MiniMax-M2.5, released February 2026, keeps the MiniMax line's signature $0.30 per Mtok input and $1.20 per Mtok output pricing and its 204,800 token context while advancing the underlying 230B efficient-MoE. Inside Atlas, MiniMax-M2.5 is the model you can put in both the main slot and the small model slot, because at thirty cents per million input tokens there is almost nothing left to save by downgrading. If you want a current-generation reasoning model for well under a dollar per million input tokens, MiniMax-M2.5 is on the shortlist.
What does MiniMax-M2.5 do well inside Atlas?
MiniMax-M2.5 holds the 230B efficient-MoE architecture at $0.30 per Mtok input, matching MiniMax-M3's input price while predating it by four months. Inside Atlas, MiniMax-M2.5 handles the full agent loop, from the read-only plan agent through the build agent, without a cost tier change in the middle.
Reasoning is on by default for MiniMax-M2.5, and the 131,072 max output tokens mean plan-then-implement fits into a single response rather than spanning turns. That single-response property matters more than it sounds. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and when the model can hold a long plan and a complete diff in one output budget, the handoff between those agents stays clean. Atlas then computes a unified diff for every file edit and surfaces it for approval before writing, so what you review on MiniMax-M2.5 is a whole change rather than the first half of one.
Should you use MiniMax-M2.5 for both model and small_model in atlas.json?
Yes. Use MiniMax-M2.5 in both slots, `"model"` and `"small_model"`, since at $0.30 per Mtok input and $1.20 per Mtok output there is little to gain from a separate cheap model. The usual reason to split slots, protecting yourself from an expensive main model, does not apply to MiniMax-M2.5.
The small model slot exists because agents make a lot of low-stakes calls: titles, summaries, and the background subagent work Atlas fans out in foreground or parallel background sessions. On a $4.00 per Mtok output model, routing those calls to a cheap tier is the difference between a manageable bill and a bad one. MiniMax-M2.5 is already the cheap tier. Running the same model in both slots also removes a category of confusion, because every call in the session shares one context limit, one output cap, and one behavior profile, which makes an unexpected result easier to attribute.
What are the limits of MiniMax-M2.5's 204,800 token window?
MiniMax-M2.5 gives Atlas a 204,800 token context window and 131,072 max output tokens. The window is large enough for a substantial service and its tests, but MiniMax-M3 offers 1,000,000 tokens at the identical $0.30 per Mtok input, so MiniMax-M2.5 is not the choice for whole-repository reasoning.
Where the 204,800 token window pays off is precision rather than breadth. Atlas indexes code by AST declarations using tree-sitter, not blind line windows, so the tokens that reach MiniMax-M2.5 are complete declarations rather than arbitrary slices. Combined with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, a 204,800 token budget covers a feature and its blast radius comfortably. What it does not cover is a survey of an entire monorepo in one pass. For that specific job, and only that job, the larger MiniMax window is the answer.
When is MiniMax-M2.5 the wrong pick?
MiniMax-M2.7 is the newer checkpoint at identical $0.30 per Mtok input and $1.20 per Mtok output pricing and the same 204,800 token context, so MiniMax-M2.5 is now a fallback rather than a first choice. Nothing about the money changes when you move to M2.7.
The second limitation is throughput. At the standard tier, MiniMax-M2.5 serves more slowly than the highspeed variant, and that difference is noticeable in an interactive TUI session where you are watching reasoning tokens arrive. MiniMax offers a same-context sibling, MiniMax-M2.5-highspeed at $0.60 per Mtok input and $2.40 per Mtok output, so you can trade cost for latency by editing one id. The rule of thumb is straightforward: keep MiniMax-M2.5 for unattended runs, batch work, and background subagents, and pay double only when a person is actually waiting on the tokens.
Setup
- 01Export MINIMAX_API_KEY, or run `atlas login` and select MiniMax.
- 02Run `atlas models minimax` and confirm `MiniMax-M2.5` appears in the registry.
- 03Set `"model": "minimax/MiniMax-M2.5"` in atlas.json. Atlas routes MiniMax through `@ai-sdk/anthropic`.
- 04Use MiniMax-M2.5 in both slots, `"model"` and `"small_model"`, since at $0.30 per Mtok input and $1.20 per Mtok output there is little to gain from a separate cheap model.
- 05If an interactive session feels slow, switch the id to `minimax/MiniMax-M2.5-highspeed` and pay $0.60 and $2.40 for the same weights.
Frequently asked questions
- how much does minimax-m2.5 cost per million tokens
- MiniMax-M2.5 costs $0.30 per Mtok input and $1.20 per Mtok output, the signature pricing of the MiniMax line, with reasoning enabled at that rate.
- what is minimax-m2.5 context window
- MiniMax-M2.5 has a 204,800 token context window and 131,072 max output tokens, which is enough for a plan and a complete implementation in one response.
- minimax-m2.5 vs minimax-m2.7 for coding
- MiniMax-M2.7 is the newer checkpoint at identical $0.30 and $1.20 pricing with the same 204,800 token context. MiniMax-M2.5 is therefore a fallback rather than a first choice.
- can i use minimax-m2.5 as the small model in atlas
- Yes. At $0.30 per Mtok input and $1.20 per Mtok output, MiniMax-M2.5 is cheap enough for both `"model"` and `"small_model"` in atlas.json, so a separate cheap tier buys you very little.
- how do i set up minimax-m2.5 with atlas
- Export MINIMAX_API_KEY or run `atlas login` and select MiniMax, run `atlas models minimax` to confirm `MiniMax-M2.5` appears, then set `"model": "minimax/MiniMax-M2.5"` in atlas.json.
- is minimax-m2.5 fast enough for interactive use
- At the standard tier MiniMax-M2.5 has lower throughput than the highspeed variant, which is noticeable in an interactive TUI session. MiniMax-M2.5-highspeed offers the same context at $0.60 and $2.40 if latency matters.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs Roo Code: Terminal AI Coding Agents in 2026
Comparing Atlas and Roo Code in 2026. Atlas offers terminal-native TUI, permission-gated tools, and diff review. Roo Code, a VS Code extension, shut down May 15, 2026.
Atlas for PyTorch: Terminal-Native AI Coding for nn.Module, Devices, and Autograd in 2026
Atlas is a terminal-native AI coding agent for PyTorch in 2026, where device placement, autograd, and DataLoader worker counts cause most bugs and most slowness.
Atlas for Nuxt: Auto-Imports, useAsyncData, and Nitro Handlers in 2026
Atlas is a terminal-native AI coding agent for Nuxt in 2026. It reads nuxt.config.ts, pages/ routes, composables/ auto-imports, and server/api/ Nitro handlers, and tests with @nuxt/test-utils.
Extract a Shared Helper from Duplicated Code with Atlas (2026 Workflow)
How to extract a shared helper from duplicated code with Atlas in 2026: codebase_search finds the copies by meaning, write creates the module, apply_patch swaps each call.
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.
Atlas for Rust in 2026
Adopt Atlas, the terminal-native AI coding agent, for Rust development in 2026. Tackle borrow checker errors and clippy lints with Atlas's secure, approval-gated assistance.
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.
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.