MiniMax M3.1 Flash vs DeepSeek V4.1 Flash: What the Two Vendor Tables Actually Agree On

Two rows. That is how much of M​iniMax's launch table survives a comparison with D​eepSeek's own published numbers, and it is worth saying so plainly before either vendor's headline gets repeated: on the two benchmarks where M​iniMax M3.1 Flash and DeepSeek V4.1 Flash — the 1017 checkpoint D​eepSeek shipped two days earlier — both labs published a score on the same evaluation, the newer model is behind. M​iniMax released M3.1 Flash on 19 October 2026 with a table spanning agentic work, coding, computer use and cybersecurity, and a claim of a generational jump over its own predecessor. D​eepSeek released V4.1 Flash 1017 on 17 October 2026 with a model card doing the same thing. The interesting part of putting them together is not who wins. It is how little of each table is built to be compared at all. Vendor tables are one kind of evidence and incident records are another, so the failures behind releases like these are catalogued in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

What M​iniMax actually put a number on

M​iniMax's post is more specific than most. It gives six figures up front: BankerToolBench 76.2, DeepSWE v1.1 73.7%, OSWorld V2 61.7, Agents' Last Exam 29.6, CyberGym 82%, and a claim of top performance on Automation-Bench without a score in the summary text. Alongside them sit two throughput claims — roughly three times faster online inference than the model it replaces, and 200-plus tokens per second with a P50 above 170 — and a cost claim of 70% lower on agentic tasks. All of that is vendor-reported and unreproduced. M​iniMax ran it, M​iniMax published it, and the harness is named only in footnotes.

The footnote is where the table starts to come apart. BankerToolBench, M​iniMax says, was evaluated with the OpenCode 1.1718 harness and judged by MiniMax-M2.7, its own older model. That is a self-judged benchmark: a M​iniMax model scoring outputs produced by a M​iniMax model, on a task set M​iniMax chose. It may be perfectly well run. It is not evidence you can hold against a lab that publishes on a harness somebody else maintains.

The two rows that line up

DeepSWE v1.1 is the one genuinely shared measurement. Both labs report it, and both name the harness: D​eepSeek says Mini-SWE-V2 on its model card, M​iniMax reports the same benchmark version. M​iniMax claims 73.7%. D​eepSeek claims 77.0% for the 1017 checkpoint, against 74.2% for the V4.1 Flash release it superseded. So the ranking here is not ambiguous, and it is not flattering to the newer model: D​eepSeek's October 17 release is ahead of M​iniMax's October 19 release on the one coding-agent benchmark the two companies measured the same way, and it is ahead of its own previous release by nearly three points while doing it.

Agents' Last Exam tells the same story with more distance. M​iniMax reports 29.6. D​eepSeek reports 35.4 for 1017, up from 31.8 for the prior checkpoint. That is a five-point gap on a benchmark designed to be hard for agents, and it is the second consecutive row where the two-day-older model is comfortably in front.

Everything else in the two tables is either measured on a different rig, judged by a different model, or published by only one side. D​eepSeek's card carries Terminal-Bench 4.0 at 42.3 in the Claude Code harness, HLE without tools at 42.6, ProgramBench 29.1, Apex Bench v1.1 72.1 and Automation-Bench v1.0.6 at 60.5. M​iniMax's post does not give a Terminal-Bench 4.0 figure at all, and its Automation-Bench row sits under a footnote about the 600 publicly released tasks rather than beside a comparable number we could attribute with confidence at the resolution the post renders. Two of those rows are checkable. The rest are two different exams with the same name on the cover.

Where M​iniMax's table is strong and D​eepSeek's is silent

It would be a mistake to read the two shared rows as the whole picture, because they are both coding-and-agent reasoning rows, and M​iniMax's pitch is broader. OSWorld V2 at 61.7 is a computer-use number D​eepSeek does not publish, and M​iniMax claims first place among Chinese models on it. CyberGym at 82% is a security-agent number D​eepSeek also does not publish. BankerToolBench 76.2 measures tool selection in a business-workflow setting that neither D​eepSeek's card nor Artificial Analysis covers. On that territory M​iniMax is not losing; it is unopposed. Unopposed is not the same as ahead, and the vendor's own summary — "reaching the level of leading models" — is a claim about parity, not about a lead.

