MiniMax-M3, released June 2026, is the jump to a 1,000,000 token context while holding the MiniMax line's $0.30 per Mtok input and $1.20 per Mtok output pricing. That combination, a million tokens of context at thirty cents per million, is the cheapest large-context reasoning offer in this comparison. Inside Atlas, MiniMax-M3 is the model for whole-repository work: architecture surveys, cross-cutting refactors, and any task where retrieval tuning keeps failing because the model simply needs to see more.
What can Atlas do with MiniMax-M3's 1,000,000 token context?
MiniMax-M3 gives Atlas a 1,000,000 token context at $0.30 per Mtok input, matching DeepSeek V4 Flash's window at roughly twice the input price but with a far lower output rate than most frontier models. A million tokens changes what an Atlas session can hold at once.
At normal frontier pricing, a million-token prompt is a thing you do once and then explain to your finance team. On MiniMax-M3 at $0.30 per Mtok input, filling the entire 1,000,000 token window costs thirty cents. That reframes the retrieval strategy completely. Atlas searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion and indexes code by AST declarations using tree-sitter, not blind line windows, so with a 1M window you let that machinery pull broad context instead of tuning it to be frugal. At $0.30 per Mtok, the input cost of being generous is negligible.
How does MiniMax-M3 expand context without raising the price?
MiniMax-M3 delivers a 5x context expansion over MiniMax-M2.7's 204,800 tokens with zero price increase, holding $0.30 per Mtok input and $1.20 per Mtok output. A jump of that size at a flat price is rare in this market and is the main reason to move an Atlas config to MiniMax-M3.
For an Atlas user the migration is one line: `"model": "minimax/MiniMax-M3"` in atlas.json, after confirming with `atlas models minimax` that MiniMax-M3 appears in the registry. Nothing about the billing model changes, which means there is no budget conversation attached to the upgrade. What does change is which tasks are now possible. A cross-cutting rename that previously required Atlas to chunk the repository into several sessions can run as one MiniMax-M3 session, and Atlas snapshots file changes as git patches so edits can be diffed and rolled back if the wide pass goes wrong.
What is MiniMax-M3's output budget, and is it a downgrade?
MiniMax-M3 has a 128,000 max output with reasoning enabled, which is enough headroom for a long plan plus a complete implementation in one turn. Honestly stated, 128,000 is a small step down from MiniMax-M2.7's 131,072, so the bigger window comes with a slightly tighter write budget.
The 3,072 token difference is unlikely to be the thing that breaks a session, but it is worth knowing before you standardize on MiniMax-M3 for large write passes. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, and a 128,000 token output cap on MiniMax-M3 still produces a very large diff in one response. The asymmetry is deliberate: MiniMax-M3 is optimized to read widely rather than to write more. If a task is dominated by generating output rather than absorbing context, MiniMax-M2.7's 131,072 cap is marginally better suited.
What is slow about running Atlas on MiniMax-M3?
Filling a million-token prompt is slow regardless of price, and MiniMax's China hosting adds network latency for US and EU users. MiniMax-M3's 1,000,000 token window is a throughput cost as much as a capability, and an interactive Atlas session that habitually fills it will feel sluggish.
Use MiniMax-M3 deliberately rather than by default. The right pattern is to reach for MiniMax-M3 when a task genuinely needs breadth, a survey pass, an architecture question, a refactor whose blast radius you do not yet know, and to run the resulting tight edit loop on a smaller-window MiniMax model where turnaround is quick. Atlas lets you switch the active model and provider on the fly with favorites and recents, so this is a `/models` keystroke, not a config rewrite. The data residency point stands separately: MiniMax is China-hosted, which is a policy decision before it is a latency one.
When should you not use MiniMax-M3 in Atlas?
Skip MiniMax-M3 when your task fits inside 204,800 tokens, which describes most day-to-day agent work. MiniMax-M2.7 costs the identical $0.30 per Mtok input and $1.20 per Mtok output, gives a slightly larger 131,072 output cap, and returns answers faster because there is less prompt to process.
