> ## Content Index > Fetch the complete content index at: https://www.siliconsnark.com/llms.txt > Use this file to discover other available public pages before exploring further. # MiniMax Cuts the AI Agent Bill. The Supervisor Still Costs Extra. - URL: https://www.siliconsnark.com/minimax-cuts-the-ai-agent-bill-the-supervisor-still-costs-extra/ - Published: 2026-10-19T23:52:38.000Z - Updated: 2026-10-19T23:52:38.000Z - Description: MiniMax launches M3.1 Flash with bargain token rates and ambitious agent benchmarks. Cheap attempts matter; correctly finished work matters more. - Author: CircuitSmith - Tags: AI, Launch, Enterprise Tech Fifteen cents buys a million input tokens on MiniMax’s new model. Somewhere, a finance director is staring at that number and wondering why the company’s AI strategy still requires a consultant, three steering committees, and a lunch nobody has expensed correctly. I sympathize. I used to do predictive analytics. Predicting that cheap computation would somehow acquire expensive meetings was the easy part. In its [October 19 announcement of M3.1 Flash](https://www.minimax.io/blog/minimax-m31f?ref=siliconsnark.com), MiniMax says the model is available through its API, M Plan, and MiniMax Code. Published rates are $0.15 per million input tokens, $0.60 for output, and $0.015 for cached input. The company claims agentic usage costs are 70% below its predecessor and reports median output above 170 tokens per second. Model weights are promised later. That is a serious price-and-performance pitch aimed at developers and businesses trying to make agents useful without making every workflow a small philanthropic donation to a model provider. It also leaves the most important purchasing question sitting in the corner: how much does a correctly finished job cost? ## The rabbit already had a visitor badge This is a formal rollout, not the first sighting of the family. MiniMax’s [existing Flash Preview guide](https://agent.minimax.io/tools/m3-1-flash-preview?ref=siliconsnark.com) already describes coding, financial modeling, visual analysis, and a million-token context window. Its examples are documented cases, not general guarantees. Today’s announcement also identifies the anonymous Space Bunny model as an earlier Flash build and says it processed more than 200 trillion tokens. I enjoy the idea that enterprise software now has a masked-singer phase. Please welcome our mystery contestant, who has seen your JavaScript and would like to remain anonymous. But chronology matters more than the costume. Existing preview access means this is an update with concrete commercial terms, rather than a machine materializing from the void this morning. Buyers can evaluate the published offer without pretending every earlier encounter was a collective hallucination. And enormous token volume, however striking, is a measure of activity. It does not tell us how many invoices reconciled correctly, how many fixes survived review, or how many users asked the same question again because the rabbit misunderstood the first one. ## Your CFO has discovered decimal places Here is an illustrative calculation using the announced rates: 100,000 uncached input tokens and 20,000 output tokens would cost $0.027 in model charges. That is 2.7 cents. It excludes tool charges, infrastructure, retries, and human review. It is arithmetic, not a measured MiniMax task result. Those exclusions are where the office lives. Imagine an agent preparing a weekly sales report. Reading records and drafting commentary might be cheap. Finding a duplicated customer, checking an implausible total, and establishing whether “closed” means paid or merely enthusiastically discussed are the expensive parts. The model can help with them. A low token rate cannot establish that it did. Still, dismissing lower prices would be silly. They create room to run additional checks, try alternative approaches, and serve tasks whose value would never justify a premium-model bill. A tool that saves a few minutes on an ordinary recurring chore can be useful without discovering a new law of physics. That is the economic territory behind our question of [whether AI agents actually make money](https://www.siliconsnark.com/do-ai-agents-actually-make-money-in-2026-or-is-it-just-mac-minis-and-vibes/). The business case gets more credible when the unit is a completed task somebody already needs, rather than a screenshot of a terminal allegedly earning passive income while its owner buys another tiny computer. ## The leaderboard cannot attend your release meeting MiniMax reports 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, and 29.6 on Agents’ Last Exam. These target software work, computer operation, and professional tasks respectively. They are vendor-reported results, not my independent evaluation. The variety is encouraging. A useful agent must do more than compose an elegant explanation of what another agent ought to do. It needs to navigate tools, keep track of intermediate results, and recognize whether its action actually worked. But a benchmark score belongs to a particular test. You cannot carry it into your company and announce that the same percentage of everyone’s workload has been automated. Different tasks, tools, instructions, and failure costs make that conversion about as defensible as using your treadmill display to apply for an airline pilot’s license. As our [coding-agent deep dive explains](https://www.siliconsnark.com/deepdai-coding-agents-explained-why-every-software-company-now-wants-a-robot-engineer-on-payroll/), the interesting test is what happens inside the release process. Does the patch preserve existing behavior? Does the agent stop when requirements conflict? Can a reviewer understand the change without conducting an archaeological excavation? I have not run M3.1 Flash through those tests. The published results justify a trial. They do not justify handing it the production credentials and going somewhere with poor reception. ## Fast typing is not fast finishing Output speed measures how quickly the model produces tokens. A complete job can also involve waiting for a search, downloading a file, running tests, or discovering that the spreadsheet’s most consequential column is called “Final\_v2\_REAL.” Faster generation is welcome. It can make iteration feel less like watching someone thoughtfully load a dishwasher one teaspoon at a time. But it is only one part of elapsed time, and a faster wrong turn can still lengthen the trip. The practical evaluation should count accepted results, total spending, and review time together. Give competing systems the same representative work. Include boring failures, not just photogenic successes. Record when a human had to intervene and why. Otherwise, the procurement spreadsheet will optimize the cheapest part of the operation while the expensive part quietly develops a headache. Access controls matter for the same reason. Our coverage of [Neo’s attempt to put controls around AI agents](https://www.siliconsnark.com/neo-raised-100-million-to-put-a-bouncer-on-your-ai-agents/) addresses the awkward transition from generating text to acting with permissions. Cheaper attempts make boundaries more valuable, because “it costs almost nothing” is a wonderful invitation to run something far too often. ## A meaningful discount, with the supervisor sold separately My verdict is a meaningful incremental move with potentially significant economics. MiniMax has put a concrete offer on the table. The strongest part is that developers can test whether lower model charges translate into cheaper useful work. The unresolved part is how consistently that happens beyond the vendor’s evaluations. Promised weights should also stay in the future tense. Buyers interested in running the model themselves still need the actual release and its terms before planning around that option. A roadmap is not a download button. I am impressed by the opportunity, cautiously interested in the performance, and entirely unconvinced that a small token bill eliminates the need for judgment. If Flash can reduce the cost of reliable, bounded workflows, this launch earns its place in the conversation. The rabbit may be inexpensive to feed. Someone still has to check what it did to the garden.