What MiniMax M3 Is
MiniMax M3 is the flagship reasoning model from MiniMax, built for agentic workflows, long-document processing, coding tasks, and complex multi-step reasoning at scale. It supports a context window of up to 1 million tokens — one of the largest available in production — and introduces multimodal input support alongside a reasoning architecture designed for frontier-tier performance.
Where previous MiniMax models excelled at conversational tasks and general instruction-following, M3 was purpose-built for environments where context depth, tool-use reliability, and sustained reasoning across long sessions actually matter. It's the first MiniMax model explicitly positioned at the agentic frontier.
Key Specs
minimax/minimax-m3MiniMax M3 Pricing
MiniMax M3 (≤512K input tokens):
Note: Current pricing is a MiniMax promo price (7-day 50% off).
MiniMax M3 (>512K input tokens):
Note: Input tokens above 512K are currently available in limited mode. Public availability is expected later.
What M3 Is Actually Built For
The 1M context window is not a marketing number — it changes what kinds of tasks are actually feasible. M3 was designed around workflows where the bottleneck is context capacity, not model intelligence: entire codebases, full research papers, lengthy conversation histories, large tool-call chains.
Agentic Workflows
Plan, execute, and iterate across multi-step tasks without losing coherence over long sessions. M3's reasoning architecture handles task decomposition, tool selection, and self-correction loops across hundreds of turns — the kind of sustained performance that breaks down in shorter-context models.
Code Generation & Engineering
Generate, review, and refactor code with full awareness of large codebases. Feed in entire repositories or multi-file projects without chunking. Useful for code migration, security audits, architecture reviews, and CI-integrated generation pipelines.
Long-Document Understanding
Feed in full contracts, research papers, financial reports, or technical specifications — not summaries, the actual documents. M3's 1M token window makes it one of the few models where "just send the whole thing" is a real strategy, not a workaround.
Multimodal Reasoning
Accepts image input alongside text, enabling document parsing with embedded figures, diagram interpretation, screenshot-based debugging, and visual context in agentic pipelines.
How It Stacks Up
M3 is positioned against Claude Opus 4.6, GPT-5, and Gemini 2.5 Pro in the reasoning-and-agentic tier. Its primary differentiators are context length and cost: at $0.39/$1.56 input/output for the base tier, it offers frontier-class context at a lower price point than most comparable models. Where it may trail is in narrow domain depth — specialized medicine, law, and finance benchmarks favor models with more training weight in those verticals.
Who Should Use M3 via API?
// 01 Agent Framework DevelopersIf you're building multi-step agent systems that need to hold large tool call histories, system prompts, and conversation context simultaneously — M3's 1M window removes the context management overhead that breaks most agent loops at scale.
// 02 Legal & Contract AnalysisOrganizations processing high-volume contracts, compliance documents, or regulatory filings benefit from sending full documents rather than chunking. M3 handles full-length legal documents as single-pass inputs.
// 03 Research & Data PipelinesTeams running evaluation pipelines, literature reviews, or large-scale data extraction workflows can feed M3 entire datasets or document collections in a single request.
// 04 Coding Infrastructure TeamsCodebases that are too large for standard context windows become tractable with M3. Repository-level code review, cross-file refactoring, and full-codebase Q&A are practical use cases at the 1M token level.
// 05 Cost-Conscious Frontier UsersFor teams currently paying Opus 4.6 or GPT-5 rates for reasoning-heavy workloads, M3's base tier pricing at $0.39 input / $1.56 output makes it worth a direct comparison before committing to higher-cost alternatives.
What You Should Know Before Committing
M3's >512K token tier is currently available in limited mode — if your use case depends on consistently sending inputs above 512K tokens, confirm availability before building a production dependency. Tool calling support is expected to be OpenAI-compatible but should be validated in your specific environment before relying on it in automated pipelines.
Measured on the AI/ML API gateway across live traffic. Figures are medians unless stated. Last updated 2026-10-12.
MiniMax M3 is documented as suited to coding tasks, including bug fixing, frontend/backend development, performance optimization, and long-horizon coding work. It is also suited to agentic workflows, including autonomous task decomposition, tool invocation, multi-step reasoning, and automated workflows. It is documented for multimodal use cases such as long-form video understanding, image and video input, and computer use on desktop applications. It is also described as suitable for office-style workflows and design or creative workflows.
import requests
r = requests.post(
"https://api.aimlapi.com/v1/chat/completions",
headers={"Authorization": "Bearer " + AIMLAPI_KEY},
json={
"model": "minimax/minimax-m3",
"messages": [
{
"role": "user",
"content": "Hello!"
}
]
},
)
print(r.json())
const r = await fetch("https://api.aimlapi.com/v1/chat/completions", { method: "POST", headers: { Authorization: `Bearer ${process.env.AIMLAPI_KEY}`, "Content-Type": "application/json", }, body: JSON.stringify({ "model": "minimax/minimax-m3", "messages": [ { "role": "user", "content": "Hello!" } ] }), }); console.log(await r.json());
curl -X POST https://api.aimlapi.com/v1/chat/completions \ -H "Authorization: Bearer $AIMLAPI_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"minimax/minimax-m3","messages":[{"role":"user","content":"Hello!"}]}'
OpenAI-compatible — swap the base URL and it works with your existing SDK.
| Type | Price |
|---|---|
| Input | |
| Output | |
| Cached input |
| Benchmark | Score | What it measures | Source | Retrieved |
|---|---|---|---|---|
| SWE-bench Verified | 80.5% | Resolving verified real GitHub issues | Source | July 12, 2026 |
| Intelligence | 29.6 | Composite score across standardised reasoning, knowledge and problem-solving evaluations, measured independently by Artificial Analysis | Source | September 12, 2026 |
| Coding | 58.6 | Composite score across standardised coding evaluations, measured independently by Artificial Analysis | Source | September 12, 2026 |
| Model | Input | Output | Context | Best for |
|---|---|---|---|---|
MiniMax M3 This page | Coding + agents | |||
| Reasoning + agents | ||||
| Balanced coding + agents | ||||
| Long-context, multimodal & agentic workflows | ||||
| Reasoning + agents |
MiniMax M3 has a 524,288 tokens context window and can return up to 524,288 tokens.
MiniMax M3 takes image, text as input and returns text.
Use minimax/minimax-m3 as the model id. Requests go to https://api.aimlapi.com/v1/chat/completions.
MiniMax M3 became available on June 1, 2026.
MiniMax M3 is priced at input $0.39 / 1M tokens, output $1.56 / 1M tokens, cached input $0.078 / 1M tokens.
Yes, MiniMax M3 can stream responses as they are generated.
Yes, MiniMax M3 accepts image input alongside text.
MiniMax M3 was built by MiniMax.
Yes, it supports function calling along with parallel tool calls and structured outputs.