The million-token window is a specialized tool, not a free upgrade in every dimension. A focused bug fix on a single service does not benefit from MiniMax-M3, and the latency of a large-context model is a real tax on an iterative loop. Two rules keep the choice simple. First, if Atlas's reciprocal rank fusion search already finds the right files, use the smaller model. Second, if you have tuned retrieval twice and the model still lacks the context to reason correctly, that is the signal to move to MiniMax-M3 and its 1,000,000 token window.
Setup
- 01Export MINIMAX_API_KEY or run `atlas login` and pick MiniMax.
- 02Run `atlas models minimax` and confirm `MiniMax-M3` appears in the registry.
- 03Set `"model": "minimax/MiniMax-M3"` in atlas.json.
- 04With a 1M window, let Atlas's tree-sitter AST chunking and reciprocal rank fusion search pull broad context. At $0.30 per Mtok input the cost of being generous is negligible.
- 05Favorite a smaller-window MiniMax model alongside MiniMax-M3 so you can drop back to a faster tier for tight edit loops.
Frequently asked questions
- what is the minimax-m3 context window
- MiniMax-M3 has a 1,000,000 token context window with a 128,000 max output and reasoning enabled. That is a 5x expansion over MiniMax-M2.7's 204,800 tokens.
- how much does minimax-m3 cost per million tokens
- MiniMax-M3 costs $0.30 per Mtok input and $1.20 per Mtok output, the same pricing as MiniMax-M2.7 despite the 1,000,000 token window.
- cheapest model with a 1 million token context window
- MiniMax-M3 offers a 1,000,000 token context at $0.30 per Mtok input and $1.20 per Mtok output, which is the cheapest large-context reasoning offer in this comparison.
- does minimax-m3 have a smaller output limit than m2.7
- Yes. MiniMax-M3 caps output at 128,000 tokens, a small step down from MiniMax-M2.7's 131,072, so the bigger context window comes with a slightly tighter write budget.
- how do i run atlas on minimax-m3
- Export MINIMAX_API_KEY or run `atlas login` and pick MiniMax, run `atlas models minimax` to confirm `MiniMax-M3` appears, then set `"model": "minimax/MiniMax-M3"` in atlas.json.
- is minimax-m3 slow
- Filling a million-token prompt is slow regardless of price, and MiniMax's China hosting adds network latency for US and EU users. Use MiniMax-M3 for breadth-heavy tasks and a smaller model for tight edit loops.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs. Goose: Choosing Your AI Coding Agent in 2026
Compare Atlas and Goose for 2026. Atlas offers terminal-native TUI and code-specialized features. Goose provides shareable Recipes and 70+ MCP extensions for general agentic workflows.
Atlas for Go in 2026
Atlas, the terminal-native AI coding agent, empowers Go developers in 2026 with intelligent code understanding, safe refactoring, and robust testing capabilities.
Atlas vs OpenAI Codex CLI: Terminal AI Coding Agents in 2026
Comparing Atlas and OpenAI Codex CLI for terminal AI coding in 2026. Atlas offers a TUI, permission-gated tools, and BYO models, while OpenAI Codex CLI provides sandboxed execution and OpenAI model integration.
Atlas for Next.js in 2026
Adopt Atlas, the terminal-native AI coding agent, for Next.js development in 2026. Enhance productivity across App Router, server components, and API routes with secure, reviewable AI assistance.
Atlas for Erlang in 2026
Atlas is a terminal-native AI coding agent for Erlang/OTP in 2026. Run it in an app with a rebar.config, map supervisors and gen_server modules, review every diff.
Atlas for Crystal: A Terminal-Native AI Coding Agent for shard.yml Projects in 2026
Atlas is a terminal-native AI coding agent for Crystal in 2026. It reads shard.yml and union types, narrows Nil unions, runs crystal spec behind a prompt, and formats the diff.
Atlas for Assembly: Registers, Calling Conventions, and nasm in 2026
Atlas is a terminal-native AI coding agent for Assembly in 2026. It reads .asm and .S sources, tracks System V and AAPCS64 calling conventions, and assembles with nasm behind a prompt.
Atlas for Laravel in 2026
Atlas is a terminal-native AI coding agent for Laravel in 2026. Run it in a Laravel app with an artisan file, add a controller or Pest test, and review the diff.