Devstral Small 2505 is the May 2025 model that started the Devstral line, and inside Atlas it is a resolver rather than a conversationalist. Unlike a general chat model, Devstral was trained on real agentic harnesses (OpenHands, Cline, Aider, and the same tool loop Atlas drives), so it fixes issues in a repository instead of answering questions about one. It runs 128,000 tokens in and 128,000 tokens out at $0.10 / 1M input tokens and $0.30 / 1M output tokens, cheap enough to drive a whole day of edits.
What is Devstral Small 2505 and why was it built for agents?
Devstral Small 2505, released by Mistral AI in May 2025, is the model that started the Devstral line. Unlike a general chat model, Devstral Small 2505 was trained on real agentic harnesses (OpenHands, Cline, Aider, and the same tool loop Atlas drives), so it resolves issues in a repository rather than answering questions about one.
Agent-first training shows up as behavior, not benchmark trivia. A chat-trained model asked to fix a bug will explain the bug. Devstral Small 2505 will open files, call tools, and produce a patch, because that is the shape of the data it learned on. That is exactly the loop Atlas runs: Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and the model is expected to drive toward that diff rather than toward a paragraph. For Atlas users coming from a general model, the change in tone is immediate and slightly startling.
How much does Devstral Small 2505 cost to run in Atlas all day?
Devstral Small 2505 costs $0.10 / 1M input tokens and $0.30 / 1M output tokens, with Apache 2.0 weights. At that price it is cheap enough to run as the default Atlas driver for a whole day of edits rather than being reserved for a handful of important turns.
The all-day economics are the point of Devstral Small 2505. Frontier models push developers into rationing: plan carefully, spend the expensive turn well. At $0.10 / 1M input tokens the rationing disappears and you can let Atlas iterate. The 128,000 tokens in and symmetric 128,000 tokens out means large multi-file patches are not clipped mid-diff, which is the failure mode that turns a cheap model into an expensive one through retries. Atlas snapshots file changes as git patches so edits can be diffed and rolled back, so an aggressive cheap model is a safe cheap model.
Why set Atlas tool permissions to ask when running Devstral Small 2505?
Set Atlas write permissions to `ask` when you first run Devstral Small 2505. Devstral Small 2505 is trained to act, not to deliberate, and every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, so `ask` puts a human in the loop on the first day.
The agent-first training that makes Devstral Small 2505 productive also makes it eager. It will reach for the edit tool early. Atlas already gates every edit behind a diff review, and pairing that with an `ask` permission rule on writes gives you two checkpoints while you calibrate how much autonomy the model has earned on your codebase. Once the diffs stop surprising you, loosen the rules. Starting at `deny` for anything destructive and `ask` for writes is the standard shape, and it costs nothing but a keystroke per turn.
Should you still pin Devstral Small 2505 in 2026?
Pin Devstral Small 2505 only if you need the exact 2505 weights. Devstral Small 2505 has been superseded twice, by Devstral Small 2507 and Devstral Small 2, both cheaper or more capable, so a new Atlas project should generally start on a later snapshot.
There are real reasons to hold the pin. Reproducibility is one: if an evaluation suite, a fine-tuned prompt, or a compliance record was built against Devstral Small 2505, moving weights moves your baseline. Apache 2.0 licensing is another, since the exact snapshot can be archived and rehosted. What is not a good reason is inertia. Atlas lets you switch the active model and provider on the fly with favorites and recents, so trying the newer snapshot against your own repository takes one session, not a migration project.
When is Devstral Small 2505 the wrong model for an Atlas session?
Devstral Small 2505 is the wrong model when the conversation is the work. Devstral Small 2505 is narrow by design and is worse than Mistral Medium at open-ended architecture discussion, so a design review or a written tradeoff analysis belongs on a general model, not on a resolver.
Knowing the boundary keeps the model useful. Devstral Small 2505 excels when the ticket is clear and the repository is the thing that has to change. It gets thin when nobody has decided what should change yet. A practical split inside Atlas: draft the plan on a general or reasoning model, then hand the plan to Devstral Small 2505 for execution at $0.10 / 1M input tokens. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, which makes that handoff a built-in step rather than a workaround.
