MiniMax M3.1 Flash Review: Pricing, Benchmarks, Limits and DeepSeek Comparison
MiniMax M3.1 Flash delivers strong agentic and coding results at half the prior cost, with 1M context and five reasoning levels, though hallucination rates remain high.
MiniMax released the M3.1 Flash model on October 19, 2026. The 428B-parameter MoE system activates roughly 23B parameters per token and supports a 1M-token context window for text, image, and video. It targets agentic workflows, coding, computer use, and cybersecurity through an always-on reasoning engine with five effort settings.
Access runs through the MiniMax API, M Plan subscription, and MiniMax Code app. Open weights are scheduled for October 23, 2026. An earlier anonymous build processed over 200T tokens during development.
MiniMax M3.1 Flash Pricing and Limits
The API charges a flat $0.15 per million input tokens and $0.60 per million output tokens regardless of context length. Cached input tokens cost $0.015 per million. These rates cut costs by roughly 70 percent versus the previous M3 model on agentic tasks and deliver inference speeds above 200 tokens per second at peak.
Third-party measurements show output speeds between 93 and 123 tokens per second. No surcharges apply above 512K input, unlike the prior version. Production users should test cache-hit ratios, because real savings depend on repeated prompts rather than one-off queries.
Typical limits include mandatory reasoning on every request and a 93 percent hallucination rate observed when the model attempts nearly every question instead of abstaining. The system offers no non-thinking mode, which can increase output token counts on simple tasks.
Benchmarks and Performance Data
Vendor tests place MiniMax M3.1 Flash at 62.2–62.8 on Automation-Bench v1.0.6, ahead of Opus 5 at 50.3. It scores 76.2 on BankerToolBench, 73.7 percent on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam, 82 percent on CyberGym, and 60.3 on LOCA-Bench 256K. Coding results remain mixed across independent suites.
Artificial Analysis Intelligence Index scores range from 42 to 43, a 13-point gain over the prior M3. The model reaches 48 percent on Humanity’s Last Exam and 22 percent on CritPt. Factual accuracy on AA-Omniscience hits 56 percent. Cost per indexed task averages $0.22.
MiniMax M3.1 Flash vs DeepSeek V4.1 Flash
Both models are low-cost 1M-context MoE systems released within days of each other. DeepSeek V4.1 Flash lists identical base rates off-peak but doubles them during peak hours and offers cheaper cache hits at roughly $0.003 per million. MiniMax maintains flat pricing at all times.
| Benchmark | MiniMax M3.1 Flash | DeepSeek V4.1 Flash |
|---|---|---|
| Automation-Bench | 62.2–62.8 | 60.5 |
| DeepSWE v1.1 | 73.7% | 77.0% |
| BankerToolBench | 76.2 | — |
| Artificial Analysis Index | 42–43 | 45 |
| Humanity’s Last Exam | 48% | 43% |
| Hallucination rate | 93% | Lower |
DeepSeek leads on several coding benchmarks such as Terminal-Bench 4.0 and NL2Repo-Bench. MiniMax ties or leads on agentic and business-oriented tests and uses fewer output tokens in some scenarios. DeepSeek already ships MIT-licensed weights; MiniMax weights arrive later in October 2026. Users running frequent coding workloads may prefer DeepSeek, while teams focused on automation and tool-use benchmarks can achieve lower effective costs with MiniMax when cache utilization is high.
Frequently asked questions
What is the context window of MiniMax M3.1 Flash?
The model supports a 1M-token context window for text, image, and video inputs. This limit applies across the API and MiniMax Code app.
How does MiniMax M3.1 Flash pricing compare to the prior M3 model?
Rates sit at $0.15 per million input and $0.60 per million output with no length surcharges. This represents roughly half the cost of M3 on long agentic tasks.
Does MiniMax M3.1 Flash require reasoning on every query?
Yes. Reasoning runs by default at maximum effort and cannot be disabled. Five tunable levels exist but the feature stays active at all times.
Which model wins on coding benchmarks between MiniMax and DeepSeek V4.1 Flash?
DeepSeek V4.1 Flash records higher scores on DeepSWE v1.1 and Terminal-Bench 4.0. MiniMax leads or ties on several agentic and business benchmarks instead.
Comments 6