MiniMax M3.1 Flash vs GLM-5.3-Flash: The Rate Cards Match, So Decide on the Rows That Don't

The two rate cards are, for practical purposes, the same card. M​iniMax M3.1 Flash charges $0.15 per million input tokens and $0.60 per million output. GLM-5.3-Flash charges $0.15 and $0.50. Blend them the way Artificial Analysis does and both land at $0.10 per million. Their independent intelligence readings are one point apart — 43 against 42 on the current index. So a comparison that stops at price and score has nothing to say, which is why most of them stop there and then recommend whichever model the writer happened to test. The decision between M​iniMax M3.1 Flash and GLM-5.3-Flash lives entirely in the rows that do not match: what you are allowed to do with the weights, what happens when your output hits a ceiling, what you can feed the model, and whether the thing is even purchasable through the endpoint you already use. Rate cards and benchmark tables both leave out how these models fail once they are in production, and the recorded incidents are catalogued in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

• Six lines, both sides

• Price — M​iniMax M3.1 Flash $0.15 input and $0.60 output per million, cached input $0.015; GLM-5.3-Flash $0.15 and $0.50, cached input $0.03. Both blend to $0.10 per million at a 7:2:1 input-to-cached-to-output mix.

• Independent score — Intelligence Index 43 for M​iniMax M3.1 Flash, 42 for GLM-5.3-Flash, on the same index version read the same day. One point is inside the noise of a single eval run.

• Cost per finished task — $0.22 per index task for M​iniMax M3.1 Flash against $0.28 for GLM-5.3-Flash. The cheaper sticker is also the cheaper task, by about 21%.

• Maximum output — not published by M​iniMax for M3.1 Flash; 128,000 tokens for GLM-5.3-Flash. Both advertise a 1M-token context window, but only one of them says how much of it can come back out.

• Inputs — text, image and video for M​iniMax M3.1 Flash; text, image, video and files for GLM-5.3-Flash. Video on both sides is the unusual part; file upload on one side is the operational part.

• Open weights — GLM-5.3-Flash is MIT-licensed, 320 billion total parameters with 18 billion active, downloadable today. M​iniMax says M3.1 Flash's weights will be released later, with no licence and no date attached to the promise.

Why one point on the index is not a reason to pick either

The Intelligence Index is a weighted composite: agents at 30%, coding at 20%, scientific reasoning at 20%, general at 30%, built from ten evaluations including AA-Briefcase and AA-Omniscience at 15% each. A single point on a composite of that shape can come from one sub-eval moving by a couple of points. Read it as a tie and you will be right far more often than you will be wrong. Read it as a ranking and you are choosing a production model on the fourth significant figure of a survey.

The per-task cost comparison is more useful than the index comparison, and it also points the same way the sticker does. At $0.22 per completed task, M​iniMax M3.1 Flash is doing the index's work for less than GLM-5.3-Flash's $0.28, despite a higher output rate per token. That is the verbosity and caching arithmetic doing its job: GLM-5.3-Flash charges a third more per output token and recovers less of it through cache, so the cheaper-looking output rate on M​iniMax's card is genuinely cheaper in aggregate. If your workload is a long-running agent loop where most of the bill is output tokens being re-read, that 21% is the number that shows up in the monthly line item.

Speed splits cleanly, and it splits against the cheaper model

This is the dimension where the two stop being interchangeable. M​iniMax M3.1 Flash measures about 105 tokens per second of output with a 1.38-second time to first token. GLM-5.3-Flash measures 76.8 tokens per second with a 2.57-second first token. M​iniMax is roughly 36% faster once it starts and roughly twice as quick to start. For an interactive surface — an editor, a chat pane, a coding assistant where the first token is what the human is waiting on — a 1.2-second advantage on TTFT is the difference between a tool that feels responsive and one that feels like it is thinking about it. For a batch pipeline the throughput gap matters more than the latency gap, and it still favours M​iniMax.

Both models make you pay for the thinking whether you want it or not. M​iniMax rejects a request that tries to disable reasoning outright. Z.ai's documentation says the same thing in different words: thinking cannot be switched off, and maximum effort is the recommended setting. So neither side offers the cost lever that a switchable-reasoning model gives you, and both bill their reasoning tokens as output. If you were hoping one of these two was the cheap-and-quiet option for a high-volume classification job, neither is. That is a property of the tier, not of either vendor.

