MiniMax-M2.1 is the December 2025 refresh of MiniMax-M2. The price stayed flat at $0.30 per Mtok input and $1.20 per Mtok output, but the window grew from 196,608 to 204,800 tokens and the output cap from 128,000 to 131,072. Inside Atlas, MiniMax-M2.1 is a reasoning model cheap enough that you stop rationing context, and the upgrade from M2 is free in every dimension that shows up in a config file.
What changed between MiniMax-M2 and MiniMax-M2.1?
MiniMax-M2.1 arrived in December 2025 as a refresh of M2 with the price held flat at $0.30 per Mtok input and $1.20 per Mtok output. The context window grew from 196,608 to 204,800 tokens and the output cap from 128,000 to 131,072, so every measurable spec improved at zero cost.
Upgrades that cost nothing are worth taking without a long evaluation. For an Atlas user the change is one string in atlas.json: pin `"model": "minimax/MiniMax-M2.1"` and the session gains 8,192 tokens of context and 3,072 tokens of output headroom. The larger output cap is the more useful half. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and a 131,072 token ceiling on MiniMax-M2.1 means a long plan plus a full implementation lands in a single response instead of spilling into a continuation turn that you then have to review twice.
How does MiniMax-M2.1 pricing compare to GLM-4.7 and Kimi K2.6?
MiniMax-M2.1 charges $0.30 per Mtok input, which is half of GLM-4.7's input price for a comparable window. On output, MiniMax-M2.1's $1.20 per Mtok is roughly half of GLM-4.7's $2.20 and under a third of Kimi K2.6's $4.00, with reasoning enabled at that rate.
The output rate is where a coding agent's bill is actually decided, because a reasoning model spends most of its tokens thinking. At $1.20 per Mtok output, MiniMax-M2.1 makes a long chain of thought affordable in a way that a $4.00 model does not. That shifts how you configure Atlas. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and on a $4.00 output model that fan-out is the thing you carefully throttle. On MiniMax-M2.1 you can let the subagents run, because three parallel background sessions still cost less than one session on Kimi K2.6.
How should you configure Atlas retrieval on MiniMax-M2.1?
Because MiniMax-M2.1 input is only $0.30 per Mtok, turn Atlas's hybrid semantic plus keyword code search loose and let it pull generous context rather than hand-picking files. Filling 100,000 of MiniMax-M2.1's 204,800 token window costs three cents, which is less than the developer time spent choosing files by hand.
Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion, and it indexes code by AST declarations using tree-sitter, not blind line windows. Both properties reward a generous token budget: the retrieved context is whole declarations, and the fusion step means precision does not collapse when you widen recall. On an expensive model you fight this by narrowing the search. On MiniMax-M2.1 the correct instinct is the opposite. Give the model the caller, the callee, and the test file, and let the reasoning do the filtering rather than doing it yourself in advance.
When should you not use MiniMax-M2.1?
MiniMax-M2.5 and MiniMax-M2.7 sit at the identical $0.30 per Mtok input and $1.20 per Mtok output price with the same 204,800 token context, so there is no cost reason to stay on MiniMax-M2.1. The only argument for M2.1 is a preference for the older, more settled checkpoint.
The second reason to look elsewhere is jurisdictional. MiniMax is China-hosted, and MiniMax-M2.1 is served behind an Anthropic-shaped API, which is convenient but is still a compatibility surface: it can lag real Anthropic on newer tool-streaming behavior. If your organization cannot send source code to a China-hosted endpoint, no price makes MiniMax-M2.1 the right answer. Atlas does offer a partial mitigation, since Atlas can build its code index with local Ollama embeddings, keeping code off third-party servers, but the chat traffic itself still goes to MiniMax.
Is the Anthropic-shaped MiniMax API a problem for Atlas?
MiniMax-M2.1 speaks an Anthropic-shaped API, which means that if you already have Atlas wired for Claude, the provider block is a base URL and key swap rather than a rewrite. The tradeoff is that the compatibility surface can lag real Anthropic on newer tool-streaming behavior.
