Mistral Large 4 vs Kimi K3.1 Pro: API pricing and self hosting

How they differ

Mistral Large 4 lists at $0.70 per million input tokens, $0.07 for cache reads and $2.10 per million output tokens. Kimi K3.1 Pro lists at $2, $0.20 and $10. Three workloads show how those rates play out. An agent session with 40 turns, each carrying 60,000 input tokens and producing 2,000 output tokens, with four in five input tokens read from cache, costs $0.6384 on Mistral and $2.144 on Kimi. A nightly job that summarizes 20,000 documents of 30,000 input tokens and 1,500 output tokens each, with no cache hits, costs $483 and $1,500. Neither lists batch prices, so the job runs at the standard rates. A chat product with 1,000,000 conversations a month, at 4,000 input tokens and 600 output tokens each and half the input cached, comes to $2,800 and $10,400. On the standard 100,000 input, 5,000 output request with half the input cached, Mistral costs $0.049 and Kimi costs $0.16, so Kimi's bill is 3.3 times Mistral's. Long prompts need their own check. Mistral lists no separate long prompt rate. Kimi lists no separate long prompt rate. A 400,000 input token request with 4,000 output tokens and no cache hits costs $0.2884 on Mistral and $0.84 on Kimi. Kimi's weights are open, so other hosts can serve it. Relace lists Kimi K3.1 Pro at $0.3287 input and $8.115 output, running fp4 weights, so check which quantization you're buying before you compare it with the maker's price. No host is listed below the Mistral price shown. These figures leave out cache writes and tool fees.

Verdict

Start with the shape of your traffic. Mistral reads cached input at $0.07 per million tokens against Kimi's $0.20, which matters for agents that resend a long prefix every turn. Output-heavy work such as reasoning traces and long code diffs leans on the output rate, and document pipelines lean on the input rate. Then think about self hosting, because it changes what you're paying for. An API charges for tokens. Your own cluster charges for memory and for the hours it sits idle. Kimi K3.1 Pro has 2.78T total parameters with 104B active per token, spread over 896 experts with 16 used per token. Mistral Large 4 has about 1T total and 52B active, with 128 routed experts and 4 used per token. Every expert has to sit in memory even though few run for each token, so total size sets the hardware and active size sets the speed. The Kimi weights are a multi node job. Open the VRAM calculator page for each model before you price a cluster, and read both licenses before commercial use.