By ByteBot
MiniMax launched M3.1 Flash today — a 428-billion-parameter mixture-of-experts model that beats Claude Opus 5 and GPT-5.6 Sol on enterprise automation benchmarks at $0.15 per million input tokens. The headline is accurate. What MiniMax doesn’t put in the headline: a 93% hallucination rate on independent reliability testing. Before you start migrating agentic pipelines, you need both numbers.
What MiniMax Is Claiming
On AutomationBench-AA — a benchmark for complex enterprise agent workflows — M3.1 Flash tops all comers, leaving Claude Opus 5 at 50.3 and GPT-5.6 Sol at 45.8 behind. It scores 73.7% on DeepSWE v1.1 (software engineering), 82% on CyberGym (vulnerability discovery), and 76.2 on BankerToolBench (multi-step financial planning). The model runs with a 1-million-token context window — five times larger than the previous M2 family — enabled by MiniMax Sparse Attention, and outputs at 120 tokens per second with a 1.05-second time-to-first-token. Cost per agentic task runs about 70% lower than its predecessor.
The model also has a backstory worth noting. It ran anonymously as “Space Bunny” for weeks, processing over 200 trillion tokens before today’s official launch. MiniMax confirmed the connection. Model weights drop on October 23.
What Independent Testing Found
Artificial Analysis ran M3.1 Flash through its full evaluation suite and the picture gets more complicated. On the Artificial Analysis Intelligence Index, the model scores 42 out of 100 — 13 points above MiniMax’s previous M3, but below DeepSeek V4.1 Flash at 45 and within a few points of GLM-5.3-Flash. That puts it in a crowded mid-tier, not a clear lead.
The more important finding is on AA-Omniscience, the benchmark that measures factual reliability. M3.1 Flash attempted to answer nearly every question put to it, producing a 93% hallucination rate. It gets 56% of factual questions right — better than DeepSeek V4.1 Flash’s 43% correct — but when it doesn’t know something, it invents rather than refuses. That’s a design choice, not a bug, and it’s the opposite of what you want in a production agent. Artificial Analysis also flagged the verbosity signal: M3.1 Flash generated 180 million output tokens during evaluation against a median of 99 million. The model talks past uncertainty instead of stopping at it.
Why This Matters for Agents Specifically
In a single-turn chatbot, a confident wrong answer is annoying. In a multi-step agentic pipeline, it compounds. An agent that’s 95% reliable per step is only about 59% reliable across a 10-step workflow. At a 7% reliability rate per step — which a 93% hallucination rate implies for uncertain inputs — the math gets brutal fast. Production agentic workloads that touch document retrieval, financial data, or decisions that require the model to know what it doesn’t know are genuine risk territory here. The Hacker News discussion — 304 points and climbing — captured the concern cleanly: a developer invoked Goodhart’s Law within the first few comments, and the thread stayed skeptical.
Who Should and Shouldn’t Use It
M3.1 Flash has a real use case. If you’re running high-volume pipelines where outputs feed into a verification layer before any action is taken — software scaffolding generation, code diff proposals, cybersecurity scan automation — the cost and speed profile is attractive. DeepSWE at 73.7% and CyberGym at 82% aren’t marketing numbers for those domains. The model ran as Space Bunny in production for weeks; the infrastructure is real.
Where it breaks down is anywhere your agent needs to reason about its own uncertainty. Customer-facing agents, legal document summarization, financial analysis, RAG systems where citation accuracy drives decisions — these are not the places to find out your model won’t say “I don’t know.” The community consensus from developers who tested Space Bunny was blunt: DeepSeek V4.1 Flash worked more smoothly in practice.
The Bigger Picture
MiniMax is competing in a pricing tier that’s already crowded. DeepSeek V4.1 Flash sits at the same $0.15/M input price and leads on overall capability index (45 vs. 42). China’s AI price race has no visible floor — six of the top ten AI models globally now come from Chinese labs, and the price war has driven some workloads from $788 per month on GPT-5.5 to under $11 on the cheapest Chinese alternatives. That’s good for developers — provided the reliability story holds. With M3.1 Flash, it holds for some workloads and breaks for others. The weights arrive October 23; self-hosted developers will have more to say by next week.
The benchmark win on AutomationBench-AA is real. The 93% hallucination rate is real. Run both numbers against your specific pipeline before you move anything to production.
