MiniMax M3.1 Benchmark: M3.1 Flash Scores and Price
MiniMax M3.1 benchmark results for M3.1 Flash from MiniMax, Vals AI and Artificial Analysis, plus price, what it cannot do and which MiniMax model makes video.
Last updated: 2026-10-21
This MiniMax M3.1 benchmark summary collects the scores published for M3.1 Flash, MiniMax's text-output reasoning model, and labels where each number comes from. M3.1 Flash does not make videos. If you want a MiniMax video model, that is MiniMax H3, which you can select in the SceneOak video workspace.
What is MiniMax M3.1 Flash?
According to the official M3.1 Flash page, M3.1 Flash is the generally available release of M3.1-Flash-Preview. It has a 1M-token context window, accepts text, images and video as input, and outputs text only. Reasoning is always on, with five effort levels (Low, Medium, High, Xhigh and Max); MiniMax recommends Max as the default. MiniMax says it is available through the MiniMax Code app, an M Plan subscription and the MiniMax developer platform API, and that model selection depends on your account access and app version.
The official page gives no release date. The Decoder reported the launch on 19 October 2026, and Vals AI lists the release date as 20 October 2026. MiniMax has not published a parameter count. We checked all sources on this page on 21 October 2026; the numbers may change as leaderboards re-run.
Official MiniMax M3.1 Flash benchmark scores
| Benchmark | M3.1 Flash score |
|---|---|
| BankerToolBench | 76.2 |
| DeepSWE v1.1 | 73.7% |
| CyberGym | 82% |
| OSWorld V2 | 61.7 |
| Agents' Last Exam | 29.6 |
| AutomationBench-AA | 62.2 |
These are vendor-reported figures from the official page; AutomationBench-AA comes from its integration guide section. MiniMax states the results reflect the cited evaluations and do not guarantee the same performance on every task. MiniMax also claims online inference is nearly 3 times faster than the previous generation and total cost on agentic tasks is 70% lower, and notes that this is not a comparison of API unit prices.
Independent results: Vals AI and Artificial Analysis
Vals AI scores MiniMax-M3.1-Flash at 53.78% (±1.07) on the Vals Index, rank 15 of 23 models.¹
| Vals AI benchmark | Score |
|---|---|
| Vibe Code Bench v1.1 | 73.94% |
| EMB v1.1 | 67.84% |
| Finance Agent v2.1 | 58.83% |
| Legal Research Bench | 34.13% |
| Tax Agent Bench | 32.14% |
| Terminal-Bench 4.0 | 25.25% |
| HLAB-Vals-1.1 | 7.50% |
We have not checked Artificial Analysis directly. According to The Decoder, Artificial Analysis gives M3.1 Flash an Intelligence Index of 42, up 13 points from M3's 29:
| Model (Artificial Analysis, via The Decoder) | Intelligence Index |
|---|---|
| DeepSeek V4.1 Flash 1017 | 45 |
| MiniMax M3.1 Flash | 42 |
| GLM-5.3-Flash | 42 |
| Gemini 3.8 Flash | 41 |
The same report gives HLE 48%, CritPt 22%, AA-Omniscience accuracy 56%, about $0.22 per Artificial Analysis task, and a 93% hallucination rate measured by Artificial Analysis. Check factual answers before relying on them.
M3.1 Flash vs M3: which numbers belong to which model
Some pages quote scores from the previous model, MiniMax M3, as if they were M3.1 results. The Decoder reported on 1 June 2026 that M3 scores 59% on SWE-Bench Pro according to MiniMax, and 83.5 on BrowseComp. Those are M3 numbers, not M3.1 Flash numbers. The official M3.1 Flash page lists neither benchmark, so the two models cannot be compared on them.
MiniMax M3.1 Flash price
The official pricing page lists MiniMax-M3.1-Flash at $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached-read tokens. No cache-write price is listed for this model.
Can MiniMax M3.1 make videos?
No. M3.1 Flash can take video as input, but it only produces text. MiniMax's video models are listed separately on the same pricing page: the H3 series (MiniMax-H3, H3-Max, H3-Regeneration and H3-Context-IR) and the older Hailuo models (2.3-Fast, 2.3 and 02).
Make a MiniMax H3 video in SceneOak
SceneOak offers MiniMax H3 text-to-video and image-to-video. H3 Max, H3 Regeneration, H3 Context-IR, Hailuo, reference-to-video and multi-image references are not available here.
| Setting | MiniMax H3 in SceneOak |
|---|---|
| Duration | 5–15 seconds |
| Resolution | 768P, 2K (default) or 4K |
| Aspect ratio | Adaptive, 21:9, 16:9, 4:3, 1:1, 3:4 or 9:16 (text-to-video only) |
| Image input | One image, used as the first frame |
| Audio | Native audio |
| Cost (5 s) | 420 credits at 768P, 550 at 2K, 830 at 4K |
Each chat message costs another 10 credits. We have not yet published H3 samples or tested its quality or speed, so this page makes no quality claim. Our showcase currently shows Kling clips with their prompts. For a lower-cost first shot, a 5-second Kling 3.0 render uses 350 credits; see the Kling image-to-video guide, Kling 4.0 status, Kling 4.0 vs 3.0 and plans and credits.
Frequently asked questions
Is MiniMax M3.1 Flash open source?
As of 21 October 2026, this is not confirmed. The Decoder reports that MiniMax plans to release open weights on 23 October; the report does not mention a license. The official page does not mention open weights.
How much does MiniMax M3.1 Flash cost?
On the official API, $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached-read tokens (pricing page checked 21 October 2026).
How does M3.1 Flash compare with GLM-5.3-Flash and Gemini 3.8 Flash?
On the Artificial Analysis Intelligence Index, as reported by The Decoder, M3.1 Flash scores 42, GLM-5.3-Flash 42, Gemini 3.8 Flash 41 and DeepSeek V4.1 Flash 1017 scores 45.
What is the context window of MiniMax M3.1 Flash?
1M tokens, according to the official page.
Is M3.1 better than M3?
On the Artificial Analysis Intelligence Index, as reported by The Decoder, it rose from 29 to 42, while MiniMax also claims nearly 3 times faster inference and 70% lower total cost on agentic tasks; these are vendor claims, not unit-price comparisons.
Does MiniMax M3.1 generate video?
No. It outputs text only. MiniMax's video models are the H3 series and the older Hailuo models.
Does SceneOak run MiniMax M3.1?
No. The MiniMax model on this site is the video model MiniMax H3.
Are the benchmark scores independent?
The official table is reported by MiniMax. Vals AI is a third-party evaluation. The Artificial Analysis numbers here come from The Decoder's report; we have not verified them on Artificial Analysis.
Try a short video project
Open the workspace, choose a model and review the cost before generating. Sign-in is required; opening the workspace alone does not start a paid render.
Open the video workspace¹ The Vals page shows 15 of 23 in its summary and table, and 15 of 22 in its updates section.
Sources checked 21 October 2026: MiniMax M3.1 Flash, MiniMax pricing, Vals AI, The Decoder on M3.1 Flash, The Decoder on M3. SceneOak is an independent service and is not affiliated with MiniMax or fal.