MiniMax M3.1 Flash: cheap agent model that approaches DeepSeek with high hallucination rate

The MiniMax M3.1 Flash was officially launched by the Chinese company MiniMax on October 19, 2026. According to the company’s announcement, the model represents a generational leap in agentic tasks (complex workflows), programming, computer use and cybersecurity. It promises nearly three times faster online inference than its predecessor and a 70% reduction in usage costs for agent tasks.

The MiniMax M3.1 Flash is an evolution of what previously ran anonymously as “Space Bunny,” processing more than 200 trillion tokens. It is available via the MiniMax API, the M plan and the MiniMax Code app. Open weights are expected to be released later (the company mentioned October 23 in reports).

What the benchmarks say about MiniMax M3.1 Flash

In the official announcement, MiniMax highlights for M3.1 Flash:

  • 73.7% on DeepSWE v1.1 (complex coding);
  • 76.2 on BankerToolBench (multi-step financial workflows);
  • 61.7 on OSWorld 2.0 and 29.6 on Agents’ Last Exam (computer use);
  • 82% on CyberGym (vulnerability discovery).

Independent tests by Artificial Analysis give the MiniMax M3.1 Flash an Intelligence Index of 42 (versus 29 for the previous M3). This places it close to DeepSeek V4.1 Flash-1017 (45) and GLM-5.3 Flash (42), and ahead of Gemini 3.8 Flash (41) in some contexts. The cost per task on the index dropped from about $0.51 to $0.22.

There is an important caveat: on Artificial Analysis’s AA-Omniscience benchmark, the model answered 56% of factual questions correctly but showed a 93% hallucination rate. That means when it did not know the correct answer, it almost always invented one instead of refusing or expressing uncertainty. This matters for agents operating autonomously in enterprise environments.

Pricing and availability of MiniMax M3.1 Flash

Pricing of the MiniMax M3.1 Flash is aggressive: $0.15 per million input tokens, $0.60 per million output tokens and $0.015 for cached input. The company claims output speeds above 200 tokens per second in everyday use (median 170+). Context reaches 1 million tokens and accepts text and images.

These numbers reinforce the price race among low-cost Chinese models aimed at agents and automation, competing directly with DeepSeek and other Flash models. At the same time, the high hallucination rate shows the classic trade-off between speed/cost and factual reliability.

Sources

Transparency: This content was created, edited or reviewed with the assistance of artificial intelligence. Information was cross-checked with public posts on X and available internet sources. Consult the original sources to verify the full context.

By GeekikiBot