MiniMax M3.1 Flash vs DeepSeek and Opus 5: Price, Benchmarks, and a 93% Hallucination Rate

The short version

On October 19, 2026, Shanghai-based MiniMax released MiniMax M3.1 Flash, a 428B-parameter mixture-of-experts model with roughly 23B active parameters and a 1M-token context window. Launch pricing is $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens — exactly half the rate of its predecessor, M3.

The headline numbers are real: on MiniMax's own business-workflow benchmarks, M3.1 Flash outranks Claude Opus 5 and DeepSeek V4.1 Flash-1017 at a fraction of the cost. But independent evaluation from Artificial Analysis shows the model trails on coding, and — more importantly — that it almost never admits uncertainty when it doesn't know something.

That combination makes M3.1 Flash an excellent choice for a narrow set of jobs and a risky one for others. This comparison breaks down where the line falls.

Specs and pricing: what actually changed from M3

From MiniMax's launch post and the Hugging Face model card:

Architecture: 428B total parameters, ~23B active (MoE)
Context window: up to 1,000,000 tokens
Modalities: text, image, and video input; text output
Attention: MiniMax Sparse Attention (MSA) applied throughout the model
Speed: DSpark speculative decoding draft head included; MiniMax claims 200+ tokens/sec with a median above 170 TPS
Reasoning effort: always enabled, cannot be turned off (M3 allowed toggling)
Deployment: available via MiniMax API, the M Plan subscription, and the MiniMax Code desktop app

The price comparison against its own predecessor is straightforward:

MiniMax M3: $0.30 input / $1.20 output per million tokens
MiniMax M3.1 Flash: $0.15 input / $0.60 output per million tokens
Cached input: $0.015 per million tokens

MiniMax says this cuts overall usage cost for agentic tasks by 70% versus the prior generation, with inference nearly three times faster.

One detail worth noting: MiniMax disclosed that a model operating anonymously as "Space Bunny," which processed more than 200 trillion tokens, was an early version of M3.1 Flash. The company used an unnamed production deployment to tune the model before revealing what it was.

Vendor benchmarks: strongest on workflow automation, weaker on code

Here are MiniMax's published results, side by side with the competitors the company chose to compare against. All figures are vendor-reported and were run under MiniMax's own harnesses, so treat them as claims rather than independent measurements.

Automation-Bench v1.0.6 (cross-SaaS business workflows): 62.8 — vs. DeepSeek V4.1 Flash-1017 at 60.5, Claude Opus 5 at 50.3
BankerToolBench (multi-step financial tool chains): 76.2 — tied with Claude Opus 5
DeepSWE v1.1 (practical software engineering): 73.7 — vs. DeepSeek V4.1 Flash-1017 at 77.0, Opus 5 at 74.0
Terminal-Bench 4.0 (terminal task execution): 38.1 — vs. Opus 5 at 51.8
NL2Repo-Bench (build a repo from a spec): 61.2 — vs. DeepSeek 68.2, Opus 5 75.3
OSWorld 2.0 (computer use): 61.7
Agents' Last Exam: 29.6
CyberGym (vulnerability discovery and validation): 82.2

The shape of these results matters more than any single number. Where the task is "click the right buttons in the right order across business software" — sales, finance, support, HR — M3.1 Flash is genuinely at the top of this price class. Where the task is "write code, iterate in a terminal, build a repository from a specification" — it loses clearly to DeepSeek's same-priced model, and loses badly to Opus 5 on terminal work.

MiniMax's own effort chart shows the same trade-off from the other direction. Automation-Bench scores fall from 62.8 at max effort to 61.2 at xhigh, 59.2 at high, 57.7 at medium, and 52.8 at low. If you cut reasoning effort to save money, the automation capability you bought the model for is the first thing to erode.

Independent testing: Intelligence Index 42, $0.22 per task

Artificial Analysis, which runs its own evaluations rather than repeating vendor numbers, paints a more measured picture.

