Devstral 2 · Mistral

What does Devstral 2 cost to run in a coding agent?

Devstral 2 costs about $0.010 per average coding-agent request on our measured traffic, at list prices of $0.40 per million input and $2 per million output tokens. It scores 9 on the Artificial Analysis Intelligence Index, 48th of the 48 models we track. Source: Artificial Analysis (artificialanalysis.ai)

List prices: OpenRouter (vendor endpoint), checked 2026-10-13. Request costs: Standard Compute gateway data, 79,301 coding-agent requests from 75 users, Sep 13 to Oct 12, 2026. How we compute this

Run current Claude, GPT, Gemini and open models in your agent

Or let the router send each step to the cheapest model that can do it, on providers we vet for price, uptime and speed. Plans from $19 a month.

See plans

What Devstral 2 costs per coding-agent request

Cost at list price	Devstral 2
Prompt under 100K tokens50% of requests, 87% of input cached	$0.0062
Prompt of 100K to 272K39% of requests, 88% of input cached	$0.015
Prompt over 272K tokens11% of requests, 86% of input cached	n/a
Average request	$0.010
100 requests	$1.01
Median session16 requests; a session ends after 10 idle minutes	$0.12
A month of regular usemedian of 20 users active on 5+ days	$12.73

Each logged request priced at the vendor’s list rates for its own token counts, including cache reads and any higher rate for long prompts. Average output: 698 tokens per request. 12% of requests were larger than Devstral 2's 262K context window and are left out of its figures.

Devstral 2 benchmarks, speed and latency

Artificial Analysis	Value	Rank
Intelligence Index10 evaluations, 0 to 100	9	48 of 48
Terminal-Bench 4.0agentic coding in a terminal	0.0%	35 of 39

Source: Artificial Analysis (artificialanalysis.ai). Measured variant: Devstral 2. Fetched October 13, 2026. Rank is among the models on this site, older versions included. Artificial Analysis runs Terminal-Bench 4.0 in its own setup, so its scores differ from the tbench.ai leaderboard.

What to watch for

Its 262K-token window is smaller than 12% of our coding-agent requests, which are left out of the figures above, so long agent sessions will hit the limit. It is weak on long agentic coding tasks (see Terminal-Bench 4.0 above), so a cheap request can turn into more steps and retries.

About Devstral 2

Devstral 2 (2512) is Mistral's open-weight model tuned for software-engineering agents, released December 9, 2025 with a 262K context, for codebase navigation, multi-file edits and tool use. Independent results are weak: it scores 9 on the Artificial Analysis Intelligence Index and 0% in Artificial Analysis's Terminal-Bench 4.0 run, and Artificial Analysis lists it as superseded. Mistral's own pricing page does not list it; the price used here is Mistral's endpoint as listed on OpenRouter.

Sources: Artificial Analysis

Devstral 2 FAQ

How much does Devstral 2 cost per coding-agent request?

$0.010 on average at list price, $1.01 per 100 requests, and $0.12 for a median 16-request session. That is 79,301 real coding-agent requests (Sep 13 to Oct 12, 2026) priced at Mistral's list rates; the median request sent 99,534 input tokens, 87% of input was read from cache, and requests returned 698 output tokens on average.

Does Devstral 2 charge more for long prompts?

No. Devstral 2 bills the same rate per token up to its 262K-token context window, which is smaller than 12% of the coding-agent requests we measured.

Is Devstral 2 good for AI agents?

Coding agents that need a Mistral open-weight model. On our scoring, which is calibrated to public benchmarks, it rates 4.4/10 for agentic ability and 3.3/10 for output quality. Open weights.

What is Devstral 2's context window?

Devstral 2 has a 262K-token context window. Mistral is the maker.

Is Devstral 2 open source?

Yes. Devstral 2 ships with open weights, so you can self-host it or run it through a hosted provider such as OpenRouter. Check the license before commercial use.

What are the downsides of Devstral 2?

The main trade-offs: 0% in Artificial Analysis's Terminal-Bench 4.0 run; 262K context, smaller than 11.7% of the coding-agent requests we measured. It's strongest for coding agents that need a Mistral open-weight model.