The 128K answer ceiling is the row that decides real work

GLM-5.3-Flash publishes a 128,000-token maximum output. M​iniMax publishes nothing for M3.1 Flash — not a low number, not a caveat, nothing on the model's page or in the API documentation we can find. In practice that means you cannot know, before you send a request, whether a 40,000-token code generation will complete or truncate. You can discover it empirically, and you can build a retry around it, but you cannot put it in a capacity plan or a contract.

For most chat-shaped traffic this is invisible. For the workloads that make a Flash tier interesting — whole-file rewrites, long agent transcripts, structured extraction from a large document — the output ceiling is the first thing that breaks, and it breaks quietly. A model that truncates at an undocumented length and returns a well-formed partial answer is worse to operate than a model that refuses the request. If any part of your pipeline generates more than a few thousand tokens in one call, this single unpublished row should outweigh the one-point index difference and the 21% cost advantage combined.

Open weights are not a philosophy question here

GLM-5.3-Flash is MIT-licensed with downloadable weights, 320 billion parameters total and 18 billion active, which puts it in the range where a serious team can self-host on a multi-GPU node and a smaller team can at least audit what it was trained to do. M​iniMax M3.1 Flash is closed today, and the promise of later weights is exactly that — a promise with no licence, no parameter count and no date. Procurement processes that ask "what is our exit if this vendor changes terms" have an answer on one side and a sentence on the other.

There is a cost angle buried in that too, and it is the one nobody puts in a comparison table. An open-weight model sets a ceiling on what any provider can charge you for it, because the alternative is always to run it yourself. A closed model has no such ceiling; the vendor sets the price and you take it or migrate. At identical blended rates today, GLM-5.3-Flash's licence is the thing that keeps its rate honest next year, and M​iniMax's is the thing that leaves the rate entirely at M​iniMax's discretion.

Availability is where the tie actually breaks

GLM-5.3-Flash has been live since 26 August 2026 and is on OrcaRouter's catalogue as z-ai/glm-5.3-flash, at $0.15 per million input and $0.50 per million output with cached input at $0.03, alongside a free SKU for evaluation and the larger z-ai/glm-5.3 at $1.26 and $3.96 if you need the flagship. M​iniMax M3.1 Flash, two days old at the time of writing, is reachable through M​iniMax's own API, the M Plan and M​iniMax Code, and it is not on our catalogue — we are not going to imply that it is, and the model page returns nothing when you go looking for it.

That asymmetry is worth more than it looks when you are comparing two models that cost the same. Getting GLM-5.3-Flash into a codebase is a model string. Getting M​iniMax M3.1 Flash in is a second vendor relationship, a second billing shape and a second set of rate limits, on top of the one you already have. And it is precisely the case where a routing layer is not a pitch but a piece of plumbing: one key, one invoice, one place where a provider's list price is passed through at 0% markup so a rate change reaches your bill the day it happens rather than at the next renewal. If M​iniMax lists M3.1 Flash on a routable endpoint later, that becomes a two-model decision instead of a two-contract one. Today it does not, and pretending otherwise would be the kind of comparison this blog exists to avoid.

Which one to call

If your work is interactive, output-heavy, and lives inside an existing OpenAI-compatible stack, GLM-5.3-Flash is the easier answer: same blended price, a published output ceiling, an MIT licence you can act on today, file and video input, and a catalogue entry that makes it a one-line change. The 2.57-second first token is the price of that, and for background agents it is a price worth paying.

If your work is latency-sensitive and output is bounded — a routing classifier, a short-form generator, an agent step that returns a plan rather than a file — M​iniMax M3.1 Flash is the faster model with the cheaper per-task cost, and its unpublished output ceiling never gets tested because you never approach it. Choose it for the throughput and the 1.38-second first token, not for the index, and go in knowing you are taking a closed model with a rate nobody has capped.

What you should not do is pick between them on the benchmark tables. Both vendors published numbers on harnesses of their own choosing, neither ran the other's, and the one benchmark both families have reported historically — DeepSWE — has been run in different harnesses across the two labs' releases often enough that a cross-column comparison is guesswork wearing a decimal point. The two figures in this article that were measured the same way for both models are the index and the cost per task, and those two say the models are the same price and the same capability. Everything that actually differs is a licence term, a documented limit, or a line in a catalogue.