Mistral Large 4 vs DeepSeek V4 Pro

Quick verdict: which model fits your job

Choose Mistral Large 4 when your prompts include screenshots, charts or documents as images, or when you run agent workflows, legal or finance tasks where it posts higher scores. Choose DeepSeek V4 Pro for very long text-only inputs beyond 524,288 tokens or when you need to self-host this month.

• Images: Mistral Large 4 accepts text and images; DeepSeek V4 Pro is text-only.
• Context: 524,288 tokens served by the Mistral API vs 1M tokens for DeepSeek V4 Pro.
• Price: during Mistral's current 50% sale, Mistral Large 4 ($0.68 / $2.09) costs about the same as DeepSeek off-peak ($0.66 / $1.98) and far less than DeepSeek peak ($1.32 / $3.96).
• Weights: DeepSeek V4 Pro is downloadable under the MIT license; Mistral Large 4 weights are due at the end of October 2026.

Specs side by side

Both are Mixture-of-Experts models with the same 49B active parameters per token, so per-token compute is similar; DeepSeek V4 Pro carries about 1.5× more total parameters (1.6T vs 1.05T).

API price: sale price vs peak and off-peak

At today's prices the two models cost nearly the same per million tokens when DeepSeek is off-peak, and Mistral Large 4 is about half the price of DeepSeek at peak. DeepSeek charges peak rates 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday; Mistral has not announced when its 50% sale ends.

DeepSeek wins on cached input: $0.022 off-peak against Mistral's $0.07 sale price, so prompts that reuse a long cached prefix lean toward DeepSeek.

Worked example: a cache-heavy daily workload

For a retrieval or coding agent that sends 10M cached input tokens, 1M fresh input tokens and 1M output tokens a day, the daily bill is $3.47 on Mistral Large 4 at the sale price, $2.86 on DeepSeek off-peak and $5.72 on DeepSeek peak.

The arithmetic: Mistral sale = 10 × $0.07 + $0.68 + $2.09 = $3.47; Mistral list = 10 × $0.14 + $1.36 + $4.18 = $6.94; DeepSeek off-peak = 10 × $0.022 + $0.66 + $1.98 = $2.86; DeepSeek peak = 10 × $0.044 + $1.32 + $3.96 = $5.72. DeepSeek's peak windows cover Asian business hours and the European morning, so plan with the column that matches when your traffic runs.

Output length matters as much as price: Artificial Analysis needed 200M output tokens to run its index on Mistral Large 4 against a median of 81M, and rates it very verbose. Setting reasoning_effort to none on simple requests keeps that in check.

Benchmarks reported by Mistral

In Mistral's launch charts Mistral Large 4 beats DeepSeek V4 Pro 0813 on most agent, coding-terminal, security and legal tests; DeepSeek V4 Pro wins SWE-Atlas-QnA and ties on Finch. These are vendor-published numbers.

Independent scores are almost level

Third-party indexes put the two models within about two points of each other. Vals AI also ran Terminal-Bench 4.0 on Mistral Large 4 and got 22.73%, lower than the 28.3% in Mistral's chart, because the test harness differs.

Images and charts: an edge DeepSeek V4 Pro cannot match

DeepSeek V4 Pro takes text only, so any task that starts from a screenshot, scanned page or chart is a Mistral Large 4 job. Mistral's vision charts compare against DeepSeek-V4.1-Flash rather than V4 Pro: on Dense200 visual grounding Mistral Large 4 scores 42.0 vs 3.3, and on ChartQA Pro 63.1 vs 59.9.

You can try Mistral Large 4 right now in the chat on our homepage: 3 messages a day without an account, 15 credits a day with a free account, or Pro at $39.90 a month for 3,000 credits.