Mistral Medium 3.1 (2508) is the cheapest quarter-million-token model in Mistral's lineup, and inside Atlas it is the model to pick when you want a large window without frontier pricing. It reads 262,144 tokens of context and can emit up to 262,144 tokens of output, at $0.40 / 1M input tokens and $2.00 / 1M output tokens. It is non-reasoning, so first-token latency in the Atlas TUI stays low, but it can rush architectural calls that a reasoning model would work through step by step.
What is Mistral Medium 3.1 (2508) best at inside Atlas?
Mistral Medium 3.1 (2508) is best inside Atlas at whole-module rewrites. The August 2025 Medium refresh carries a 262,144 token input window and a 262,144 token output ceiling, so Atlas can read a large service and hand back the entire replacement file in a single tool call instead of chunking.
The symmetric ceiling is the part that changes how an Atlas session feels. Most models cap output far below their input window, which forces Atlas to split a refactor across several turns and re-establish context each time. With 262,144 tokens available on both sides, a build agent driven by Mistral Medium 3.1 (2508) can absorb a service plus its tests and return one patch. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so a large single-turn rewrite still lands in a review pane rather than straight on disk. Mistral Medium 3.1 (2508) is also non-reasoning, which keeps first-token latency low in the TUI compared to the Magistral line.
How much does running Atlas on Mistral Medium 3.1 (2508) cost?
Mistral Medium 3.1 (2508) costs $0.40 / 1M input tokens and $2.00 / 1M output tokens. Filling all 262,144 tokens of its context costs roughly $0.10 per turn before Atlas generates a single output token, so a padded context is the expensive mistake, not the model choice.
Cost control on Mistral Medium 3.1 (2508) is mostly context hygiene. The $0.40 / 1M input tokens rate is cheap per token, but 262,144 tokens is a lot of tokens, and Atlas will happily fill the window if you let it. Keep the context trimmed: Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so it can inject the three functions that matter instead of six files. Output at $2.00 / 1M output tokens is the other lever. A full-module rewrite is worth it, a chatty explanation of a rewrite is not.
Mistral Medium 3.1 (2508) vs Medium 3 pricing: what changed?
Mistral Medium 3.1 (2508), released in August 2025, kept the $0.40 / $2.00 per million token price of Medium 3 and doubled the context window to 262,144 tokens in and 262,144 tokens out. Same bill, twice the window, which is why it is the value pick in Mistral's Medium tier.
For Atlas users already pinned to a Medium-tier Mistral model, Mistral Medium 3.1 (2508) is a pure window upgrade at unchanged pricing. Nothing about the per-token math shifts: input is still $0.40 / 1M input tokens and output is still $2.00 / 1M output tokens. What shifts is what fits. A repository slice that had to be summarized down now fits whole, and a rewrite that used to be split across turns can land as one diff. Atlas lets you switch the active model and provider on the fly with favorites and recents, so the change is a pin edit and a restart, not a migration.
When should an Atlas user pick a different model than Mistral Medium 3.1?
Pick a different model than Mistral Medium 3.1 (2508) when Atlas needs to reason, not type. Mistral Medium 3.1 (2508) is non-reasoning, so it can rush architectural decisions that Magistral Medium would work through step by step. For a gnarly root cause, route the plan agent to a reasoning model.
The honest boundary is between deciding and doing. Mistral Medium 3.1 (2508) shines when the plan is already clear and Atlas has to produce a lot of correct code across a big surface. It is weaker when the plan itself is the hard part: cross-service root-cause work, subtle invariants, and design tradeoffs are where a step-by-step reasoning trace earns its extra output tokens. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, which makes the split easy to exploit. Reason on one model, build on Mistral Medium 3.1 (2508), and use `model.cycle_recent` to swap without leaving the session.
How do you pair Mistral Medium 3.1 (2508) with a cheap small_model in Atlas?
Pair Mistral Medium 3.1 (2508) with `"small_model": "mistral/ministral-3b-latest"` at $0.04 / Mtok. Atlas routes titles and summaries to the cheap slot while the $0.40 / 1M input tokens Medium model handles real edits, so high-volume housekeeping calls never touch the larger model.