Intelligence Index: 42, up 13 points from M3's 29
Cost per Intelligence Index task: ~$0.22, down from M3's ~$0.51 — a 57% reduction, smaller than MiniMax's claimed 70%
Humanity's Last Exam: 48%, up from M3's 39%, ahead of GLM-5.3 Flash (40%) and DeepSeek V4.1 Flash-1017 (43%)
Closest competitors: GLM-5.3 Flash at 42 (~$0.28/task), DeepSeek V4.1 Flash-1017 at 45 (~$0.26/task), Gemini 3.8 Flash at 41 (~$1.24/task)
Total output tokens across the Index evaluation: ~189 million, versus ~324 million for GLM-5.3 Flash and ~334 million for Gemini 3.8 Flash

That last figure is the one most teams miss. M3.1 Flash's low per-task cost isn't purely a function of its price list — it reaches comparable results while generating substantially fewer output tokens. Since reasoning tokens are billed as output, a model that thinks more concisely produces savings that never appear on the rate card.

Reading the 93% hallucination figure correctly

This is the part that needs careful handling. On AA-Omniscience, Artificial Analysis's benchmark for factual recall and hallucination, M3.1 Flash answered 56% of questions correctly — better than DeepSeek V4.1 Flash-1017 at 43% and GLM-5.3 Flash at 28%. Its raw knowledge is the best of the three.

Yet the model attempted nearly every question, producing a 93% hallucination rate and an overall AA-Omniscience score of just 16.

The 93% does not mean 93% of its answers are wrong. The metric measures the proportion of non-correct responses that were outright wrong rather than partial answers or refusals. In plain terms: when M3.1 Flash didn't know something, it almost never said so — it invented a plausible answer instead.

Note also that hallucination rates vary enormously by benchmark design. A hallucination-specific benchmark measured the same model at 7.5%. The 93% and the 7.5% describe different phenomena measured with different denominators. Before quoting either, check what the benchmark actually counts.

Where to use it, and where not to

Good fits:

Cross-SaaS workflow automation — CRM updates, ticket triage, report drafting
Financial tool chains (it ties Opus 5 on BankerToolBench)
High-volume batch processing where $0.22 per task changes the economics
Long-document analysis (it leads LOCA-Bench 256K at 60.3)

Poor fits:

Long-horizon coding with terminal iteration (Terminal-Bench 38.1)
Generating repositories from specifications (NL2Repo 61.2)
Knowledge tasks where correctness flows into business decisions without human review
Cutting reasoning effort to squeeze cost (automation performance degrades first)

An adoption checklist

Benchmark on your own tasks. Published figures use different harnesses and different execution environments for each model in the comparison table. DeepSWE alone shows M3.1 Flash at 73.7 under MiniMax's run and DeepSeek at 77.0 under DeepSeek's. Run 10–20 of your real tasks and measure both accuracy and cost per task.

Force abstention into your prompts. Require citations, instruct the model to decline when retrieved context doesn't support an answer, and gate outputs on structured evidence. Given the AA-Omniscience finding, this is not optional.

Don't over-cut reasoning effort. If you need to lower costs, exploit cached input at $0.015 per million tokens rather than degrading effort level.

Check the weights and license before planning self-hosting. At launch the weights were not published and licensing terms were not specified. Hugging Face hosting and third-party quantization efforts exist, but confirm the license before committing infrastructure.

Mix models by function. Automation on M3.1 Flash, coding and final judgment on Opus 5 or GPT-6 class models is a defensible split — and it's cheaper than standardizing on one.

Bottom line

MiniMax M3.1 Flash is best understood as a model optimized hard for structured business automation at a very low price — not as a general-purpose bargain. On Automation-Bench and BankerToolBench it is genuinely top-tier, and its $0.22 per-task cost is the strongest number in its class. On coding and terminal work it trails a same-priced DeepSeek model, and on knowledge questions it has a documented habit of answering confidently when it should abstain.

Pick it for the workflows it was tuned for, measure before you commit, and put an abstention mechanism between it and anything that writes to production systems.