The other thing worth noticing about the launch table is what it does with competitors. The published graphic sets M​iniMax's column beside D​eepSeek, Kimi-K3, GLM-5.3, GPT-5.6 Sol and Opus 5, and the numbers in those columns are the other labs' own published figures rather than numbers M​iniMax ran. That is normal practice and it is not dishonest, but it means every cross-column comparison in that image inherits whatever harness each lab happened to use. A table that lines up six models invites you to read across rows. Only two of those rows were built to be read across.

The price story runs the other way, and it is real

On money, M​iniMax has the cleaner card. M3.1 Flash is $0.15 per million input tokens and $0.60 per million output, with cached input at $0.015 — a 90% discount — and no time-of-day variation. DeepSeek V4.1 Flash 1017 lists $0.30 and $1.20 during its weekday peak windows, halving to $0.15 and $0.60 off-peak, with cache hits at $0.006 and $0.003 respectively. During peak hours D​eepSeek costs twice what M​iniMax charges for the same million tokens. Off-peak the two cards are identical. A workload that runs on a schedule you control can chase D​eepSeek's cheap window; a workload that runs when your users do cannot.

Artificial Analysis puts a number on the difference that survives the pricing games: $0.22 to complete one Intelligence Index task on M3.1 Flash against $0.26 on DeepSeek V4.1 Flash 1017. That is a 15% gap in M​iniMax's favour on the same evaluation, and it is the one economic claim in this comparison that is independently measured rather than vendor-asserted. It matters because the per-token card does not tell you this. D​eepSeek's $0.30/$1.20 looks like a 2x penalty and behaves like a 15% one, because M​iniMax burns 180 million output tokens running the index against D​eepSeek's 210 million and because D​eepSeek is roughly twice as fast at 208.7 tokens per second against about 105. Two models, two pricing philosophies, and a real cost difference of eighteen cents an eval.

Which is also where a routing layer earns its place, in a way that is easy to state and hard to do by hand. OrcaRouter passes provider list price through at 0% markup, so when either lab moves a rate the number on your bill moves the same day rather than at the next contract renewal — and on a card with a peak window, that is the difference between a budget you set once and a budget you re-check hourly. We do not carry M​iniMax M3.1 Flash or DeepSeek V4.1 Flash 1017 on the catalogue today, and this article is not a claim that we do; the D​eepSeek family we do route is a different checkpoint, which is a distinction worth keeping straight when the model string is what your code is pinned to.

Two numbers the benchmark table does not contain

The first is speed. M​iniMax claims 200-plus tokens per second and P50 above 170. Artificial Analysis measures roughly 105 tokens per second on its own harness, with a 1.38-second time to first token. D​eepSeek measures 208.7 tokens per second at 1.07 seconds. Vendor throughput claims are almost always measured on the best case a serving stack can produce; the independent number is the one that shows up in your p95. On this dimension M​iniMax's own claim and the measured result are not close, and the measured result is half of D​eepSeek's.

The second is licence. DeepSeek V4.1 Flash 1017 is MIT-licensed with weights on Hugging Face and 552 billion total parameters against 16 billion active. M​iniMax says M3.1 Flash's weights "will be released later" and has attached no date, no licence and no parameter count to that promise. If your procurement runs through a security review that asks what happens if the vendor disappears, the benchmark table is not the question you are answering, and on the table you were given the answer is not there.

What would change this reading

Three things, in order of how much they would matter. Independent runs of DeepSWE v1.1 and Agents' Last Exam on M​iniMax's checkpoint would settle whether the 73.7 and 29.6 are conservative or generous — right now nobody outside the two labs has published either. A published Automation-Bench score from M​iniMax, on the same 600-task release D​eepSeek used, would give the agentic claim the number it currently lacks. And a weights release with a licence attached would make the model comparable on the axis where D​eepSeek currently has no competition at all.

Until then the honest summary is narrower than either launch post. M​iniMax M3.1 Flash is a cheaper model to finish a task with, a slower one to talk to, a closed one to procure, and — on the only two benchmarks both companies measured identically — the second-place one. That is not the story the table was drawn to tell, but it is the story the table contains. If you are choosing between them this week, choose on the price card and the licence, because those are the two rows nobody had to be persuaded of.