MiniMax released its M3.1 Flash model on Monday, claiming it reduces agent costs by 70 percent, though independent testing suggests the savings may be smaller. The company published new pricing at $0.15 per million input tokens and $0.60 per million output tokens, half the cost of its predecessor, and says inference runs nearly three times faster than M3. The 70 percent cost reduction depends partly on per-token rates and partly on how many tokens an agent uses to complete a task, but MiniMax hasn't disclosed which workloads it tested to reach that figure.
Independent analysis from Artificial Analysis found the average cost per Intelligence Index task at $0.22 for M3.1 Flash, down from $0.51 for M3, a 57 percent drop that falls short of MiniMax's claim. A separate study of token use in agentic coding tasks found that repeated runs of identical tasks can consume up to 30 times as many tokens, with input tokens making up most of the usage. On coding benchmarks, M3.1 Flash scored 73.7 percent on DeepSWE v1.1, just behind the 74 percent MiniMax lists for Claude Opus 5 and a point ahead of GPT-5.6 Sol at 73 percent, while VentureBeat places DeepSeek V4.1 Flash 1017 at 77 percent. On Artificial Analysis's Intelligence Index, M3.1 Flash scored 42, tied with GLM-5.3 Flash, ahead of Gemini 3.8 Flash at 41 and behind DeepSeek V4.1 Flash 1017 at 45. The model scored 61.7 on OSWorld V2 for computer use, and MiniMax reports 82 percent on CyberGym along with strong ratings from internal security experts for code auditing, patching, penetration testing, malware analysis, and reverse engineering, though the launch post doesn't explain how those evaluations were conducted.
The most notable finding comes from Artificial Analysis's AA-Omniscience benchmark, where M3.1 Flash answered 56 percent of factual questions correctly, ahead of DeepSeek V4.1 Flash at 43 percent. According to The New Stack's coverage, the model attempted nearly every question, and about 93 percent of its non-correct responses were wrong answers rather than abstentions, with an overall Omniscience score of 16. Artificial Analysis summarized the result as high accuracy offset by a high hallucination rate, and The Decoder described the model as attempting every test question without declining when uncertain.
The discrepancy between MiniMax's 70 percent savings claim and Artificial Analysis's 57 percent reduction matters because cheaper pricing only delivers value if the agent completes tasks in fewer loops. The study showing 30-times variation in token consumption across identical tasks makes it hard to assess real-world savings without knowing which specific workloads MiniMax tested. DeepSeek V4.1 Flash offers the same $0.15 input and $0.60 output rates during off-peak hours, but doubles those prices during weekday peak hours from 01:00 to 04:00 and 06:00 to 10:00 UTC, meaning teams in New York or San Francisco working daytime hours stay in DeepSeek's off-peak window while overnight batch jobs hit peak rates. MiniMax's cached input costs $0.015 per million tokens against DeepSeek's $0.003, and since the study found input tokens drive most agent costs, the cache rate may matter more than headline pricing. The model supports a one-million-token context window, fitting large codebases in a single prompt, and the weights are reportedly coming on October 23, though MiniMax's own post just says "later" without publishing the license yet, an important consideration for anyone planning to run the model on their own hardware.
MiniMax says a stronger Pro model is also on the way with no date yet, and the company confirmed that Space Bunny, the mystery model that processed more than 200 trillion tokens, was an early build of M3.1 Flash. The high hallucination rate means businesses using the model for knowledge-heavy work might want to add verification, retrieval, guardrails, or human review before deploying it in production. The model's willingness to answer nearly every question without declining when unsure creates risk in applications where accuracy matters more than coverage, making the 93 percent wrong-answer rate on non-correct responses a critical factor for deployment decisions. The gap between promised and measured savings, combined with the aggressive answer behavior, suggests organizations should run their own evaluations on their specific workloads before committing to migration.