Mistral Medium 3 (2505) is the May 2025 model that reset Mistral's price-to-performance curve. Medium 3 was the first Mistral tier to pair a 131,072 token window with a fully symmetric 131,072 token output budget at $0.40 / 1M input tokens. Inside Atlas, Mistral Medium 3 (2505) is a cheap EU-hosted workhorse whose generations are never clipped by an artificial output cap. Output bills at $2.00 / 1M output tokens.
What do symmetric limits mean for Atlas on Mistral Medium 3 (2505)?
Mistral Medium 3 (2505) offers symmetric limits: 131,072 tokens of context and 131,072 tokens of output. A long generation from Mistral Medium 3 is never clipped by an artificial output cap, which is unusual in a tier priced at $0.40 / 1M input tokens.
Symmetry removes a whole class of workflow friction. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and on a model with a small output ceiling that diff gets chopped into fragments that you approve one at a time. Mistral Medium 3 (2505) can write back as much as it read, so a change stays whole. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, which means the 131,072 tokens going in are complete declarations, and the 131,072 tokens available to come back are enough to rewrite what was read.
How cheap is Mistral Medium 3 (2505) compared to Mistral Large 2.1?
Mistral Medium 3 (2505) charges $0.40 / 1M input tokens, which is 5x cheaper than Mistral Large 2.1 for the same 131,072 token window. Output on Mistral Medium 3 (2505) is $2.00 / 1M output tokens, a third of what the older flagship asks.
That price gap is why Medium 3 reset Mistral's price-to-performance curve in May 2025. For Atlas, a 5x cheaper input rate means the retrieval side of the agent stops being a budget concern. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and at $0.40 / 1M input tokens you can let that search fill the window rather than hand-selecting two files. The remaining discipline is on output: at $2.00 / 1M output tokens, Atlas's parallel background subagents still add up, so route the mechanical work to a smaller model.
Can you self-host Mistral Medium 3 (2505)?
Mistral Medium 3 (2505) is deployable self-hosted on customer infrastructure under a Mistral enterprise agreement, which is unusual: most US frontier APIs offer no equivalent. For an Atlas deployment inside a regulated environment, that makes Mistral Medium 3 a genuinely different option rather than a cheaper variant of the same thing.
Self-hosting closes the last gap that a cloud API cannot. Atlas already lets you keep the retrieval layer private, since Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, and every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs. Adding a self-hosted Mistral Medium 3 (2505) behind that means neither the index nor the chat traffic leaves your infrastructure. Teams that cannot send source code anywhere at all should evaluate Mistral Medium 3 on this axis before comparing token prices.
Should you use Mistral Medium 3 (2505) or its successor?
Mistral Medium 3 (2505) is superseded by Medium 3.1 (2508), which doubles the window to 262,144 tokens at the identical price. A free doubling of context means Mistral Medium 3 (2505) is now a deliberate choice rather than a default, and it needs a specific reason behind it.
The reason usually comes down to pinning. Pin the dated id in atlas.json as `"model": "mistral/mistral-medium-2505"`, because the -latest alias now points at a newer generation and you may not want your Atlas config to drift underneath a validated pipeline. If you do not have that constraint, take the larger window. Verify with `atlas models mistral` that mistral-medium-2505 resolves before you rely on it, since an id that silently fails to resolve is the most common cause of a session starting on the wrong model.
When should you not run Atlas on Mistral Medium 3 (2505)?
Mistral Medium 3 (2505) has no reasoning mode, so multi-hop debugging benefits from Magistral instead. Any task where Atlas's read-only plan agent needs to chain several inferences together is the wrong job for Mistral Medium 3, regardless of how attractive $0.40 / 1M input tokens looks.
The practical pattern is to split the work by phase. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and Atlas lets you switch the active model and provider on the fly with favorites and recents. Use a reasoning model for the plan, then cycle to Mistral Medium 3 (2505) for the build pass, where its symmetric 131,072 token output budget and $2.00 / 1M output tokens rate do good work. Treating Mistral Medium 3 as a competent executor rather than a planner is how you get value out of it.
