Rate card · September 2026
How the bill actually moves
Large-language-model APIs bill by the token. A rough English rule is 4 characters, or about three-quarters of a word, per token. Input and output are priced apart, and output is usually the expensive side.
Input and output
Input is the prompt, the system text, and any documents you attach. Output is the reply. On GPT-5.6 Sol, $4 per million in and $20 per million out means a short answer can cost more than a long prompt.
Worked example, GPT-5.6 Sol
1,000 input tokens cost $0.004. 500 output tokens cost $0.010. The call is $0.014. Three-quarters of that call is output, even though the prompt was longer. Shortening replies is usually the fastest cut. At 100 calls a day for 30 days, that same shape is $42.00 a month before cache or batch.
Cache
Repeated prefixes are often billed near 10% of the input rate on OpenAI, Anthropic, and Google. The calculator mixes list input and cached input by the hit rate you set. It does not model cache writes, minimum sizes, or expiry.
Batch
Where the card says batch is listed, the monthly tab halves input and output. Grok and DeepSeek stay at list price, and the row says so.
Self-host floor
A GPU rented by the hour costs the same when it sits idle. The self-host tab divides that monthly rent by the tokens the box can actually serve at the utilization you enter. If the volume where rent equals the API price is larger than that capacity, one box never wins. The number is still a floor.