Read the units before comparing the numbers
These are selected text-model API rates in USD per one million tokens, checked on September 30, 2026. They are not ChatGPT subscription prices, a complete catalog of every OpenAI service, or a promise about what a particular job will cost. The official pricing page is the source of the following Standard rates for prompts up to 272K input tokens.
| Model | Fresh input | Cached input | Cache write | Output |
|---|---|---|---|---|
| GPT-6.1 Sol | $2.00 | $0.10 | $2.50 | $10.00 |
| GPT-6 Astra | $10.00 | $1.00 | $12.50 | $50.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.125 | $0.50 |
A cache read and a cache write are different billing categories. A repeatable workflow can have a cheap read after an earlier, more expensive write. A budget that assumes every input token is a cache hit will be misleading when the prefix changes or the first request establishes the cache.
A calculation, not a measured task cost
Assumed example: one request with 100,000 fresh input tokens and 10,000 billable output tokens, no cache writes or reads, Standard processing, and no separately charged tools. Applying the table gives:
- GPT-6.1 Sol: 0.1 × $2 + 0.01 × $10 = $0.30.
- GPT-6 Astra: 0.1 × $10 + 0.01 × $50 = $1.50.
- GPT-6 Luna: 0.1 × $0.10 + 0.01 × $0.50 = $0.015.
Those are arithmetic results under identical token assumptions, not evidence that all three models complete the same job equally well. A retry, a longer answer, or more review work can change the practical comparison. Use billable output from the usage record, not a word count of the answer you see.
When the headline rates do not apply
The GPT-6.1 Sol reference says requests exceeding 272K input tokens apply higher rates to the whole request: input and cache categories double, and output is 1.5 times the short-context rate. That is a pricing boundary, not simply a charge on the tokens above the threshold.
The pricing page lists Fast at twice Standard, Batch and Flex at half Standard, and a 10% premium for eligible regional processing. Since the October 8 release, both Astra and GPT-6.1 Sol also offer Ultrafast at six times Standard. These conditions are not a set of discounts and premiums to stack indiscriminately. Check supported combinations and the selected service tier.
Budget from completed work
For a publishing workflow, collect the total billed consumption for research, drafting, verification, and revision. Divide that total by the number of articles that pass review. Failed drafts belong in the numerator; counting only the final successful request hides their cost.
For a repeated source prefix, compare measured fresh, cache-write, and cache-read amounts before and after a configuration change. Keep frequently changing material separate in your experiment. Do not assume a cache will help until the usage record shows reuse.
A practical budget has three parts: observed cost per accepted result, the number of planned results, and a reserve for revisions. This is a suggested accounting method, not a vendor guarantee. It remains useful when a rate changes because you can replace the price without pretending the workload disappeared.
Compare with the Claude API price table or the focused GPT-6.1 Sol versus Opus 5.5 guide. For included Work and Codex allowances, read the separate ChatGPT subscription guide.