Comparisons
MiniMax M3.1 Flash vs DeepSeek V4.1 Flash
Same measurement run, same providers, two very different trade-offs — and one place where the vendor-supplied data caveat matters most.
The headline
M3.1 Flash is cheaper per task and scores higher on the composite index. DeepSeek V4.1 Flash is roughly twice as fast and already ships open weights. Neither wins outright, and the interesting differences are in the columns the headline numbers hide.
Same-run comparison
Both rows come from the same Artificial Analysis run on 21 October 2026, so the numbers are comparable to each other.
| Metric | MiniMax M3.1 Flash | DeepSeek V4.1 Flash (Max) |
|---|---|---|
| Intelligence Index | 42.71 | 39.46 |
| Humanity’s Last Exam | 47.96% | 39.25% |
| AA-LCR | 79.67% | 84.00% |
| GDPval-AA | 47.88% | 55.00% |
| Terminal-Bench 4.0 | 25.76% | 26.77% |
| Automation-Bench (partial) | 72.52% | 68.89% |
| Omniscience accuracy | 56.2% | 46.4% |
| Omniscience hallucination rate | 92.81% | 96.46% |
| Input / output per 1M | $0.15 / $0.60 | $0.30 / $1.20 |
| Cached input per 1M | $0.015 | $0.006 |
| Weighted cost per task | $0.22 | $0.571 |
| Time per task | 615 s | 283 s |
| Median output speed (long prompts) | 104.7 t/s | 233.5 t/s |
| Context window | 1,000,000 | 1,000,000 |
| Weights | Not published | Open, MIT licence, 552B parameters |
| Release date | 2026-10-19 | 2026-09-10 |
Where each one wins
M3.1 Flash wins on cost per finished task. $0.22 against $0.571 is a 2.6× difference, and it comes from both a lower token price and a lower reasoning bill. For a workload that runs hundreds of agent tasks a day, this is the number that decides the month.
M3.1 Flash wins on the composite and on automation. The Intelligence Index gap is 42.71 against 39.46, Automation-Bench is 72.52% against 68.89%, and HLE is 47.96% against 39.25%.
M3.1 Flash wins on grounding in hard facts. Accuracy on Omniscience is 56.2% against 46.4%, and the hallucination rate is lower (92.81% against 96.46%). Both numbers are bad in absolute terms — the practical lesson is that neither model should be trusted on long-tail facts without checking.
DeepSeek wins on throughput. A median 233.5 tokens per second against 104.7 is more than twice the speed on long prompts, and it is consistent across prompt lengths. If your bottleneck is wall-clock time on a long agent run, that matters.
DeepSeek wins on open weights. It is MIT licensed with 552B parameters published on Hugging Face, while M3.1 Flash has no public weights at all — see the Ollama page for the three-source check on that.
DeepSeek wins on price-sensitive reverse caching. Its cached input price ($0.006) is lower than M3.1 Flash’s $0.015, though both are far below their standard input rates.
DeepSeek wins on the business-document and professional-knowledge evaluations. GDPval-AA is 55.0% against 47.88%, and AA-LCR is 84.0% against 79.67%.
The caveat that matters here
Artificial Analysis flags M3.1 Flash’s performance and speed data as first-party — supplied by MiniMax — while the same battery for DeepSeek is measured against the provider’s API. In practice this means the speed comparison above is partly a comparison of two vendors’ claims, not of two independently instrumented pipelines.
We are keeping the numbers because they are the best same-run data available, and flagging them so you do not over-read the 2× throughput gap.
Choosing
| If you need | Pick |
|---|---|
| Lowest cost per agent task | M3.1 Flash |
| Long-context work on a large codebase | M3.1 Flash (flat price, 1M context, no tier step-up) |
| Fastest wall-clock completion of long runs | DeepSeek V4.1 Flash |
| Self-hosting or on-premise deployment | DeepSeek V4.1 Flash |
| Business-document and knowledge work | DeepSeek V4.1 Flash |
| Better grounding on hard facts | M3.1 Flash, with verification either way |
A note on the “Flash” naming
Both vendors ship a tier called Flash, and several other labs do too. The label means fast and cheap, not weak — but it also means the comparison above is between two specific, dated models rather than between categories. Re-check the numbers when either side ships a new revision.
Next: benchmarks and their reliability, or pricing in detail.
Sources and version notes
Checked October 21, 2026. We label official claims, third-party measurements, our own testing and community reports separately — see the sources policy for what each label means.