In day-to-day Atlas use the shim is nearly invisible. Run `atlas models minimax`, locate `MiniMax-M2.1`, pin it in atlas.json, and the agent loop behaves normally. Where it shows up is at the edges: a streaming or tool-call feature that Atlas exercises against real Claude may not have an exact counterpart on the MiniMax endpoint. The practical habit is to treat any odd tool-streaming behavior as a compatibility gap first and a model quality issue second, and to keep Atlas's permission gating on allow, ask, and deny rules so that a malformed tool call is caught before it runs anything.
Setup
- 01Export MINIMAX_API_KEY or run `atlas login` and choose MiniMax.
- 02Run `atlas models minimax` and locate `MiniMax-M2.1` in the resolved registry.
- 03Pin `"model": "minimax/MiniMax-M2.1"` in atlas.json.
- 04Because input is only $0.30 per Mtok, turn Atlas's hybrid semantic plus keyword code search loose and let it pull generous context rather than hand-picking files.
- 05Keep Atlas's allow, ask, and deny permission rules active so tool calls from the Anthropic-shaped endpoint are gated before they run.
Frequently asked questions
- minimax-m2.1 vs minimax-m2 what is the difference
- MiniMax-M2.1 is the December 2025 refresh of M2. The price stayed at $0.30 per Mtok input and $1.20 per Mtok output, while the context grew from 196,608 to 204,800 and the output cap from 128,000 to 131,072.
- how much does minimax-m2.1 cost
- MiniMax-M2.1 costs $0.30 per Mtok input and $1.20 per Mtok output with reasoning enabled, which is roughly half GLM-4.7's $2.20 output rate.
- what is the minimax-m2.1 context window and max output
- MiniMax-M2.1 has a 204,800 token context window and a 131,072 token max output, enough for a plan and a complete implementation in one response.
- how do i run atlas on minimax-m2.1
- Export MINIMAX_API_KEY or run `atlas login` and choose MiniMax, run `atlas models minimax` to locate `MiniMax-M2.1`, then pin `"model": "minimax/MiniMax-M2.1"` in atlas.json.
- should i use minimax-m2.1 or minimax-m2.5
- MiniMax-M2.5 costs the identical $0.30 and $1.20 with the same 204,800 token context, so there is no cost reason to stay on MiniMax-M2.1. Pick M2.1 only if you prefer the older checkpoint.
- is minimax-m2.1 hosted in china
- Yes. MiniMax is China-hosted, and MiniMax-M2.1 is served behind an Anthropic-compatible surface. Teams with data residency requirements should weigh that before pinning it in atlas.json.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs Gemini CLI: A Developer's Guide to Terminal AI Agents in 2026
Atlas vs Gemini CLI in 2026: Compare terminal AI coding agents. Atlas offers permission-gated tool calls and diff-based approvals. Gemini CLI provides a million-plus token context and a free tier.
Run the Test Suite and Triage the Failures with Atlas in 2026
How to triage a failing test suite with Atlas in 2026: bash truncates at 2000 lines or 50 KB and saves the full log, then grep groups failures by root cause.
Atlas for Polars: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for Polars. Build LazyFrame chains, push scan_csv predicates into the reader, and read explain() plans in 2026.
Refactor a Legacy Module with Atlas in 2026
How to refactor a legacy module with Atlas in 2026: findReferences maps every callsite, apply_patch refuses to apply against a drifted file, and bash proves behavior.
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 for C in 2026
Atlas is a terminal-native AI coding agent for C in 2026. Run it in a project with a Makefile, have it find memory leaks or add Unity tests, and review the diff.
Atlas for Clojure: A Terminal-Native AI Coding Agent for deps.edn and Kaocha in 2026
Atlas is a terminal-native AI coding agent for Clojure in 2026. It reads deps.edn aliases and namespace requires, runs clojure -M:test with Kaocha, and applies cljfmt.
Atlas vs Qwen Code: Terminal AI Coding Agents in 2026
Atlas and Qwen Code are leading terminal AI coding agents in 2026. Compare Atlas's TUI, permission-gated tools, and free core with Qwen Code's 1M token context, Agent Teams, and Alibaba Cloud hosting.