We may earn from vendors via affiliate links or sponsorships. This might affect product placement on our site, but not the content of our reviews. See our Terms of Use for details.
China’s MiniMax is taking aim at the cost of autonomous AI, introducing a new model it says can handle agent tasks faster and more cheaply than its predecessor.
The company launched M3.1 Flash on Monday, claiming the model operates nearly three times faster than its previous version while reducing agent task costs by 70%. MiniMax also revealed that its anonymously tested model, known as “Space Bunny,” had processed more than 200 trillion tokens before the official release.
The launch adds another low-cost contender to the increasingly competitive AI model market. But independent evaluations raise an important question for businesses: whether lower inference costs can offset the additional work needed to verify an agent’s output.
Benchmarks and the price ledger
Available immediately through MiniMax’s API, desktop coding environment, and subscription plans, M3.1 Flash targets multi-step enterprise workflows. MiniMax prices standard input tokens at $0.15 per million, output at $0.60 per million, and cached reads at $0.015 per million.
The vendor highlights competitive parity with leading frontier systems on specialized agent benchmarks. On Automation-Bench v1.0.6, M3.1 Flash posted a 62.8 score across 600 multi-application business workflows, edging out DeepSeek V4.1 Flash-1017’s 60.5. It also scored 76.2 on BankerToolBench, tying Claude Opus 5 on financial compliance and chained tool executions. On OSWorld 2.0, it logged 61.7, leading domestic Chinese rivals.
Independent evaluations from Artificial Analysis gave M3.1 Flash an Intelligence Index score of 42 at maximum reasoning, bringing the evaluator’s average benchmark task cost down from $0.51 on M3 to approximately $0.22.
Independent assessments reveal substantial operational hurdles. Artificial Analysis tracked a concerning tendency: while M3.1 Flash answered over half of its factual inquiries correctly, it demonstrated an unusually high propensity to invent false answers rather than acknowledge uncertainty. On the evaluator’s AA-Omniscience index, it scored 16.
Technical friction also complicates advertised savings. MiniMax attributes median output speeds exceeding 170 tokens per second to sparse attention architecture. Yet faster generation speeds only shorten raw generation cycles. Multi-step agents frequently stall while awaiting third-party tool responses, running automated tests, or recovering from execution errors.
The savings from prompt caching will depend on each workload’s mix of input and output tokens, how often prompts are reused, and the number of tool calls required to complete a task.
The true cost of cheap autonomy
The fundamental tension of low-cost agent models lies between raw compute pricing and the downstream expense of human oversight. Lower token rates offer limited value if an automated system introduces errors into accounting software or submits code changes that require extensive human review.
When execution costs drop to pennies, businesses inevitably run higher volumes of automated tasks. However, unchecked throughput amplifies verification debt. If an agent executes hundreds of operations with unacknowledged errors, the human labor required to audit the pipeline quickly erodes initial API savings.
More must-read AI coverage
- SS&C Intralinks DealCentre AI vs. Datasite: Which platform is built for the future of dealmaking?
- SS&C Intralinks FundCentre AI vs. Juniper Square: Which platform better supports modern private markets fund managers?
- Why Data, Not Models, Determines AI Success
- The Rise of the AI-Native Factory: How Physical AI Is Transforming Manufacturing
What this means for developers and enterprises
For engineering teams and enterprise buyers, M3.1 Flash could offer a lower-cost option for bounded, easily verifiable tasks such as draft code generation, preliminary log parsing, and structured data processing. However, its suitability for more complex or sensitive workflows will depend on reliability testing and the level of oversight required.
Organizations evaluating the model should consider safeguards such as human approval for sensitive actions, schema validation for tool calls, and custom evaluations using representative company data.
Rather than comparing models solely by advertised token prices, technical teams should measure the cost per successful task, including retries, error recovery, and human verification. That provides a more meaningful indication of whether lower inference costs translate into real operational savings.
Other news: Apple’s iPhone Duo packs a 4,883mAh dual-cell battery, placing it between the iPhone 18 Pro and Pro Max in capacity.