MiniMax Introduces M3.1 Flash: A Budget-Friendly Option for AI Agents
MiniMax entered the competitive AI model market with a new offering. The launch happened on Monday. The model promises superior performance over the previous version. Company representatives state it operates nearly three times faster. It reduces agent task costs by 70%. The prototype, known as „Space Bunny,” processed over 200 trillion tokens before the public release.
Breaking news
How long can you expect your original Steam Deck to last?
Google Unveils TorchTPU, Targeting Nvidia’s AI Dominance
Rust Rewrite of a JavaScript Compiler Targets AI Agents
AI-Assisted Gaming Mods Create Wild Character CrossoversThe model is available immediately via the MiniMax API, desktop coding environment, and subscription plans. It targets enterprise workflows involving multiple steps. The pricing structure is aggressive. Standard input tokens cost $0.15 per million. Output tokens cost $0.60 per million. Cached reads are priced at just $0.015 per million. The producer highlights competitive parity with frontier systems on specialized agent benchmarks. On Automation-Bench v1.0.6, M3.1 Flash scored 62.8 across 600 multi-application business workflows. It slightly outperformed DeepSeek V4.1 Flash-1017, which recorded 60.5. On BankerToolBench, the model reached 76.2 points. This equals Claude Opus 5 in financial compliance and tool chain execution. On OSWorld 2.0, the score of 61.7 placed it ahead of domestic Chinese rivals.
Independent evaluations by Artificial Analysis gave M3.1 Flash an intelligence index of 42 at the maximum reasoning level. This reduced the average benchmark task cost from $0.51 on M3 to approximately $0.22. However, these evaluations highlight a critical question for purchasing firms. Can lower inference costs offset the extra effort needed to verify agent outputs?
Independent reports reveal significant operational hurdles. Artificial Analysis identified a concerning trend. Although M3.1 Flash answered more than half of factual questions correctly, the model showed an unusual propensity to invent false answers instead of acknowledging uncertainty. On the AA-Omniscience index, the score was only 16. Technical frictions further complicate the advertised savings. While MiniMax attributes average output speeds exceeding 170 tokens per second to its sparse attention architecture, faster short-generation speeds only shorten brute-force generation cycles. Multi-step agents frequently stall waiting for third-party tool responses. They also struggle during automated test runs or error recovery. Prompt cache savings depend on the input-output token mix per task, prompt reuse frequency, and the number of tool calls needed to complete a job.
The fundamental tension in low-cost agent models
The fundamental tension in low-cost agent models lies between raw compute prices and the real cost of human oversight. Lower token rates offer limited value if an automated system introduces errors in accounting software or sends code changes requiring extensive human review. When execution costs drop to a few cents, firms will inevitably run more automated tasks. However, uncontrolled volume amplifies the verification burden. If an agent performs hundreds of operations with unexplained errors, the human work needed to audit the workflow rapidly erodes the initial API savings.
For engineering teams and enterprise buyers, M3.1 Flash may offer a cheaper option for bounded, easily verifiable tasks. Examples include draft code generation, preliminary log analysis, and structured data processing. Its suitability for more complex or sensitive workflows will depend on reliability testing and the required level of supervision. Organizations evaluating the model should consider safety measures. These include human approval for sensitive actions, schema validation for tool calls, and custom evaluations using representative company data. Instead of comparing models solely on advertised token prices, technical teams should measure cost per successful task. This includes retries, error recovery, and human verification. This provides a better indication of whether lower inference costs translate into real operational savings.
Additional news: Apple’s i Phone Duo features a dual-cell 4,883mAh battery, placing it between the i Phone 18 Pro and Pro Max in terms of capacity.