Setup
- 01Export MISTRAL_API_KEY or run `atlas login` and select Mistral.
- 02Verify with `atlas models mistral` that mistral-medium-2505 resolves.
- 03Pin the dated id in atlas.json: `"model": "mistral/mistral-medium-2505"`. The -latest alias now points at a newer generation.
- 04Enable Atlas's local Ollama embedder if you want the codebase index to stay on-device while chat goes to Mistral's EU endpoint.
- 05Favorite a reasoning model alongside Mistral Medium 3 (2505) so the plan agent can hand off to it for multi-hop debugging.
Frequently asked questions
- how much does mistral medium 3 cost
- Mistral Medium 3 (2505) costs $0.40 / 1M input tokens and $2.00 / 1M output tokens, which is 5x cheaper on input than Mistral Large 2.1 for the same 131,072 token window.
- what is the mistral-medium-2505 context window
- Mistral Medium 3 (2505) has symmetric limits: a 131,072 token context window and a 131,072 token output budget, so a long generation is never clipped by an artificial output cap.
- can mistral medium 3 be self hosted
- Yes. Mistral Medium 3 (2505) is deployable self-hosted on customer infrastructure under a Mistral enterprise agreement, unlike most US frontier APIs.
- does mistral medium 3 support reasoning
- No. Mistral Medium 3 (2505) has no reasoning mode. Multi-hop debugging benefits from Magistral instead, with Mistral Medium 3 handling the build pass.
- how do i pin mistral medium 3 in atlas
- Export MISTRAL_API_KEY, verify with `atlas models mistral` that mistral-medium-2505 resolves, then pin `"model": "mistral/mistral-medium-2505"` in atlas.json. The -latest alias now points at a newer generation.
- is mistral medium 3 still worth using in 2026
- Mistral Medium 3 (2505) is superseded by Medium 3.1 (2508), which doubles the window to 262,144 tokens at the identical price. Stay on 2505 only when a pinned dated id is a requirement.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs Traycer in 2026: Terminal Agent That Writes Code vs a Planning Layer Above One
Atlas vs Traycer in 2026. Traycer plans and verifies but writes no code itself, so you pay it on top of an agent. Atlas plans and writes, free and open source.
Atlas vs Magic.dev: Terminal AI Coding Agents in 2026
Compare Atlas, the terminal-native AI coding agent with permission-gated tools and diff review, against Magic.dev's research claims of 100 million token context models in 2026.
Atlas for TensorFlow: Terminal-Native AI Coding for Keras 3 and tf.data in 2026
Atlas is a terminal-native AI coding agent for TensorFlow and Keras 3 in 2026, where tf.function graph tracing and tf.data pipelines are what make training fast.
Atlas for React in 2026
Adopt Atlas, the terminal-native AI coding agent, for React development in 2026. Enhance your workflow with intelligent code search, refactoring, and testing for React components and hooks.
Atlas vs PearAI: Choosing Your AI Coding Agent in 2026
Compare Atlas, the terminal-native AI coding agent, with PearAI, a VS Code fork. Evaluate features, pricing, and community support for developers in 2026.
Atlas for Go in 2026
Atlas, the terminal-native AI coding agent, empowers Go developers in 2026 with intelligent code understanding, safe refactoring, and robust testing capabilities.
Atlas for Java in 2026
Adopt Atlas, the terminal-native AI coding agent, for Java development in 2026. Enhance your workflow with intelligent code search, refactoring, and robust safety features for Maven and Gradle projects.
Run Atlas Headless in CI with Atlas (2026 Workflow)
How to run Atlas headless in CI in 2026: atlas run sends one prompt and exits when the session goes idle, with --format json, --command, and --continue for pipeline steps.