | 2× applicable rates where available. | Pay for latency; it does not turn a token into a different model-quality guarantee. |
Apply one example only after selecting the mode. For instance, if a qualifying 1M-uncached-input/20K-output job uses Fast and the endpoint applies the documented 2× Fast rate to the already calculated Standard total, the illustrative token subtotal is $43.00 ($20 input + $1.50 output, then ×2). Confirm how the selected endpoint reports mode and usage. The current latest-model guidance says Fast is unavailable with EU data residency, so do not include it in an EU-residency forecast without a documented eligible route.
Cache-write and context planning
The model page lists cache writes at 1.25× uncached input. Keep three quantities separate in a spreadsheet: fresh input, cache writes and cache-hit reads. A useful row structure is:
| Job | Fresh input | Cached read | Cache write | Output | Mode |
|---|---|---|---|---|---|
| First research pass | Source tokens | 0 | Reusable instructions/source prefix | Answer tokens | Standard or Batch |
| Second question | New question | Eligible prefix | 0 or new prefix | Answer tokens | Standard |
Do not pad a prompt to use the 1.05-million-token context window. Select only sources needed for the decision, summarize stable material, and measure whether a larger context reduces review or retrieval work. If a request crosses 272K, apply the full-request multiplier to the relevant input/cache and output classes.
Compare Astra with GPT-5.6 without mixing billing surfaces
Astra’s current API rates are materially different from GPT-5.6 Sol, Terra and Luna. The GPT-5.6 guide documents its own model routing and the pricing history; do not substitute a ChatGPT or Work allowance for API tokens. Compare a fixed job with the same source pack, output contract, retries, review time and accepted-result rate. A lower per-token price can lose if it requires more correction; a higher-priced model can lose if the capability is unnecessary.
Five controls for a realistic estimate
- Count input, cache-hit, cache-write and output tokens separately.
- Split jobs that cross 272K from jobs below the threshold.
- Record Standard, Batch, Flex or Fast explicitly in the forecast.
- Include retries, tool calls, storage, retrieval and human review where applicable.
- Reconcile a sample of provider usage with the forecast before scaling.
Frequently asked questions
How much is one million uncached input tokens above 272K?
At Standard rates, $20 for the input portion: 1M × $10/M × 2. Output is additional and uses the 1.5× long-context multiplier.
How do Batch and Flex affect the estimate?
OpenAI lists Batch and Flex at 50% of Standard rates where the request qualifies. Confirm queue, availability and deadline behavior before using that discount in a production forecast.
Is cached input always $1 per million?
$1/M is the Standard base cached-input rate. A request over 272K uses the documented 2× input/cache multiplier; cache writes are a separate class.
Does API pricing tell me how many ChatGPT messages I get?
No. API token billing and subscription, Work or Codex usage allowances are separate surfaces.
Bottom line
Use the four token classes, then apply the long-context and processing-mode rules. For a 1M uncached-input request above 272K, input is $20 in Standard mode before output—not $10. Keep the arithmetic visible and test it against provider usage before committing budget.
For capability and price context beyond the token arithmetic, compare the frontier model guide and the older-model comparison; each keeps provider-specific benchmark and compatibility claims attributed.
Do not transfer these API examples to Enterprise Chat, Work or Codex billing. OpenAI’s separate rate card currently lists Fast at 2.5× for that surface and describes Codex Astra without the API page’s >272K multiplier or separate cache-write charge. Use the rate card for the named surface and the model page for API estimates.
API rates and processing rules are current as of September 5, 2026; see the Astra model page for API examples and the Enterprise Chat/Work/Codex rate card for its separate surface rules.
