Mistral Medium 3.5 Review: Pricing, Context & Verdict

Mistral Medium 3.5 is Mistral AI's first flagship "merged" model — a dense 128B network that folds instruction-following, reasoning, and coding into one set of open weights, with a 256k-token context and an official API price of $1.50 in / $7.50 out per million tokens. It is the new default in Le Chat and the engine behind Mistral's Vibe CLI coding agent. This review covers what it is, where it sits against Mistral's own lineup and the frontier, the benchmarks (clearly split into vendor-reported vs independent), the real price, and a plain verdict. We also ran it through our own executed 9-task coding benchmark (it scored 9/9) — where a number is self-reported, we say so.

The short version

• What it is: Mistral Medium 3.5 (model ID mistral-medium-3-5-26-04) is a frontier-class multimodal model optimized for agentic and coding use, released as open weights under a Modified MIT license around April 28, 2026.
• The headline design: Mistral's first flagship "merged" model — a dense 128B network (vendor-stated) with a 256k-token context that combines chat, reasoning, and coding in one set of weights. It is now the default in Le Chat and replaces Devstral 2 in the Vibe CLI.
• The price: $1.50 in / $7.50 out per million tokens, per Mistral's official docs. On DataLLM Lab the rate matches that first-party price exactly — no gateway markup.
• Benchmarks: vendor-reported SWE-Bench Verified 77.6% and tau-3 Telecom 91.4; independently, Artificial Analysis measures an Intelligence Index of 30 (#2 of 62 open-weights models) at ~132 tokens/sec.
• First-party test: we ran it through our executed 9-task coding benchmark — 9/9, at ~$0.87 per 1,000 tasks and 2.9s average latency (the fastest model in that run). The 128B parameter count and the SWE-Bench/tau-3 scores remain self-reported by Mistral; verify against your own workload.

What Mistral Medium 3.5 is

It is Mistral's open-weights flagship for agentic and coding work — one dense model, multimodal, with a large context. The spec sheet, straight from Mistral's official model card and news page (as of July 2026):

Every figure above is from Mistral's official model card, except the two marked vendor-stated. Two things stand out for builders. First, this is genuinely open-weights — you can download and self-host it, which most frontier-class models still do not allow. Second, it is built for tool use, not just chat: function calling, agents/conversations, and structured outputs are first-class, which is why Mistral leans on it for agentic workloads. One caveat worth flagging: "Modified MIT" is not the same as a standard OSI MIT license — read the terms before you assume unrestricted commercial reuse.

The 'merged model' positioning

Mistral calls it their first "merged" model: one dense checkpoint that replaces a stack of specialists. Rather than shipping separate models for instruction-following, reasoning, and coding, Medium 3.5 folds all three into a single set of weights. Per Mistral's own announcement, this has two concrete consequences:

• It becomes the default in Le Chat — Mistral's consumer assistant — so it is the model most Mistral users now hit by default.
• It replaces Devstral 2 in the Vibe CLI, Mistral's terminal coding agent. That is a strong internal signal: Mistral is retiring a dedicated coding checkpoint in favor of the merged model.

For a gateway user, the practical upshot is fewer routing decisions. Instead of splitting traffic between a chat model, a reasoning model, and a coding model, you point mixed workloads at one endpoint and one price. That is the same one-model-for-everything argument you weigh when choosing a coding LLM — the trade is convenience and consistent behavior against the peak scores a narrow specialist might hit on its home turf.

Benchmarks: vendor vs independent

The vendor-reported coding and agentic scores are strong; the independent signal is a top-tier open-weights ranking, and we add our own executed coding result. Keep the sources separate — we do not treat a vendor number as verified.

What each says. On SWE-Bench Verified — the standard real-world software-engineering benchmark — Mistral claims 77.6%, ahead of its own Devstral 2 and of models like Qwen3.5 397B A17B (Mistral's framing). The tau-3 Telecom score of 91.4 targets multi-step agentic tool use, which lines up with the model's stated purpose. Both come from Mistral, so treat them as a claim to reproduce on your own tasks, not a settled result.

