# Atlas with Mistral Small 3.2 (2506): The Cheap Slot Done Right in 2026 > Mistral Small 3.2 (2506) costs $0.10 / 1M input tokens with a 128,000 token window, which makes it the natural small_model slot for an Atlas session. Mistral Small 3.2 (2506) is the natural small_model slot for an Atlas session. At $0.10 / 1M input tokens and $0.30 / 1M output tokens with a 128,000 token window, it is cheap enough to absorb every session title, summary, and subagent call. The June 2025 update targeted the three failure modes that break coding agents: bad function calling, sloppy instruction following, and repetition loops. Its output is capped at 16,384 tokens, so it cannot emit huge diffs. ## Key takeaways - Mistral Small 3.2 (2506) costs $0.10 / 1M input tokens and $0.30 / 1M output tokens, cheap enough for every Atlas title, summary, and subagent call. - The June 2025 release hardened function calling and cut repetition errors, the two failure modes that break a tool loop. - Input window is 128,000 tokens but output is capped at 16,384 tokens, so it cannot emit huge diffs. - 24B dense weights ship under Apache 2.0, so the same model can run on your own hardware. - On hard reasoning it is outclassed by Mistral Medium 3.1 and by the Magistral line, so keep it in the small_model slot. ## What is Mistral Small 3.2 (2506) best at inside Atlas? Mistral Small 3.2 (2506) is best inside Atlas as the small_model slot. At $0.10 / 1M input tokens it is cheap enough to run every title, summary, and subagent call, and the June 2025 release specifically hardened function calling and cut repetition errors, which matters when Atlas loops tools. An Atlas session generates a surprising amount of low-stakes traffic: naming the session, summarizing a long turn, and feeding subagents. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and every one of those fanned-out calls is a bill. Mistral Small 3.2 (2506) is sized for exactly that traffic. The 2506 release matters here more than the price does: agent harnesses die on malformed function calls and on models that repeat themselves until the turn times out, and those are the two things Mistral hardened. A cheap model that fails a tool schema is not cheap. ## Why is the 16,384 token output cap on Mistral Small 3.2 a problem for diffs? Mistral Small 3.2 (2506) caps output at 16,384 tokens, far below its 128,000 token input window. Atlas can therefore feed the model a large slice of a repository but cannot get a large patch back in one turn, so big edits get split across turns. The asymmetry is the single most important thing to plan around with Mistral Small 3.2 (2506). Reading 128,000 tokens and writing at most 16,384 means the model is a reader, not a writer, at scale. Watch the Atlas diff review pane: Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and when a change exceeds the output budget you will see the work arrive in pieces. For a multi-file refactor, that is friction. For a summary, a commit message, or a single-function fix, 16,384 tokens is more than enough. ## How much does Mistral Small 3.2 (2506) cost to run in Atlas? Mistral Small 3.2 (2506) costs $0.10 / 1M input tokens and $0.30 / 1M output tokens. Against a frontier driver in the primary slot, routing summarization and subagents to Small 3.2 is what keeps an all-day Atlas session from billing frontier rates on housekeeping turns. The economics only work if you actually use both Atlas model slots. Set a frontier model in `"model"` and let Atlas route summarization and subagents to Mistral Small 3.2 (2506) via `"small_model"`. The $0.30 / 1M output tokens rate is the one to watch, because subagent turns produce output, and a wide fan-out multiplies it. Keeping the context lean helps on both sides: Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, so a subagent can get the two relevant declarations instead of the whole file tree. ## Can Mistral Small 3.2 (2506) run on your own hardware? Yes. Mistral Small 3.2 (2506) ships 24B dense weights under Apache 2.0, so the same model that serves the API can run on your own hardware. For a team that cannot send source to a third party, that combination of open weights and a 128,000 token window is the draw. Self-hosting Mistral Small 3.2 (2506) pairs well with the rest of Atlas's local story. Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, so the retrieval layer and the model layer can both stay in-house. The Apache 2.0 license means no per-seat negotiation and no usage terms to read before a proof of concept. The catch is capability, not licensing: 24B dense is outclassed on hard reasoning by Mistral Medium 3.1 and by the Magistral line, and self-hosting does not change that. ## When should you pick a different model than Mistral Small 3.2 in Atlas? Pick a different model than Mistral Small 3.2 (2506) when the task is hard reasoning or a large patch. 24B dense is outclassed on hard reasoning by Medium 3.1 and by the Magistral line, and the 16,384 token output cap rules out single-turn rewrites of a whole module. Mistral Small 3.2 (2506) is a supporting model, and it is excellent in that role. Promote it to the primary slot only for small, well-scoped work. For multi-hop debugging, reach for a reasoning model. For a rewrite that has to land in one diff, reach for a model whose output budget matches its input window. Atlas lets you switch the active model and provider on the fly with favorites and recents, so keeping Mistral Small 3.2 (2506) in the cheap slot while cycling the primary model per task is a one-keystroke habit rather than a config rewrite. ## Setup 1. Export MISTRAL_API_KEY before launching Atlas. 2. Confirm the model resolves with `atlas models mistral`. 3. Set it as the cheap slot in atlas.json: `"small_model": "mistral/mistral-small-2506"`. 4. Keep a frontier model in `"model"` and let Atlas route summarization and subagents to Small 3.2. 5. Watch the diff review pane: the 16,384 token output cap means Atlas will split large edits across turns. ## FAQ ### what is mistral small 3.2 2506 good for Mistral Small 3.2 (2506) is built for the high-volume, low-stakes calls in an agent session: titles, summaries, and subagent turns. At $0.10 / 1M input tokens with a 128,000 token window it is the natural small_model slot for Atlas. ### what is the output token limit of mistral small 3.2 Mistral Small 3.2 (2506) caps output at 16,384 tokens, far below its 128,000 token input window. Inside Atlas that means large edits get split across turns rather than arriving as one diff. ### how do I set mistral small 3.2 as small_model in atlas Export MISTRAL_API_KEY, run `atlas models mistral` to confirm the model resolves, then set `"small_model": "mistral/mistral-small-2506"` in atlas.json while keeping a frontier model in `"model"`. ### is mistral small 3.2 open source Mistral Small 3.2 (2506) releases 24B dense weights under Apache 2.0, so the same model served by the API can be run on your own hardware alongside Atlas's local Ollama embedding index. ### does mistral small 3.2 handle function calling reliably The June 2025 update to Mistral Small targeted bad function calling, sloppy instruction following, and repetition loops specifically. Those are the failure modes that break coding agents, and hardening them is why 3.2 works as an Atlas subagent model. ### mistral small 3.2 vs mistral medium 3.1 for coding Mistral Small 3.2 (2506) at $0.10 / 1M input tokens is the cheap slot. Mistral Medium 3.1 is the stronger reasoner and writer, and 24B dense Small is outclassed by it on hard reasoning, so run Small as small_model and Medium as the driver. ### how much does mistral small 3.2 cost per million tokens Mistral Small 3.2 (2506) is priced at $0.10 / 1M input tokens and $0.30 / 1M output tokens, roughly the cheapest place to put an Atlas session's summarization and subagent traffic without giving up a 128,000 token window. --- Canonical HTML: https://runatlas.sh/resources/models/mistral-small-3-2 Source of truth: aeo_pages row `/resources/models/mistral-small-3-2` (segment: Models) (this file is generated from it, never hand-edited). Licence: Atlas is proprietary with a free core. It is not open source and there is no public source repository.