← Tools I Use

MiniMax M3: an open-weight coding model that undercuts the frontier

2026-06-02 · ~6 min read · Conor Dobbs

A Chinese lab shipped an open-weight coding model on June 1 that beats GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro, carries a one-million-token context, and costs a fraction of what the frontier labs charge. That is the MiniMax M3 pitch, and most of it checks out. The interesting part is not the leaderboard. It is what an open-weight model at this tier does to the build-versus-buy math for anyone running coding agents at scale.

What MiniMax shipped

M3 is built on a new attention design MiniMax calls MSA, short for MiniMax Sparse Attention. The point of MSA is to dodge the quadratic cost of standard attention at long context. The payoff MiniMax reports: at a one-million-token context, per-token compute drops to roughly one-twentieth of the previous M2 generation, with prefill more than nine times faster and decoding more than fifteen times faster. If those numbers hold outside the lab, long-context coding stops being a billing event and starts being routine.

The headline specs:

That last point is the one that matters. Frontier-level coding behind open weights is a different product than the same scores behind a closed API.

The benchmarks, and the asterisk

On the numbers MiniMax published, M3 is genuinely competitive at the top:

Here is the asterisk, and it is the same one I put on every model launch. These are vendor-reported figures from the lab that built the model. A benchmark is a marketing surface until someone independent reproduces it. SWE-Bench numbers in particular swing hard on the harness, the retry budget, and how partial credit is counted. Treat 59.0% as "plausibly frontier-adjacent," not as settled fact, until third parties run it. The honest read is that M3 is in the conversation, which a year ago no open-weight model was.

There is a second asterisk on the word "open." At launch, the weights were not out. MiniMax promised them within ten days. Open-weight is a commitment here, not yet a download. If you are betting infrastructure on self-hosting M3, wait for the weights to land and for someone to confirm the released checkpoint matches the API model.

Why the price is the real story

MiniMax listed M3 on OpenRouter at roughly $0.60 per million input tokens and $2.40 per million output, with a launch promotion cutting that to about $0.30 and $1.20. Set that next to what the closed frontier charges for output tokens and the gap is an order of magnitude, not a rounding error. There are also subscription tiers running from $20 to $120 a month for billions of tokens.

For a single developer, model price barely registers. You pay for the tool that finishes the task, and a few dollars of token spend is noise against an hour of your time. For anyone running agents in a loop, at volume, across a team, the math flips. Output tokens are where coding agents burn money, because the model writes far more than it reads. A ten-times cheaper output token, on a model that lands within a few points of the frontier on the benchmark you care about, is the kind of thing a finance team notices. This is the same pressure I wrote about in how to cut your Claude Code bill: the cost of an agent is dominated by how much it writes and how often you run it, and a cheaper per-token rate compounds across every run.

How I would decide

The instinct on a launch like this is to swap your stack. Do not. Here is the decision I would run instead.

Stay where you are for your daily driver. If Claude Code or your current agent already wins your real tasks, M3 does not change your day. Raw benchmark proximity is not the same as the lived experience of an agent that handles your codebase's size, respects your conventions, and recovers from its own mistakes. Those qualities do not show up on SWE-Bench Pro. I walked through that gap between leaderboard and lived workflow in Cursor vs Claude Code vs Aider, and it holds here.

Pilot M3 where price and context dominate. Two cases make it worth a real trial. First, jobs that genuinely need the long context: dumping a large monorepo or a sprawling log into a single window where the one-million-token math and the cheap tokens both pay off. Second, high-volume batch work where you run the same agent thousands of times and a ten-times output discount becomes a budget line rather than a footnote.

Wait for the weights if self-hosting is the point. The reason to care about open weights is control: running the model inside your own boundary, no data leaving for a third-party API, no vendor able to deprecate it out from under you. That case only exists once the checkpoint is downloadable and verified. Until then you are using a closed API that happens to promise openness later.

One more consideration that has nothing to do with quality. M3 comes from a Chinese lab, and for some organizations the API endpoint's jurisdiction is a procurement question before it is a capability one. Self-hosting the eventual open weights sidesteps that, which is part of why the open-weight release matters beyond ideology.

The honest bottom line

MiniMax M3 is the clearest sign yet that open weights are closing on the closed frontier instead of trailing it by a generation. An open-weight model that argues with GPT-5.5 and Gemini 3.1 Pro on a real coding benchmark, at a tenth of the token price, with a million-token window, would have read as fiction a year ago.

What it is not is a reason to rip out a working agent. The frontier labs still hold an edge on the parts that never make a benchmark: reliability across a long session, the multi-step autonomy you trust without watching, the failure modes you have learned to predict. M3 narrows the gap on price and context. Whether it narrows it on the work in front of you is something only your own evals, run on your own tasks once the weights are public, can tell you. Verify, do not trust the launch chart.

Sources