The independent read comes from Artificial Analysis, whose proprietary Intelligence Index (a composite across their eval suite) puts Medium 3.5 at 30 — good enough for #2 of the 62 open-weights models they track — with a measured throughput around 132 tokens/second. A composite index is a rough guide, not a per-task guarantee, but a #2 open-weights placement is a real, third-party signal that this is a capable model, not just a well-marketed one.

Pricing & context

$1.50 input / $7.50 out per million tokens — and the gateway does not mark it up. Mistral's official first-party API rate and the DataLLM Lab catalog rate are identical, so there is no gateway premium on the published price. That matters because the mid-tier is exactly where price discipline pays off: at a 4:1-ish output-to-input ratio, a typical agentic task on Medium 3.5 lands well below flagship models like Claude Opus while carrying a 256k context most cheap models can't match.

Where it sits on the cost curve: it is not a rock-bottom model — plenty of open-weights options undercut it on raw per-token price. It is a value model: frontier-adjacent capability and a large context at a mid-tier rate. If your priority is the lowest possible bill, our cheapest LLM API guide ranks the sub-dollar tier; if it is capability-per-dollar with real tool-use quality, Medium 3.5 is a strong pick. Full catalog pricing is on the pricing page.

How it compares to rivals

Against Mistral's own lineup it is the new top of the stack; against the frontier it competes on value, not on peak raw scores. A few honest comparisons:

• vs Devstral 2 (Mistral's coding specialist): Medium 3.5 replaces it in the Vibe CLI and, per Mistral, beats it on SWE-Bench Verified — so for Mistral-native coding it is now the default choice.
• vs closed frontier models (Claude, GPT): the closed flagships still tend to top the independent leaderboards, but they are pricier and you cannot self-host them. If you are weighing a mid-tier against a flagship in general, the same logic in our Claude Sonnet vs Opus breakdown applies — most workloads do not need the most expensive tier, and a capable mid model at a fraction of the price wins on total cost.
• vs other open-weights models: Artificial Analysis ranks it #2 of 62 open-weights models on their index — so among models you can actually download, it is near the top. Other served open-weights options on the gateway (Grok 4.3, GLM 5.2, Nemotron 3 Ultra) trade blows on different axes; benchmark them on your own workload rather than trusting any single composite.

Pros & cons

The case for and against, distilled.

Calling it on the gateway

Mistral Medium 3.5 is served on DataLLM Lab — call it with the OpenAI-compatible API and the same key you use for every other model. Point the base URL at the gateway and pass the model slug:

from openai import OpenAI

client = OpenAI(
    base_url="https://www.datallmlab.com/v1",
    api_key="YOUR_DATALLMLAB_KEY",
)

resp = client.chat.completions.create(
    model="mistralai/mistral-medium-3-5",
    messages=[{"role": "user", "content": "Refactor this function and add tests."}],
)
print(resp.choices[0].message.content)

Because the gateway is OpenAI-compatible, switching to Medium 3.5 is a one-line model-string change — no new SDK, no separate key. See the full model page at /models/mistralai/mistral-medium-3-5/ or browse all models.

Verdict

Mistral Medium 3.5 is the best-value open-weights model in Mistral's lineup for agentic and coding work — pick it when you want frontier-adjacent tool use, a big context, and the option to self-host, without paying flagship prices. The merged-model design is a real convenience: one checkpoint, one price, consistent behavior across chat, reasoning, and code. The independent #2 open-weights ranking gives it credibility beyond Mistral's own numbers.

Where it is not the answer: if you need the single highest score on an independent leaderboard, the closed frontier models still edge it; if you need the absolute cheapest per-token rate, sub-dollar open-weights models undercut it. Mistral's headline coding/agentic scores are still vendor-reported, but our own executed coding run (9/9, and the fastest model we tested) is one first-party data point in its favor — still, validate it on your own tasks before you commit a production workload. For most teams building agents or coding tools who value ownership and cost, though, it earns a place on the shortlist. (Volatile facts above are current as of July 2026.)

FAQ