MiniMax-M3.1-Flash Scores 42 on the Intelligence Index — and Costs $0.22 a Task

MiniMax-M3.1-Flash landed on 19 October 2026 with an Artificial Analysis Intelligence Index score of 42, thirteen points above MiniMax-M3 at 29, at half the per-token rate and — the figure that actually matters — less than half the cost per task, $0.22 against $0.51. It ties GLM-5.3-Flash at 42, sits three points behind DeepSeek V4.1 Flash 1017 at 45, and costs less than either to complete a benchmark task. Those four numbers are the whole story in one line. The interesting part is what happens when you stop reading the index and start reading the invoice, because this model is cheap per token and unusually expensive per token at the same time. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

The number M​iniMax did not put in the blog post

M​iniMax's launch post is a competent piece of vendor marketing: faster inference, stronger agentic scores, 70% lower agentic-task cost than the predecessor. What it does not contain is a single number that lets you compare the model to anything it did not make. That has now been supplied from outside, by Artificial Analysis, which ran MiniMax-M3.1-Flash through its standard evaluation suite and published on 19 October 2026.

• Intelligence Index — 42, ranked fourth of 173 models on the tracker

• Cost per Intelligence Index task — $0.22, against $0.51 for MiniMax-M3 and $0.28 for GLM-5.3-Flash

• List price — $0.15 per million input tokens, $0.60 per million output, $0.015 per million cached reads

• Blended price — $0.10 per million on the tracker's 7:2:1 cache-hit/input/output mix

• Context window — 1,000,000 tokens

• Output speed — 125.5 tokens per second, with a 1.04-second time to first token

• Verbosity — 180 million output tokens consumed across the Index, against a median of 99 million

• Input modalities — text and image, per the tracker; M​iniMax's own API reference also lists video

The rank is the weak part of the headline. Fourth of 173 sounds like a podium, and in the narrow sense that the tracker compares models of every size and price, it is. But the three models above it — DeepSeek V4.1 Flash 1017 at 45, and two larger frontier systems — are not its competition in a procurement decision. The models it will actually be shortlisted against are the ones at the same price point, and there the picture is a dead heat with GLM-5.3-Flash and a narrow loss to DeepSeek's newest Flash.

42 twice, for two different reasons

GLM-5.3-Flash and MiniMax-M3.1-Flash both score 42 on the same index, which makes the tie less useful than it looks. The two get there by opposite routes.

GLM-5.3-Flash is an open-weight model: 320 billion total parameters, 18 billion active, MIT licence, weights on Hugging Face since its 26 August 2026 release. It runs at 78 tokens per second with a 2.52-second time to first token, costs $0.15 per million in and $0.50 out, and burns $0.28 per index task. MiniMax-M3.1-Flash is closed today — M​iniMax says weights "will be released later" — runs at 125.5 tokens per second, and burns $0.22 per task. So the tie is a tie on capability and a split decision on everything a production team cares about: one you can download and quantise, the other is 60% faster and 20% cheaper per task.

DeepSeek V4.1 Flash 1017, released two days earlier on 17 October, is the harder comparison. It scores 45, runs at 222 tokens per second — nearly double — and costs $0.30 in and $1.20 out, exactly double M​iniMax's rate card, with a 98% cache discount that pulls its cost per task down to $0.26. It is also MIT-licensed open weights at 552 billion total and 16 billion active parameters. On raw capability and raw speed it wins outright. On money per completed task M​iniMax is still ahead, by four cents, which is a margin thin enough that caching behaviour and output length will decide it in real workloads rather than the list price will.

Worth saying plainly: the older DeepSeek V4.1 Flash, released 10 September 2026, scores 39 — below MiniMax-M3.1-Flash. The "1017" suffix matters. Anyone comparing against the September checkpoint is comparing against a weaker model.

The vendor benchmarks, and what they are worth

M​iniMax published six evaluations alongside the launch. None of them are independently reproduced, and the post gives no comparison column, so the numbers are useful mainly as a profile of what M​iniMax thinks this model is for.

• BankerToolBench — 76.2, an enterprise-workflow suite covering multi-step planning, document consistency and chained tool calls

• DeepSWE v1.1 — 73.7%, code understanding, feature work, bug fixing and test validation

• OSWorld V2 — 61.7, computer operation across applications; M​iniMax claims first among Chinese models

• Agents' Last Exam — 29.6, long-horizon professional tasks; same "first among Chinese models" claim

• CyberGym — 82%, vulnerability discovery and validation

• AutomationBench-AA — described as a top performer, no score published

That last line is the tell. AutomationBench is the benchmark underwriting the "agentic" positioning, and it is the one number the post declines to print. The other five are real and specific, but a benchmark score without a comparison set tells you what a model can do in a lab and nothing about whether it beats the model you are already paying for. Note also that the model was in testers' hands for three weeks as MiniMax-M3.1-Flash-Preview before this launch, and no independent evaluation of that checkpoint was published either — the first third-party measurement of any M3.1 model is the 42 that appeared on launch day.

