Mistral Large 4 API pricing and cost calculator

How this is calculated

The calculator opens on a request of 100,000 input tokens and 5,000 output tokens. Uncached, the input part costs $0.07 and the output part costs $0.0105, so one call comes to $0.0805. With half the input read from cache it's $0.049, and 100 calls a month add up to $4.90. At that mix the prompt side costs $0.0385 and the answer side costs $0.0105, so look at which half of your own traffic is heavier before you decide whether to trim context or cap reply length. Three other shapes show how the same rates behave. An agent session of 40 turns, each sending 60,000 input tokens with about nine tenths read from cache and getting 1,500 tokens back, costs $0.4452. The same session with a cold cache on every turn costs $1.806. A nightly document pipeline that reads 20,000 files of 8,000 tokens each and writes 600 tokens per file, with nothing cached, costs $137.20 per run. A chat product with 200,000 conversations a month, each sending 3,000 input tokens with half cached and receiving 400 output tokens, costs $399 a month. Those figures cover tokens only. Cache writes, tool fees, retries and any provider routing fees come on top, so treat them as a floor and compare them with real billing records after a week of traffic. Output limits matter too. Large 4 can generate up to 262,144 tokens in one reply, and a runaway generation after a 10,000 token prompt would cost $0.5575024. Set a max output setting that fits your product instead of leaving it at the ceiling. No batch price is listed for Large 4 right now, so a nightly job runs at the standard rates. Check Mistral's pricing page before you assume one exists. The prices above are the maker's own. Open weights mean other hosts can list the model too, and the calculator will show a lower host price here if one appears.

Verdict

Large 4 suits teams that want a large open weight model through an API and may want the option to host it later. The weights sit on Hugging Face under the Mistral Ecosystem Open-Weight License, so read that license before any commercial use. Start by logging prompt length, cached tokens and output length per finished task, then price those numbers here. The context window is big enough that a careless prompt can grow quietly, and every extra input token is billed. If self hosting is on your list, the VRAM calculator page for this model shows what the weights and cache need, and it's a multi node job at this size.