Mistral Medium 3.5 Review: What's New, Benchmarks & Pricing
Mistral released Medium 3.5 on May 22, 2026 โ a 128B merged model with open weights, 77.6% on SWE-Bench Verified, and $1.5/M input API pricing. Here's what changed, what the benchmarks show, and whether it belongs in your stack.
Reviewed 2026-06-11 ยท Source: Mistral announcement
Quick Verdict
Mistral Medium 3.5 is the most competitive open-weights frontier model available as of June 2026. The 77.6% SWE-Bench Verified score is genuine โ it beats Devstral 2 and Qwen3.5 397B A17B, models with far more parameters. The $1.5/$7.5 API pricing is meaningfully cheaper than Claude Opus 4.8 and GPT-4o. The real story is open weights: a 128B model you can self-host on four GPUs, with no cloud data residency, is a different product category than anything OpenAI or Anthropic offers. For regulated industries or cost-sensitive agentic workloads, this is worth serious evaluation.
What's New in Medium 3.5
Merged model architecture
Mistral Medium 3.5 is described as their first 'flagship merged model' โ a single dense 128B set of weights handling instruction-following, reasoning, and coding. Most prior frontier models use separate expert routing or separate model files for different capabilities. The merged architecture means no routing overhead and consistent behavior across task types.
Configurable reasoning effort
Reasoning effort is now configurable per API request. You can dial down effort for fast conversational replies and dial up for complex agentic runs โ the same model, same weights, same pricing per token, but with control over how hard it thinks. This brings Mistral to parity with Claude's effort control and OpenAI's reasoning models for users who need latency/quality tradeoffs.
Mistral Vibe remote coding agents
Medium 3.5 powers the new async cloud coding sessions in Mistral Vibe. Sessions run in isolated sandboxes in the cloud, can be started from the CLI or Le Chat, and can run in parallel. The agent opens pull requests on GitHub and notifies you when done โ no need to watch the terminal. Integrates with GitHub, Linear, Jira, Sentry, Slack, and Teams.
Le Chat Work mode
Work mode is a new agentic mode in Le Chat powered by Medium 3.5. It connects to email, calendar, documents, and other tools to handle multi-step tasks: inbox triage, meeting prep, research briefs, cross-tool workflows. Sessions persist across many turns and every tool call is visible with explicit approval before sensitive actions like sending messages or modifying data.
Open weights under modified MIT license
Medium 3.5 weights are available on Hugging Face under a modified MIT license. The model can be self-hosted on as few as four GPUs โ an unusually low bar for a 128B model. This makes it practical for regulated industries (healthcare, finance, government) that need on-premise AI without cloud data residency concerns. Review the modified license terms before commercial deployment.
Benchmarks
Key performance numbers from Mistral's announcement (May 22, 2026).
| Benchmark | Medium 3.5 | Notes |
|---|---|---|
| SWE-Bench Verified (coding) | 77.6% | Beats Devstral 2 and Qwen3.5 397B A17B |
| ฯยณ-Telecom (agentic) | 91.4 | Strong multi-step agentic task performance |
| Context window | 256k tokens | Long-document and extended session support |
| Self-hosting requirement | 4 GPUs minimum | Accessible at smaller GPU count than most 100B+ models |
Pricing
Access to Medium 3.5 via Le Chat web interface. No API access. Usage limits apply.
Direct API access. Configurable reasoning effort. JSON mode, function calling, vision support.
Vibe remote coding agents, Work mode, extended sessions, and priority access.
Download from Hugging Face. Modified MIT license. Requires ~4 GPUs minimum for inference.
Who should use Mistral Medium 3.5?
- โRegulated-industry teams โ healthcare, finance, and government orgs that need on-premise LLM deployment without cloud data residency. Self-hosting on four GPUs is a low hardware bar for a 128B model.
- โCost-sensitive agentic workloads โ $1.5/$7.5 per million tokens is meaningfully cheaper than Claude Opus 4.8 ($5/$25). For high-volume agentic pipelines, this gap compounds fast.
- โAsync coding task delegation โ developers who want to offload module refactors, test generation, and dependency upgrades to remote cloud agents while working on other things.
- โHighest reasoning quality โ if you need the absolute best reasoning on hard math, logic, or research tasks, Claude Opus 4.8 and o3 still lead Mistral benchmarks.
- โCasual chat users โ Le Chat is fine, but GPT-4o and Claude have more mature assistant features for consumer use cases.
Frequently Asked Questions
What is Mistral Medium 3.5?
Mistral Medium 3.5 is Mistral AI's new flagship model, released May 22, 2026. It is a dense 128B parameter model with a 256k token context window. Unlike previous Mistral models, it is a 'merged model' โ a single set of weights trained to handle instruction-following, reasoning, and coding simultaneously. It is available via the Mistral API, on Le Chat, and as open weights on Hugging Face under a modified MIT license.
How does Mistral Medium 3.5 compare to Claude and GPT-4o?
On SWE-Bench Verified (real-world coding), Medium 3.5 scores 77.6% โ competitive with frontier models. Its API pricing ($1.5/$7.5 per million tokens) is significantly cheaper than Claude Opus 4.8 ($5/$25) and closer to GPT-4o pricing. The key differentiator is open weights: Medium 3.5 can be self-hosted, which Claude and GPT-4o cannot. For teams that need cloud API performance at lower cost, or on-premise deployment, Medium 3.5 is a strong option. For tasks requiring the absolute highest reasoning quality, Claude Opus 4.8 and GPT-4o still lead on most benchmarks.
Can I self-host Mistral Medium 3.5?
Yes. Weights are on Hugging Face under a modified MIT license. Mistral says inference is practical on as few as four GPUs โ notably low for a 128B model. This is not standard MIT; review the modified license terms before deploying commercially. Self-hosting gives you full data residency, no API costs per token, and no rate limits beyond your own hardware.
What is the API pricing for Mistral Medium 3.5?
Mistral Medium 3.5 is priced at $1.5 per million input tokens and $7.5 per million output tokens via the API. This puts it significantly below Claude Opus 4.8 ($5/$25) and below GPT-4o standard pricing. For high-volume agentic workloads, this cost gap compounds meaningfully.
What are Mistral Vibe remote agents?
Mistral Vibe remote agents are async cloud coding sessions powered by Medium 3.5. You start a session from the Mistral Vibe CLI or Le Chat, and the agent runs in an isolated cloud sandbox โ working through coding tasks (refactors, test generation, dependency upgrades, bug fixes) without you staying in the loop. When done, it can open a GitHub pull request and notify you on Slack or Teams. Multiple sessions can run in parallel. Available on Pro, Team, and Enterprise plans.
What is Le Chat Work mode?
Work mode is a new agentic mode in Le Chat that runs on Medium 3.5. It connects to email, calendar, documents, and tools via built-in connectors, then handles multi-step tasks: meeting prep, inbox triage, research reports, and cross-tool workflows. Every tool call is visible and Le Chat asks for explicit approval before sensitive actions like sending emails or modifying data. Available in Le Chat Pro and above.
Try Mistral Medium 3.5
Available on Le Chat, the Mistral API, and as open weights on Hugging Face.
Try Mistral โAPI model ID: mistral-medium-3-5 ยท Open weights on Hugging Face (modified MIT)
Affiliate disclosure: Some links on this page are affiliate links. If you sign up through them, AISO Tools may earn a commission at no extra cost to you. This never affects our rankings or reviews.
๐ฌ Get the best new AI tools delivered weekly
One concise email with fresh launches, trending picks, and featured standouts.
Join thousands of professionals who discover the best AI tools every week. No spam โ unsubscribe anytime.