Why the token price and the task price disagree

Here is the trap in this model, and it is not a small one. MiniMax-M3.1-Flash's rate card is exactly half of MiniMax-M3's: $0.15/$0.60 against $0.30/$1.20 up to 512,000 input tokens. The natural inference is that a workload costs half as much. It does not, for two compounding reasons.

The first is verbosity. On the Intelligence Index the model consumed 180 million output tokens where the median model consumes 99 million. It thinks at the default effort level of max, and thinking tokens are billed as output and count against max_tokens. Run the arithmetic on a workload that is output-heavy: at 180 million output tokens, MiniMax-M3 at $1.20 per million bills $216. Ten million at $0.60 bills $108 — the halving holds. So the rate is honest; the volume is what moves.

The second reason is that you cannot turn thinking off at all. M​iniMax's API reference is explicit: sending thinking: {"type": "disabled"} returns HTTP 400, and reasoning_effort: "none" is rejected the same way. The five-rung ladder — low, medium, high, xhigh, max — is the only control you get, and omitting the field means max. MiniMax-M3, by contrast, lets you disable thinking entirely and defaults to it being off on the Anthropic-compatible endpoint.

The practical consequence: for a classification, extraction or formatting job where you were paying for sixty output tokens and no reasoning, MiniMax-M3 with thinking disabled at $1.20 per million output can still undercut MiniMax-M3.1-Flash at $0.60, because the newer model will spend a few hundred thinking tokens before it answers. For agentic runs — multi-step tool use, long coding sessions, computer-use loops — the newer model wins, and the independent cost-per-task number confirms it: $0.22 against $0.51 is a 57% cut, not a 50% one, because the shorter trajectories more than pay for the extra reasoning.

The right way to read a launch like this is per completed task, not per token, and the tracker's $0.22 is the only figure in this article that already does that arithmetic for you.

Getting the cost right without rewriting your stack

Cost per task is a routing decision more than a purchasing decision. The models in this comparison are not interchangeable on a single axis: MiniMax-M3.1-Flash is fast and cheap per task but closed; GLM-5.3-Flash is the same capability with downloadable weights; DeepSeek V4.1 Flash 1017 is faster and stronger for twice the token rate. Most teams end up running two of the three, splitting traffic by task type.

That split is the case OrcaRouter exists for: one API key across 200+ models, with provider list prices passed through at 0% markup, so when M​iniMax halves a rate card the number on your invoice moves the same day rather than at the next contract renewal. MiniMax-M3.1-Flash itself is not on our catalogue yet — it is callable today through M​iniMax's own API, M Plan and M​iniMax Code, and nowhere else we can vouch for. What is already routable on OrcaRouter is the rest of the shortlist: MiniMax-M3, GLM-5.3-Flash and DeepSeek V4.1 Flash 1017, all three on one key, so the comparison above can be run on your own traffic instead of on a benchmark.

Automatic failover matters more than usual when the newest model in a line is also the one with the least operational history. MiniMax-M3.1-Flash is two days old, its predecessor's migration path is still in flight, and the vendor has promised a more capable Pro model later — which is exactly the situation where you want a second model behind the first in the routing table rather than a rewrite in three weeks.

What to watch

Three things will decide whether 42 holds up as the number people quote in six months. The weights, first: M​iniMax has committed to releasing them "later", with no date, and open weights would make this the cheapest-per-task downloadable model in its class overnight. The Pro model, second: M​iniMax has said a stronger version is coming, which usually means the Flash gets repositioned on price. And the third is independent replication of the agentic numbers — OSWorld V2, ALE and CyberGym are exactly the benchmarks that vary most between harnesses, and right now every figure on them comes from the company selling the model.

Until then, the defensible summary is narrow and useful: at $0.22 per task, MiniMax-M3.1-Flash is the cheapest way to buy an Intelligence Index score of 42, it ties GLM-5.3-Flash on that score while being faster and closed, and it loses to DeepSeek V4.1 Flash 1017 on capability and speed while beating it on cost per completed task. If your workload is agentic, that trade is worth making today. If it is a thousand short classification calls, the model it replaced may still be the cheaper one.