GPT-6 Sol API pricing: calculate cache and task costs

GPT-6 Sol’s direct OpenAI Standard rates start at $2 per million ordinary input tokens and $10 per million output tokens. Those two rates are not the whole bill. Cache writes, cache reads, long inputs and processing mode change the calculation.

This guide is for developers estimating API usage. Rates were checked on September 23, 2026 against the model page and OpenAI pricing documentation. All amounts below are USD per million text tokens. They are direct OpenAI rates, not Ofox catalog prices or subscription allowances.

Start with the right rate row

The long-context threshold changes the applicable rates for the full request, not just tokens above 272K. The model’s overall context capacity is a separate limit. You can remain within its technical capacity while crossing a pricing threshold.

Batch and Flex use half the applicable Standard rates; Fast uses twice the applicable rates. These are different processing options, not discounts to stack indiscriminately. Regional processing adds a premium where supported, and EU data residency is limited to Standard processing in the current documentation. Check the mode you actually send, rather than selecting the cheapest row from a table after the request has run.

Do not bill cache writes twice

The prompt-caching guide separates ordinary input, cache reads and cache writes. A token charged as a cache write uses that rate instead of the ordinary-input rate. Do not charge it once at $2 and again at $2.50.

For a request with non-overlapping token categories:

cost = (ordinary_input × input_rate
      + cache_read × read_rate
      + cache_write × write_rate
      + billed_output × output_rate) / 1,000,000

If your usage export begins with total input, subtract the cache categories to obtain ordinary input, using the provider’s definitions. Never assume that a field named input means uncached input. Record billed output, including billed reasoning usage where applicable, rather than counting only the text visible to the reader.

Work through a reproducible example

Assume a short-context Standard request contains 20,000 ordinary input tokens, 60,000 cached-read tokens, 20,000 cache-write tokens and 5,000 billed output tokens. The four input/output categories are disjoint. This is an arithmetic example, not a measured model run.

Download the rate sheet and example costs or the local cost calculator. The calculator reproduces the published arithmetic without calling an API. Replace the sample usage with your own mutually exclusive categories; it does not retrieve your bill or infer cache hits.

A request with 300,000 ordinary input tokens and 5,000 output tokens crosses the threshold. At Standard long-context rates it costs 0.3 × $4 + 0.005 × $15 = $1.275, before separate charges. Applying short-context input rates to the first 272K would undercount that example.

Compare with GPT-5.6 Sol carefully

The launch announcement compares GPT-6 Sol’s $2/$10 ordinary input/output rates with the earlier promotional $4/$20 rates. That is a like-for-like 50% reduction for those two rate fields. It is not a promise that a migration halves your bill: the new run may use a different amount of context, reasoning, tools or retries.

Keep your old model’s historical invoice intact. Repricing its usage at a new rate answers a hypothetical question; it does not change what you paid. For migration decisions, compare the cost of an accepted result and the manual work needed to get there.

Budget for completed tasks

A useful task ledger records the model ID, supplier, processing mode, input size, cache categories, billed output, separate tool charges and acceptance result. Include failed attempts. Dividing the cost of successful attempts alone by the number of successes hides the expense of the work you rejected.

Use the Sol versus Opus 5.5 comparison for a two-model cost framework. Use Sol, Luna or Astra when deciding how much model capacity a task needs. These guides do not substitute a leaderboard for your own acceptance criteria.

Before buying more capacity, check whether long conversation histories or repeated uncached prefixes explain the bill. The cache-cost worksheet guide covers the same accounting principle across providers. Confirm each supplier’s current fields rather than transferring OpenAI’s rates to an Ofox route.