MiniMax M3.1 Flash scores 42 on AA Intelligence Index at $0.15/$0.60 per 1M tokens, open weights Oct 23
MiniMax has released M3.1 Flash, a reasoning model that scores 42 on the Artificial Analysis Intelligence Index, up from 29 for M3. It costs $0.15 per 1M input tokens and $0.60 per 1M output tokens, and open weights are planned for October 23.
MiniMax M3.1 Flash — Quick Specs
MiniMax has released M3.1 Flash, a reasoning model that scores 42 on the Artificial Analysis (AA) Intelligence Index. That is 13 points above its predecessor M3 (29), at less than half the cost per task. The API is live now, and MiniMax plans to publish open weights on October 23.
Specs and pricing
| Spec | Value |
|---|---|
| Context window | 1,000,000 tokens |
| Input modalities | Text and images |
| Input price | $0.15 per 1M tokens |
| Cached input price | $0.015 per 1M tokens |
| Output price | $0.60 per 1M tokens |
| Parameter count | Not disclosed in available reporting |
| Training cutoff | Not disclosed in available reporting |
The API is available at platform.minimax.io. MiniMax also offers a desktop app called MiniMax Code.
Benchmark results
The figures below come from Artificial Analysis unless noted.
- AA Intelligence Index: 42. This ties GLM-5.3-Flash, edges Gemini 3.8 Flash (41), and trails DeepSeek V4.1 Flash 1017 (45).
- Humanity's Last Exam: 48%, ahead of DeepSeek (43%) and GLM-5.3-Flash.
- CritPt (research-level physics, 71 problems): 22%. GLM-5.3-Flash scores 15%, Gemini 3.8 Flash 18%, and DeepSeek 24%.
- AA-Omniscience accuracy: 56% correct, versus 43% for DeepSeek and 28% for GLM-5.3-Flash.
- DeepSWE v1.1: up to 73.7% at under $0.30 per task. This figure comes from MiniMax, not an independent evaluation.
Cost per task and speed
AA puts M3.1 Flash at about $0.22 per task. The model uses fewer output tokens than its rivals: roughly 189 million across the AA benchmark suite, versus 224 million for DeepSeek V4.1 Flash 1017. That makes it slightly cheaper per task than DeepSeek ($0.26), even though the two models' per-token prices are not directly compared in the available reporting.
MiniMax claims the model runs three times faster than M3. That figure has not been independently verified.
Weakness: hallucination rate
AA measures a 93% hallucination rate for M3.1 Flash. The model tries to answer nearly every test question rather than declining when uncertain. Its high AA-Omniscience accuracy (56%) therefore comes with a strong tendency to give wrong answers when it does not know.
Versus DeepSeek V4.1 Flash 1017
Neither model wins across the board. DeepSeek leads on the Intelligence Index (45 vs. 42) and CritPt (24% vs. 22%). MiniMax leads on Humanity's Last Exam (48% vs. 43%), AA-Omniscience accuracy (56% vs. 43%), and cost per task ($0.22 vs. $0.26). DeepSeek's weights are already on Hugging Face. MiniMax's will follow on October 23.
What this means
M3.1 Flash moves MiniMax from the back of the Flash-tier pack to parity with GLM-5.3-Flash and within 3 points of DeepSeek. The main gain is efficiency: a 13-point index jump alongside a lower cost per task suggests the improvement came from more economical reasoning, not just a stronger model.
The 93% hallucination rate matters for builders. Pipelines that need calibrated refusals, such as retrieval, compliance, or customer-facing answers, will need guardrails or verification layers. The 1M-token context and image input widen the use cases, but reliability is the constraint.
The open-weight release on October 23 is the next checkpoint. Self-hosted results will show whether the AA numbers and MiniMax's speed claims hold outside the first-party API.
Related Articles
Mistral releases Timestral-209M, a time series model claiming top zero-shot scores on three benchmarks
Mistral AI has released Timestral-209M, the largest of four time series foundation models (5M, 28M, 123M, 209M parameters). According to the model card, it leads GIFT-Eval, TIME and fev-bench in zero-shot forecasting, with a context of up to 32,768 time steps and 21-quantile probabilistic output.
Tencent releases Hy-Memory-30B-A3B, a 3B-active model that writes agent memory for Mem0 and MemOS
Tencent Hunyuan released Hy-Memory-30B-A3B, a 30B-parameter mixture-of-experts model with about 3B active parameters, built to run the write path in external memory frameworks. Tencent reports a 55.49 cross-harness average, ahead of Claude Opus 5 (54.62) and DeepSeek V4.1 Flash (53.64) on its own fixed 20% benchmark subsamples.
DeepSeek V4.1 Flash 1017 launches with 552B CED backbone, 1M context, $0.15/$0.60 per 1M tokens
DeepSeek has released V4.1 Flash 1017, an October 17, 2026 snapshot built on its new Causal Encoder-Decoder architecture. It uses a 552B-parameter sparse MoE backbone with a 1M-token context window, native image understanding, and pricing of $0.15 input / $0.60 output per 1M tokens on OpenRouter.
Black Forest Labs releases FLUX 3 Image Klein, a 5.6B-parameter open-weights image model
FLUX 3 Image Klein is a 5.6B-parameter open-weights image model covering text-to-image, multi-reference generation and image editing. Its developer claims it leads its own blind evaluations in all three tasks. Pricing, license terms and benchmark data were not disclosed in the announcement.
Comments
Loading...