GPT-6.1 Sol Cost Calculator

Cost per request = (uncached input tokens × input rate + cached input tokens × cached rate + output tokens × output rate) ÷ 1,000,000.

The batch switch applies the provider's batch discount to every token type. Requests above a long-context threshold use the higher long-context rates for the whole request. A month is 30 days.

Here is how the GPT-6.1 Sol cost calculator gets its numbers, using the RAG Q&A preset. Each request sends 4,000 input tokens (instructions, retrieved passages and the question) and gets a 500-token answer, and the assistant handles 20,000 requests a month.

1. Input: 4,000 × $2 ÷ 1M = $0.008
2. Output: 500 × $10 ÷ 1M = $0.005
3. One request: $0.013. One month: $0.013 × 20,000 = $260

If the first 1,000 tokens are a fixed system prompt, they can be served from the prompt cache at $0.10 per 1M instead of $2. That brings the month to $222.

Sending the same requests through the Batch API, if answers can wait, costs $130 a month.

The presets are a starting point. To estimate your own GPT-6.1 Sol bill, replace them with numbers from your app:

1. Tokens per request. Every API response reports the input and output tokens it used. Log them for a day of real traffic and enter the averages. Before launch, count the system prompt, any chat history or documents you send, and a typical reply.
2. Requests per month. Multiply daily requests by 30. For an agent, count each model call, not each task: one task can take dozens of steps.
3. Cached share. Work out how much of each prompt is identical from one request to the next, such as instructions and reference documents, and enter that share with caching turned on.
4. Batch. Turn on the batch switch for the part of your traffic that can wait for results.
5. Headroom. Retries, reasoning tokens and traffic spikes all add to the bill, so budget above the estimate rather than exactly at it.

Reply length

Output costs 5 times as much as input on GPT-6.1 Sol. Every extra 100 output tokens per request adds $10.00 per 10,000 requests, so asking for shorter answers and setting a maximum output length pays off quickly.

Prompt caching

Cached input costs $0.10 per 1M tokens, 5% of the normal input rate. Writing a prefix to the cache costs $2.50 per 1M tokens, and that extra cost is recovered after the first cache hit. Keep instructions and reference documents at the start of the prompt and unchanged between requests so they are served from the cache.

Batch processing

Requests sent through the Batch API cost 50% less. Results arrive asynchronously instead of in real time, which suits evaluations, backfills, document processing and nightly reports.

Long prompts

Once a request passes 272K input tokens, the whole request is billed at $4 input and $15 output per 1M tokens. A request with exactly 272,000 input tokens and a 1,000-token answer costs $0.55; 10% more input costs $1.21. If you work with long documents, trimming them below the threshold saves more than the token count suggests.

For the RAG example above, GPT-6.1 Sol comes to $260 a month. Among similar models, GPT-6 Luna would cost $13.00 and GPT-6 Astra $1,300 for the same traffic. The cheapest of the 49 models available in the API for this workload is Qwen3.7-Flash, at $3.70 a month.

Price is only half the decision: a cheaper model that needs a second attempt or a longer prompt to get the answer right can end up costing more. Run your own token counts through the table above, which prices the same workload on every model we track, and test the two or three cheapest candidates on real requests before you switch.