MiniMax M3 makes a claim that should make every closed-model vendor look twice. It says an open-weight model now beats GPT-5.5 and Gemini 3.1 Pro on a hard coding benchmark, and lands close to Claude Opus 4.7. If that holds, the math of building agentic coding tools changes overnight. You’d get frontier-class results from weights you can download, run, and price however you like.
Here’s the honest version up front. Most of the numbers behind that claim come from MiniMax itself. They’re vendor-reported, and independent leaderboard confirmation is still pending. So this isn’t a coronation. It’s a look at what M3 says it can do, how that stacks against two closed frontier models, and how to decide which one belongs in your stack. For the full background on the model, see what is MiniMax M3, and the source figures live in the MiniMax M3 announcement.
The contenders at a glance
Three models, three different bets. M3 goes open and cheap. Opus 4.7 goes for reliability and ecosystem. GPT-5.5 goes for the default-platform position inside the OpenAI stack.
| Attribute | MiniMax M3 | Claude Opus 4.7 | GPT-5.5 |
|---|---|---|---|
| Weights | Open (release due ~10 days) | Closed | Closed |
| Context window | 1,000,000 tokens | Large (see Anthropic docs) | Large (see OpenAI docs) |
| Multimodal | Native: image, video, computer use | Image + text | Image + text |
| Architecture | MSA (~1/20 per-token compute vs prev gen) | Not disclosed | Not disclosed |
| Pricing model | Plans $20 / $50 / $120 + usage API | Per-token, Anthropic pricing | Per-token, OpenAI pricing |
| Param counts | Not disclosed | Not disclosed | Not disclosed |
The open-versus-closed split is the headline. You can’t self-host Opus 4.7 or GPT-5.5. With M3, MiniMax says weights and a technical report ship within about ten days, which puts on-prem deployment and full price control back on the table.
Coding benchmarks: where M3 leads, and where it doesn’t
Coding is where M3 stakes its biggest claim. The standout is SWE-Bench Pro, a test of real-world software engineering tasks. Here are the MiniMax-reported figures.
| Benchmark (MiniMax-reported) | MiniMax M3 | Positioning MiniMax claims |
|---|---|---|
| SWE-Bench Pro | 59.0% | Above GPT-5.5, above Gemini 3.1 Pro, approaches Opus 4.7 |
| Terminal-Bench 2.1 | 66.0% | Strong agentic terminal score |
| SWE-fficiency | 34.8% | Efficiency on resolving issues |
| KernelBench Hard | 28.8% | Low-level kernel generation |
| PostTrainBench | 0.37 | Behind Opus 4.7 (0.42) and GPT-5.5 (0.39) |
Read that table carefully, because it cuts both ways. On SWE-Bench Pro, M3’s 59.0% is the number that would let an open-weight model sit in frontier company. You can check the public SWE-Bench leaderboard to see how that lines up once third parties verify it. But on PostTrainBench, M3 trails. Opus 4.7 leads at 0.42, GPT-5.5 follows at 0.39, and M3 sits at 0.37. MiniMax is behind on that one, and pretending otherwise would do you a disservice.
So the picture isn’t “M3 wins coding.” It’s “M3 reaches frontier range on the headline coding benchmark while still trailing on others.” That’s a meaningful step for an open model. It’s not a clean sweep. We’ve seen this pattern before with strong open releases. If you tracked the Qwen 3.7 vs GPT-5.5 vs Opus 4.7 comparison, the shape is familiar: open models close the gap on specific tasks faster than they close it everywhere.
One more caveat worth repeating. These are MiniMax’s own runs. Benchmark harnesses, scaffolding, and prompt setups vary between vendors, and small methodology choices move scores by points. Treat the comparison as directional until independent leaderboards report their own numbers.
Agentic and tool use: the long-horizon bet
If coding is the headline, agentic behavior is where M3’s architecture earns its keep. The model scores 74.2% on MCP Atlas, a test of tool orchestration through the Model Context Protocol, and MiniMax reports the highest score in the field on Claw-Eval, an agentic evaluation.
The demos are the part that get attention. MiniMax shows M3 running a 24-hour CUDA kernel optimization task that lands a 9.4x speedup, and an autonomous paper reproduction that produced 18 commits and 23 figures without a human in the loop. Long-horizon agentic work like that is exactly where most models drift, lose context, or burn tokens on dead ends.
The reliability of an agent depends as much on the harness around the model as on the model itself. How you structure tool calls, context, and recovery loops decides whether a 24-hour run finishes or falls over. Our breakdown of Claude Code agent harness architecture covers that scaffolding in depth, and the same principles apply whichever model sits at the center. A strong agentic score on a vendor benchmark is promising. Watching it hold up across your own multi-step workflows is the real test.
Multimodal and document understanding
M3 ships native multimodal support out of the box: image, video, and computer use. That’s a wider input surface than the image-plus-text setups on Opus 4.7 and GPT-5.5.
Two benchmarks back the claim. On SVG-Bench, which tests structured graphics generation, MiniMax reports M3 above Opus 4.7. On OmniDocBench, a document-understanding test, it reports M3 above Gemini 3.1 Pro. Pair that with computer use, and M3 positions itself for workflows that read documents, parse screens, and act, not only chat. As always, these sit in the vendor-reported column until someone else runs them.
Context window and the cost of long context
M3 carries a 1,000,000-token context window, and the way it gets there matters more than the number. The model uses an architecture MiniMax calls MSA, which it says cuts per-token compute to roughly 1/20 of the previous generation, with more than 9x faster prefill and more than 15x faster decode.