Setup
- 01Export MISTRAL_API_KEY in the shell that launches Atlas.
- 02Run `atlas models mistral` and confirm devstral-small-2505 is in the registry.
- 03Pin the exact dated snapshot in atlas.json: `"model": "mistral/devstral-small-2505"`.
- 04Set tool permissions to `ask` for writes at first: Devstral is trained to act, and Atlas gates every edit behind a diff review.
- 05If you have no reason to hold the 2505 weights, compare against Devstral Small 2507 before committing.
Frequently asked questions
- what is devstral small 2505
- Devstral Small 2505 is the May 2025 model from Mistral AI that started the Devstral line. It was trained on real agentic harnesses including OpenHands, Cline, and Aider, so it resolves issues in a repository rather than answering questions about one.
- how much does devstral small 2505 cost
- Devstral Small 2505 costs $0.10 / 1M input tokens and $0.30 / 1M output tokens, with a 128,000 token context window and Apache 2.0 weights. That is cheap enough to run as the default Atlas driver all day.
- is devstral small 2505 still worth using in 2026
- Devstral Small 2505 has been superseded twice, by Devstral Small 2507 and Devstral Small 2, both cheaper or more capable. Pin 2505 only when you need the exact 2505 weights, for example to hold an evaluation baseline steady.
- how do I pin devstral small 2505 in atlas
- Export MISTRAL_API_KEY, run `atlas models mistral` to confirm devstral-small-2505 appears, then pin the exact dated snapshot in atlas.json with `"model": "mistral/devstral-small-2505"`.
- why does devstral write code instead of explaining it
- Devstral Small 2505 was trained agent-first on real coding scaffolds rather than on chat transcripts, so its default behavior is to call tools and produce a patch. Atlas surfaces that patch as a unified diff for approval before anything is written.
- what permissions should I set for devstral in atlas
- Set writes to `ask` when starting on Devstral Small 2505. Every Atlas tool call is permission-gated against allow, ask, and deny rules before it runs, and Devstral is trained to act, so a human checkpoint on the first sessions is worth the keystroke.
- is devstral small 2505 good for architecture discussions
- No. Devstral Small 2505 is narrow by design and is worse than Mistral Medium at open-ended architecture discussion. Plan on a general or reasoning model, then hand the plan to Devstral Small 2505 for the build pass.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs Fine.dev: Terminal AI Coding Agents in 2026
Compare Atlas and Fine.dev for terminal AI coding in 2026. Atlas offers a free core TUI with permission-gated tools, while Fine.dev provides asynchronous cloud agents and prebuilt libraries.
Atlas vs Sourcegraph Cody: Terminal AI Coding Agents in 2026
Comparing Atlas, the terminal-native AI coding agent, with Sourcegraph Cody, an editor extension for whole-codebase search, for developers in 2026. Explore features, pricing, and workflow.
Diagnose a Hanging or Long-Running Command with Atlas in 2026
How to diagnose a hanging command with Atlas in 2026: the bash tool races every command against a timeout and tells you whether it is slow or blocked on input.
Atlas vs Qodo: Choosing Your AI Coding Agent in 2026
Comparing Atlas, the terminal-native AI coding agent, with Qodo 2.0, the multi-agent PR reviewer, for developers in 2026. Evaluate features, pricing, and workflow.
Atlas vs Pieces for Developers: AI Tools for Developers in 2026
Comparing Atlas, a terminal-native AI coding agent, with Pieces for Developers, an OS-level memory layer, for developers in 2026. Evaluate code generation, safety, and context management.
Plan a Multi-File Change Before Editing with Atlas in 2026
How to plan a multi-file change with Atlas in 2026: the plan agent denies all edit tools, you research with codebase_search and lsp, then plan_exit hands off.
Atlas vs GitHub Copilot CLI: A Developer's Guide for 2026
Comparing Atlas and GitHub Copilot CLI for developers in 2026. Explore pricing, privacy, agentic workflows, and extensibility to choose the best terminal AI coding agent.
Atlas for Haskell in 2026
Atlas is a terminal-native AI coding agent for Haskell in 2026. Run it in a project with a .cabal file or a stack.yaml, let cabal build drive it, and review the diff.