MiniMax made agent work cheaper and faster. The receipt still starts with the task

THE SIGNAL IN ONE SENTENCE

Fast is useful. Cheap is useful. Neither is a completed agent task. MiniMax released M3.1 Flash on October 19 with the sort of numbers that make an API spreadsheet sit up straight. The company lists a flat price of $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. It says everyday output can exceed 200 tokens per second, with median speed above 170 tokens per second. The company also says agentic task costs are 70 percent lower than the model's predecessor and online inference is nearly three times faster. Those are company claims. They are concrete enough to test, which is much more useful than a cloud of adjectives. The plain signal is this: M3.1 Flash could make long, tool-heavy workflows cheaper and less annoying, but the meaningful unit is not a token. It is a task completed correctly, safely and with a receipt. That distinction matters because agent work creates costs that a token price cannot see. An agent may read ten files, open a browser, call a sales tool, revise a spreadsheet, ask another model for help, retry a failed action and wait for a person to approve the final step. A fast first answer can still lead to an expensive run if the model loops, loses state, calls the wrong tool or leaves cleanup for a human. The cheap model is the one that completes the job with the fewest failures. MiniMax positions M3.1 Flash for exactly these longer workflows. Its release page highlights complex automation, coding, computer use and cybersecurity. The company reports 76.2 on BankerToolBench, 73.7 percent on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam and 82 percent on CyberGym. Each number describes a particular test setup. None is a universal agent grade. BankerToolBench is meant to exercise planning, documents, files, compliance and chained tool calls. DeepSWE focuses on software work that includes understanding code, changing it and validating the result. OSWorld V2 tests computer operation across applications. Agents' Last Exam targets professional tasks with economic value. CyberGym tests vulnerability discovery and validation. That spread is interesting because it pushes beyond trivia and polished chat. It asks a model to keep a goal intact while the environment changes around it. It also creates more places for a benchmark harness to matter. A score can change with the prompt, tool descriptions, allowed retries, context window, computer image, network access, scoring rule and the amount of hidden help in the wrapper. A chart that puts several models in one row is not proof that every model received an identical operating environment. MiniMax says its scores place M3.1 Flash alongside leading models and first among Chinese models on two computer-use tests. The release does not turn those comparisons into an independent audit. The company is reporting its own model in its own launch material. That is normal. It is also why the buyer should bring a frozen local task set. Start with twenty jobs that resemble the work the agent will actually do. Include easy jobs, common jobs, expensive jobs and the weird failures everybody remembers. Freeze the files, prompts, tool permissions, browser state and acceptance criteria. Run each model with the same retry cap and the same rule for asking a human. Then count complete outcomes. A useful receipt should record the exact model identifier, endpoint date, prompt template, tools exposed, number of calls, input tokens, cached tokens, output tokens, elapsed time, retries, tool errors, human-review minutes and whether the result passed. That final pass is the important part. Suppose Model A costs half as much per output token but succeeds on fourteen of twenty tasks. Model B costs twice as much but succeeds on eighteen. Model A might still win if failures are cheap to repair. It might lose badly if each failure sends an employee into a thirty-minute cleanup. The receipt converts marketing prices into operating prices. M3.1 Flash also arrives in an awkward but common middle state. It is available through the MiniMax API, M Plan and MiniMax Code, while the company says downloadable weights will come later. Available now does not mean inspectable in every way. API users can test behavior and economics immediately, but they cannot yet reproduce the serving stack from released weights or inspect how the model behaves under their own inference configuration. Teams that require local deployment, sovereign hosting or a pinned checkpoint should treat the future weight release as a separate event. Do not build today's approval memo around tomorrow's file. The same caution applies to speed. More than 200 output tokens per second sounds quick because it is quick. Median speed above 170 tokens per second gives a more useful center than a single peak. Neither number tells a team how long a complete workflow takes under its own concurrency, location, rate limits and tool latency. The model may finish thinking before the browser finishes loading. Measure first-token latency, sustained output speed and task completion time separately. A coding assistant that starts instantly but spends five minutes running tests has a different bottleneck from a customer-service agent that waits eight seconds before saying hello. An operations agent may spend most of its life waiting on SaaS APIs that the model provider does not control. Speed can also change behavior. A faster model makes more retries feel cheap. That can be good when retries are intentional and bounded. It can be terrible when an agent repeatedly clicks, resubmits or probes a system without understanding why the last action failed. Cheap loops are still loops. Give each task a call budget, retry budget and stop condition. Require the agent to name the error before trying again. Make write actions idempotent where possible. Separate reading from changing. Put confirmation between preparation and submission. This is especially important for the cybersecurity claims. MiniMax says its internal security experts saw strong performance in code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering. The release also reports an 82 percent CyberGym score. Those capabilities can help defenders. They also belong behind authorization, isolated test environments, logging and explicit scope. A benchmark result is not permission to scan somebody else's system. The useful deployment is a bounded security assistant working on code and infrastructure the operator is authorized to test. It should preserve evidence, distinguish a suspected flaw from a reproduced flaw and keep a person in the loop before any potentially disruptive action. The same principle fits computer use. An agent that can navigate a desktop should be judged on what happens when the button moves, the session expires, the file name changes or a dialog appears in front of the expected window. Success on a clean benchmark desktop is not a promise about a busy finance machine with old browser tabs and a nervous human moving the mouse. Test the interruption. MiniMax also says an early version of M3.1 Flash operated anonymously as Space Bunny and processed more than 200 trillion tokens. That is an eye-catching history note, not a reliability metric. Token volume does not disclose the mix of tasks, error rate, user population, safety incidents or the share of work that required correction. Scale is evidence of traffic. Quality needs a denominator. The attractive thing about M3.1 Flash is not that it settles the model race. It makes a class of experiments cheaper. Small teams can run more complete comparisons. Open-source projects can test an API model without spending the month's budget in one afternoon. Researchers can study long agent traces. Companies can reserve expensive models for tasks where local evidence shows the difference is worth paying for. The routing policy should remain boring. Send the task to the lowest-cost model that has passed that task class. Escalate when a check fails, uncertainty crosses a threshold or the action carries real consequence. Record the route. Do not let a model choose a more powerful model merely because it got stuck. There is a practical opportunity here for Chinese developers and for teams elsewhere. MiniMax lists official contacts in Beijing, Shanghai and Shenzhen while presenting itself as a global foundation-model company. Its release is another sign that low API prices and rapid agent inference are becoming competitive features, not side notes. That competition can make useful automation more accessible. It can also tempt teams to treat a favorable token line as a complete business case. Resist the temptation. Pick a real workflow. Freeze the harness. Count human repair. Keep the stop signs. Save the receipt. If M3.1 Flash is genuinely faster and cheaper on the work that matters, the local task ledger will make the case without needing the launch adjectives.