MiniMax M3 is worth testing as a value coding model, not treating as a proven replacement for Opus, GPT, Gemini, Kimi, or GLM. The value case is unusually strong on paper: MiniMax’s launch post documents a $20/month Plus Token Plan with ~1.7B M3 tokens/month, a 1M-token context window, native multimodal input, coding-agent positioning, and vendor-reported benchmark results near premium coding lanes.

That is enough to put MiniMax M3 on the shortlist. It is not enough to move production coding workflows without your own repo evals, latency checks, quota checks, and tool compatibility testing.

Best next click: Smart Spend Guide

Quick Facts

SpecMiniMax M3
ProviderMiniMax
Model IDMiniMax-M3
PositioningCoding, agents, long-context, native multimodal input
Context windowUp to 1M tokens, per MiniMax
Architecture claimMSA, MiniMax Sparse Attention
Subscription anchorPlus Token Plan: $20/month for ~1.7B M3 tokens/month, per MiniMax’s launch post
PAYG anchorStandard: $0.60 input / $2.40 output per 1M tokens for up to 512K input; $1.20 / $4.80 above 512K input. Priority is separate at $0.90 / $3.60 and $1.80 / $7.20.
Tool pathsMiniMax Code, Claude Code, OpenClaw, OpenCode, Cursor, Kilo Code, Cline, Roo Code, TRAE, others
Evidence posturePrimary-source specs and pricing; benchmarks are MiniMax-reported

Short Verdict

MiniMax M3 is a serious value candidate for coding-agent users because the Token Plan quota is large relative to premium API prices. The practical buying question is not “does it win a benchmark chart?” It is:

  1. Does M3 produce acceptable patches in your repo?
  2. Does the Token Plan key work cleanly in your preferred tool?
  3. Do rolling windows, weekly windows, latency, and output quality hold up under your real workload?
  4. Does it reduce premium GPT or Claude spend without increasing review cost?

If the answer is yes after testing, M3 can become a cheap daily coding lane. If the answer is no, it still may be useful as a secondary long-context or multimodal coding model.

What MiniMax Claims

MiniMax’s launch post says M3 is built around three capabilities:

  • MSA attention: MiniMax Sparse Attention, positioned as the mechanism that makes 1M context practical.
  • Coding and agentic work: MiniMax frames coding, terminal execution, multi-step collaboration, and long-horizon agent tasks as core training targets.
  • Native multimodality: MiniMax says M3 supports image and video input and can operate a desktop computer.
  • Open-weight plan: MiniMax calls M3 open-weight, but the same launch post says the technical report and weights will be released over the following 10 days. Until the weights and report are live and independently inspected, treat open-weight availability as pending, not confirmed.

The strongest source-backed statement today is narrow: MiniMax has launched M3 on its own product/API surfaces and says the model combines 1M context, coding-agent strength, and native multimodality.

MiniMax-Reported Benchmark Signals

These are MiniMax-reported numbers and relative claims from the M3 launch post. They are useful shortlist signals, not independent proof.

SignalMiniMax-reported resultHow to use it
SWE-Bench Pro59.0%Treat as a coding-agent signal to test, not a universal quality score
Terminal-Bench 2.166.0%Relevant if your workflow involves shell execution and repair loops
SWE-fficiency34.8%Useful for coding efficiency comparisons, but still vendor-reported
KernelBench Hard28.8%Useful signal for low-level/code optimization tasks
MCP Atlas74.2%Relevant to tool-using and agentic workflows

MiniMax also says M3 surpasses GPT-5.5 and Gemini 3.1 Pro and approaches Claude Opus 4.7 on SWE-Bench Pro, surpasses Opus 4.7 on SVG-Bench, scores above Gemini 3.1 Pro on OmniDocBench, and leads on Claw-Eval. Do not rewrite that into “MiniMax beats Opus” or a blanket winner claim. The comparison is vendor-published, benchmark-specific, and tied to MiniMax’s Opus 4.7 comparison, not current Opus 4.8 pricing or capability.

The $20 Plan Math

MiniMax’s launch post lists:

TierPriceMonthly M3 token usage
Plus$20/month~1.7B tokens
Max$50/month~5.1B tokens
Ultra$120/month~9.8B tokens

For the Plus tier, a naive flat division is:

1
$20 / 1,700M tokens = about $0.012 per 1M included M3 tokens

That number is useful for intuition, but it is not the same thing as API PAYG pricing. The Token Plan uses a Subscription Key, shared resource quota, 5-hour rolling windows, weekly windows, and product-specific access rules. PAYG uses API billing at the listed per-token rates.

Here is the safer comparison: if you were paying API list prices for a mixed coding workload up to 512K input per request, the headline rates look like this.