That speedup is the quiet headline. Long context is cheap to advertise and expensive to actually use. Every token you stuff into a prompt costs compute on every step of an agent loop, which is why long-running agents get slow and pricey fast. If M3’s per-token cost really is a fraction of prior models, feeding it a large codebase or a long document trail becomes far less punishing.
That economics question applies to all three models. Before you assume a 1M window is free to fill, read how to reduce agent token costs in the CLI. The cheapest token is the one you never send, regardless of which model you pick.
Pricing reality
This is where open and closed diverge hardest. M3 has token plans at $20 (Plus), $50 (Max), and $120 (Ultra), plus an API with a standard rate for inputs up to 512K tokens and a long-context rate above that, across standard and priority tiers. MiniMax hasn’t published an exact per-token price yet, so treat the plan tiers as the concrete signal for now.
Opus 4.7 and GPT-5.5 price per token, and you should pull the current numbers straight from the source: Anthropic’s pricing page and OpenAI’s pricing page. Prices move, and hardcoding them here would only mislead you later.
The structural tradeoff is the durable point. With M3’s open weights, you can self-host and turn API cost into infrastructure cost, which pays off at high volume if you have the ops capacity. With Opus 4.7 and GPT-5.5, you rent inference at a known per-token rate and skip the infrastructure entirely. This open-weight pricing pressure is part of a larger shift; the Chinese LLM price war of 2026 traces how aggressive open releases are dragging frontier costs down across the board.
Which one should you pick
Match the model to your constraint, not to the leaderboard.
| Your situation | Pick | Why |
|---|---|---|
| Cost-sensitive or need self-hosting | MiniMax M3 | Open weights, cheap plans, full price and deployment control |
| Maximum reliability and mature ecosystem | Claude Opus 4.7 | Proven tooling, leads PostTrainBench, deep integration support |
| Already standardized on OpenAI | GPT-5.5 | Stays inside your existing stack, tools, and billing |
| Long agentic runs on a budget | MiniMax M3 | 1M context plus MSA efficiency cuts long-horizon cost |
| Data residency or air-gapped needs | MiniMax M3 | Only option you can run on your own hardware |
If you’re risk-averse and shipping to production today, the vendor-reported caveat matters, and Opus 4.7’s track record carries weight. If you’re cost-driven, building at volume, or need control over where the model runs, M3’s open weights are hard to ignore once they land. There’s no single winner here, only the right fit for your constraints.
How to benchmark them yourself
Vendor numbers tell you what’s possible. Your own prompts tell you what’s true for your workload. The fastest way to settle it is to run identical prompts against all three model APIs and compare the actual output, latency, and token usage side by side.
You can set this up in one Apidog project. Create a request for each provider’s chat endpoint, drop in the same prompt and parameters, save them as a test scenario, and run the batch. Apidog shows you response time and full output per request, so you compare M3, Opus 4.7, and GPT-5.5 on the same task in one window instead of juggling three playgrounds. Add a few assertions and you can even check that each model returns valid JSON or hits a structure your app expects. Download Apidog to follow along, and use environment variables to swap API keys cleanly between the three.
When you’re ready to wire up M3 specifically, our guide on how to use the MiniMax M3 API walks through auth and the request shape. From there, running the same suite against Opus 4.7 and GPT-5.5 in Apidog is a copy-paste away.
FAQ
Is MiniMax M3 really better than GPT-5.5? On SWE-Bench Pro, MiniMax reports M3 at 59.0%, above GPT-5.5. On PostTrainBench, GPT-5.5 leads at 0.39 versus M3’s 0.37. So it depends on the task, and these are vendor-reported figures awaiting independent confirmation. M3 isn’t uniformly ahead.
Is MiniMax M3 open source? M3 is open-weight, with weights and a technical report due within about ten days of the announcement. You’ll be able to download and run the model. MiniMax hasn’t disclosed parameter counts, and open-weight isn’t always the same as a fully open-source license, so read the release terms when they land.
Can M3 replace Opus 4.7 for agentic coding? Possibly, for cost-sensitive or self-hosted setups. M3 posts strong agentic numbers (66.0% Terminal-Bench 2.1, 74.2% MCP Atlas) and long-horizon demos. But Opus 4.7 leads PostTrainBench and has a more proven production track record. Test both on your own workflows, ideally with a solid harness, before you switch.
Are these benchmark numbers independent? Mostly no. The figures here are largely MiniMax’s own reported results. Public leaderboards like SWE-Bench will let you cross-check the headline coding claim once third parties run M3. Until then, treat the comparison as directional.
What’s the catch with M3’s 1M-token context? The window is real, and the MSA architecture is built to make filling it cheaper, with more than 9x faster prefill and more than 15x faster decode. But long context still costs compute on every agent step across any model, so prompt discipline still matters.
How do I compare all three without committing to one? Run the same prompts against each API and measure output, latency, and cost. A single Apidog project with one request per provider gives you a side-by-side view without writing throwaway scripts.
The bottom line
MiniMax M3 is the most serious open-weight challenge to the frontier we’ve seen, and its SWE-Bench Pro claim would reset expectations if independent leaderboards confirm it. But the data is mostly MiniMax’s own, and PostTrainBench shows Opus 4.7 and GPT-5.5 still ahead. Pick M3 if cost, self-hosting, or control drive your decision. Pick Opus 4.7 for proven reliability, or GPT-5.5 if you live in the OpenAI stack. Then run all three against your own prompts before you commit, because your workload is the only benchmark that ships.