Atlas exposes two model slots, and using both is the difference between a predictable bill and a surprising one. Session titles, summarization, and routine subagent chatter are high-frequency and low-stakes, which is exactly what a $0.04 / Mtok model is for. Reserve Mistral Medium 3.1 (2508) for the turns that write code. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, so the cheap slot can absorb real volume. Use `/models` in the TUI to favorite Mistral Medium 3.1 (2508), then cycle against Devstral for agentic edit passes when a task is more tool loop than long rewrite.
Setup
- 01Export MISTRAL_API_KEY in the shell where you launch Atlas.
- 02Run `atlas models mistral` and confirm mistral-medium-2508 appears in the registry.
- 03Pin `"model": "mistral/mistral-medium-2508"` in atlas.json.
- 04Pair it with `"small_model": "mistral/ministral-3b-latest"` at $0.04 / Mtok for titles and summaries.
- 05Use `/models` in the TUI to favorite it, then `model.cycle_recent` to swap against Devstral for agentic edit passes.
Frequently asked questions
- what is the context window of mistral medium 3.1 2508
- Mistral Medium 3.1 (2508) has a 262,144 token context window, with an output ceiling that is also 262,144 tokens. The August 2025 refresh doubled the window Medium 3 offered while holding the same price.
- how much does mistral medium 3.1 cost per million tokens
- Mistral Medium 3.1 (2508) is priced at $0.40 / 1M input tokens and $2.00 / 1M output tokens. Feeding its full 262,144 token window costs roughly $0.10 per turn before any output is generated.
- how do I set mistral medium 3.1 as the model in atlas
- Export MISTRAL_API_KEY, run `atlas models mistral` to confirm mistral-medium-2508 resolves, then pin `"model": "mistral/mistral-medium-2508"` in atlas.json. You can also favorite it from `/models` inside the Atlas TUI.
- is mistral medium 3.1 a reasoning model
- No. Mistral Medium 3.1 (2508) is non-reasoning, which keeps first-token latency low in the Atlas TUI compared to the Magistral line. The tradeoff is that it can rush architectural decisions Magistral Medium would work through step by step.
- best mistral model for large refactors in a coding agent
- Mistral Medium 3.1 (2508) fits large refactors because both its input window and its output ceiling are 262,144 tokens, so Atlas can read a whole module and return the full rewrite in one tool call rather than splitting it across turns.
- what small_model should I pair with mistral medium 3.1 in atlas
- Pair Mistral Medium 3.1 (2508) with `"small_model": "mistral/ministral-3b-latest"` at $0.04 / Mtok. Atlas then routes titles and summaries to the cheap slot and keeps the $2.00 / 1M output tokens model for real code edits.
- can I switch models mid session in atlas
- Yes. Atlas lets you switch the active model and provider on the fly with favorites and recents. Favorite Mistral Medium 3.1 (2508) with `/models` in the TUI, then use `model.cycle_recent` to swap against Devstral for agentic edit passes.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Write Unit Tests for Untested Code with Atlas in 2026
How to write unit tests for untested code with Atlas in 2026: the lsp tool enumerates exported symbols, grep copies repo conventions, and bash actually runs the suite.
Atlas for Angular in 2026
Adopt Atlas, the terminal-native AI coding agent, for your Angular projects in 2026. Enhance development with intelligent code search, secure local embeddings, and granular control over AI actions.
Atlas vs Amp: Terminal AI Coding Agents in 2026
Compare Atlas, a terminal-native AI coding agent with free core and local embeddings, against Amp, Sourcegraph's agent featuring Oracle and Orbs, for developers in 2026.
Atlas vs GitHub Copilot: Terminal AI Coding Agents in 2026
Comparing Atlas, the terminal-native AI coding agent, with GitHub Copilot's editor extension and chat features for developers in 2026. Explore planning, pricing, and privacy.
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 Mistral Vibe for Code: Terminal AI Coding Agents in 2026
Compare Atlas and Mistral Vibe for Code in 2026. Atlas offers terminal-native TUI, explicit diffs, and BYO models. Mistral Vibe for Code provides a four-model stack, multi-platform access, and EU data sovereignty.
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.
Atlas vs Kilo Code: Terminal AI Coding Agents in 2026
Atlas and Kilo Code in 2026: Compare terminal-native TUI vs VS Code/JetBrains agents. Evaluate pricing, code safety, deployment, and model routing for AI coding.