Model/API laneInput / output per 1M tokens80% input / 20% output blended cost50% input / 50% output blended cost
MiniMax M3 Standard PAYG, up to 512K input$0.60 / $2.40$0.96 per 1M total tokens$1.50 per 1M total tokens
MiniMax M3 Priority PAYG, up to 512K input$0.90 / $3.60$1.44 per 1M total tokens$2.25 per 1M total tokens
Claude Opus 4.6 / 4.7 / 4.8 API$5.00 / $25.00$9.00 per 1M total tokens$15.00 per 1M total tokens
GPT-5.5 API standard, short context$5.00 / $30.00$10.00 per 1M total tokens$17.50 per 1M total tokens

MiniMax’s PAYG page also showed a time-limited-looking 50% off M3 row when checked on June 5, 2026. This guide uses list pricing for the durable comparison. For any purchase, the MiniMax console and checkout page are the final authority.

What The Math Means

If your Token Plan workflow actually fits inside the Plus quota and supported tools, $20/month could cover far more coding-agent token volume than premium API spend. That is the value story.

The caveat is just as important: do not flatten subscription quota into guaranteed API savings. Token Plan usage, PAYG API usage, model routing, resource mix, and quota windows are separate operational questions.

Where MiniMax M3 Fits

Workflow choiceUse MiniMax M3 when…Keep the alternative when…
GLM-5.1You want a 1M-context, multimodal, MiniMax-supported lane and the Token Plan works in your toolYou already rely on Z.AI’s cheaper supported-tool GLM coding lane and it passes your evals
Kimi K2.6 / Kimi k2.5You want to test MiniMax’s long-context coding-agent stack or MiniMax CodeYou prefer Kimi’s API/tooling, Kimi Code quota, or open-weight ecosystem
Gemini high-context lanesYou need another 1M-context candidate with coding-agent focus and Token Plan economicsYou are doing Google-native workflows, AI Studio experiments, or Gemini-specific document analysis
GPT-5.5You need to reduce API spend on routine coding loops and can accept a model trial laneYou need OpenAI-native tooling, Codex workflows, or the premium model’s behavior on hard work
Claude Opus 4.6 / 4.7 / 4.8You want a cheaper daily lane while keeping Claude for review, architecture, and hard debuggingYou need Claude Code-native behavior, highest-confidence review, or established Anthropic enterprise posture

The practical pattern is the same as other value models: route routine work to the cheaper lane only after it passes a real eval, then keep the premium model for arbitration.

Access And Setup

MiniMax documents two billing paths:

  • Token Plan: Subscription Key, monthly quota, shared quota across supported text/image/speech/music resources, and MiniMax Code access.
  • PAYG: API key billed at per-resource list prices, with M3 pricing split between up to 512K input and above 512K input calls.

MiniMax also documents M3 setup paths for coding tools:

  • Claude Code: MiniMax marks this as the recommended coding-tool path.
  • OpenClaw: MiniMax documents OAuth setup with MiniMax as the provider and optional extra API keys for tools.
  • OpenCode: MiniMax documents provider setup and an Anthropic-compatible endpoint at https://api.minimax.io/anthropic/v1.
  • Cursor: MiniMax documents adding MiniMax-M3 as a custom model for Chat / Composer / Edit modes, while noting Cursor Tab uses Cursor’s own model.
  • Kilo Code, Cline, Roo Code, TRAE, Droid, Zed: MiniMax documents additional setup paths.
  • Codex CLI: MiniMax’s own table lists a Codex CLI path as “Not Recommended,” so do not treat it as the default route.

For API compatibility, MiniMax’s API overview lists Anthropic-compatible messages, OpenAI-compatible chat completions, and OpenAI-compatible model surfaces.

Evaluation Plan

Do not buy the first month and immediately route important work through it. Use a short eval set:

TestAsk M3 to doPass signal
Bug fixFix a real failing test with repository contextSmall patch, correct diagnosis, tests pass
RefactorMove behavior across 2-4 filesPreserves local style and avoids unrelated churn
ReviewReview a risky PR or local diffFinds concrete issues without inventing policy
Long contextAnalyze a large module or docs setUses context accurately instead of summarizing vaguely
MultimodalInspect a screenshot or UI artifact, then change codeProduces changes that match the visible artifact

If M3 passes, graduate it to low-risk routine work. If it fails, keep it as an experimental lane and continue using GPT, Claude, Gemini, Kimi, or GLM for the tasks they already handle well.

Referral Disclosure

MiniMax’s official referral docs say the Co-builder Referral Event runs through June 30, 2026. They also say referral purchases receive a 10% checkout discount, while the referrer receives Open Platform vouchers worth 10% of the invitee’s actual payment amount.

The privacy disclosure matters: MiniMax says the referrer may be aware of the invitee’s purchase timing and amount. If you use a referral link, assume the referrer may receive that purchase metadata.

AIHackers does not yet have an owned MiniMax referral code. Treat the official referral mechanics as documented, but do not promote a third-party code here. AIHackers-owned code pending.

Sources


Last verified: June 5, 2026. Pricing, quota windows, discounts, supported tools, and model availability can change quickly. Treat checkout and provider docs as final authority.