GPT-6.1 Sol costs $2 per million uncached input tokens, $0.10 per million cached input tokens and $10 per million output tokens at Standard API rates for requests with up to 272,000 input tokens. A hypothetical request using 100,000 uncached input tokens and 10,000 output tokens has a $0.30 token subtotal, before other charges.
Those rates answer the price question. Following the request through retries, tools and review answers a different one: what does it cost to finish the job? OpenAI’s GPT 6.1 Sol announcement introduced the model on September 29 for coding, computer use and tasks spanning several steps.
Standard API token prices
The following USD rates were checked against OpenAI's API pricing on October 23, 2026. They apply to Standard processing, per one million tokens in the named column.
| Input size of the request | Uncached input | Cached input | Output |
|---|---|---|---|
| Up to 272,000 input tokens | $2.00 | $0.10 | $10.00 |
| More than 272,000 input tokens | $4.00 | $0.20 | $15.00 |
Crossing 272,000 input tokens changes the rates for the whole request, not just the tokens above the threshold. Cache writes, tools, storage and other applicable charges are additional. Cached input here means an eligible cache hit, rather than the cost of creating the cache.
The model reference identifies the model as gpt-6.1-sol and gives the processing qualifications. Fast costs twice Standard, while Batch and Flex are 50% lower. Ultrafast, described as forthcoming at launch, is now documented at six times Standard. Regional processing can add a further premium. Choose the applicable mode before using the table as a cost estimate.
Worked request costs
Multiply each token count by its rate per million, then add input and output. These are hypothetical Standard API subtotals, not observed bills or measured task costs. They exclude cache creation and other charges.
The Sol and Luna comparison connects these per-token numbers to a practical test of cost per accepted result.
| Assumed request | Input cost | Output cost | Token subtotal |
|---|---|---|---|
| 10,000 uncached input and 1,000 output | $0.02 | $0.01 | $0.03 |
| 100,000 uncached input and 10,000 output | $0.20 | $0.10 | $0.30 |
| 100,000 eligible cached input and 10,000 output | $0.01 | $0.10 | $0.11 |
| 90,000 eligible cached input, 10,000 uncached input and 10,000 output | $0.029 | $0.10 | $0.129 |
| 300,000 uncached input and 10,000 output | $1.20 | $0.15 | $1.35 |
The final row uses long-context rates for all 300,000 input tokens and the output. The partial-cache row adds $0.009 for the eligible cache hit to $0.02 for the uncached input, then $0.10 for output. That calculation assumes the stated hit actually appears in usage, rather than treating all repeated context as free. The 11-cent row assumes the entire input qualifies for the cache-hit price. A changing prompt or an additional call can produce a different total even when the intended task sounds identical.
API billing and subscription usage are separate
OpenAI's ChatGPT Work and Codex pricing guidance separates API token prices from included subscription usage and says not to use them to estimate included tasks. Work and Codex share usage under that product's pricing and limits. A 30-cent API example therefore does not tell you how many tasks a subscription includes.
At launch, OpenAI said Sol was available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. That historical product announcement and the API model's billable rates describe different forms of access. Check the current plan and mode for the surface you use.
A large language model generates tokens from its context. In an agent workflow, output can request a tool action. The tool returns information, the context grows and another model call follows. The completed job can therefore involve several token subtotals, tool charges and a person's time checking the result.
A stable prefix changes the input bill
OpenAI’s prompt-caching guide explains that reuse depends on an exact matching prefix. Put stable instructions and examples before changing task details when that ordering also makes the task clear. Similar meaning is insufficient if the tokens at the beginning differ.
Expected reuse belongs in the plan. Observed cached input belongs in the cost record. Eligibility and retention conditions affect what is actually reused, so check the usage fields before treating the lower rate as an achieved saving.
Keeping a prefix stable also makes a comparison easier to read. If the task instructions change between runs, a different result may come from those changes. Hold the common definition still while testing the variable you mean to compare.
Benchmarks describe their own tasks
OpenAI reports gains over the previous Sol model on coding and computer-use evaluations. Those results are evidence about the vendor’s specified tasks and settings. They do not establish identical reliability across every repository, document format, or desktop workflow.
Choose a small set of real tasks and define completion before comparing models. A code change may need to pass existing tests and preserve an interface. A document may need accurate figures and conclusions tied to evidence. A browser task may need the final page state checked. These give “finished” something concrete to mean.
Now count the attempts that reach that standard, the ones that need repair and the ones abandoned. A fluent answer can look ready early. Checking the required outcome tells you whether the task has actually advanced.
Close the record when the work is accepted
For one task, keep all model calls in the same record as tool expenditure and human review. Then mark the outcome: accepted, repaired or abandoned. You can compare equivalent finished work without letting a cheap first attempt stand in for the whole cost.
Different tasks may lead to different choices. An easily checked operation may work well on the lower-priced model. A task that repeatedly needs repair may justify another option. Let the observed work and its acceptance criteria determine where that tradeoff falls.
A useful cost record has a final row: what was accepted, and what did it take to get there? The 30-cent subtotal belongs at the beginning of that record. Keeping the rest beside it makes the next model comparison much easier to trust.
FAQ
Why can a request just above 272,000 input tokens cost more than expected?
The higher input and output rates apply to the entire request at that threshold. A complete estimate also needs to account for charges outside the listed token subtotal.

