On October 19, 2026, MiniMax made its AI model MiniMax M3.1 Flash generally available. "Space Bunny," a model that ran anonymously and processed more than 200 trillion tokens, was reportedly an early version of it. The company emphasizes the model's ability to develop software and carry out tasks that span multiple apps. It says total usage costs for agent work are down 70% compared with the previous generation, and that online inference is about three times faster. However, low per-token prices, output speed, and the cost and time required to complete a task should each be evaluated separately. A calculation using the company's published prices makes the difference clear.
Cached input costs one-tenth as much, but total API costs don't fall by the same proportion
According to MiniMax's official October 19 announcement, M3.1 Flash API pricing per million tokens is US$0.15 for standard input, US$0.60 for output, and US$0.015 for cached input. Tokens are the units AI uses to process text and other information, and they don't necessarily match character counts. Cached input refers to the portion that reuses content that was previously loaded, and it is billed at a lower rate than standard input.
| Billing item | Price per million tokens |
|---|---|
| Standard input | US$0.15 |
| Output | US$0.60 |
| Cached input | US$0.015 |
The cached input price is one-tenth of standard input, but total API costs do not fall by the same proportion. For example, with one million input tokens and one million output tokens, the total without caching is 0.15 + 0.60 = US$0.75. If all of the input is cached, the total is 0.015 + 0.60 = US$0.615. The difference of US$0.135 equals 18% of the original cost. In other words, even though the input price drops by 90%, output accounts for 80% of the original cost under these conditions, so the total cost falls by only 18%.
This 18% figure is a calculation in which the same model's input and output volumes are held constant and only the input billing category changes; it is not a measurement of actual usage. Caching will not necessarily apply to all input, and the calculation excludes taxes, external tool fees, and labor costs. Meanwhile, MiniMax's "70% reduction" refers to the total usage cost of agent work compared with the previous generation, so its comparison target and basis differ from the 18% calculation. The official announcement does not specify which version of the previous-generation model was used for comparison, the input-to-output ratio, or detailed measurement conditions, including retries, so you may not get the same level of savings in your own workloads.
Faster output doesn't necessarily mean shorter task completion times
MiniMax says M3.1 Flash's output speed exceeds 200 tokens per second in everyday use, and exceeds 170 tokens per second at the median (P50). It credits the speedup to even greater sparsity, which narrows computation to the parts that are needed. However, because P50 is the median of the measured values, it does not guarantee a speed of 170 tokens per second or more for every request.
Faster output speed directly helps only the step in which the model generates text or code. For example, when a software development agent fixes code, it may test the generated code and then revise it again based on the results. Because waiting for external tool responses and redoing fixes also take time, even if inference becomes about three times faster, the total time to complete a task will not necessarily shrink to one-third. Latency until the first output is returned also needs to be measured separately from output tokens per second.
According to MiniMax's official model list, M3.1 Flash also supports multimodal input and has a context window of one million tokens. The depth of reasoning can reportedly be adjusted as well. Because the context window indicates how much information the model can reference at once, being able to read a large volume of material is a separate matter from being able to process it quickly and return results.
Performance claims span multi-app operation, software development, and security
What MiniMax stresses about M3.1 Flash is not simply the ability to answer questions but its capability as an agent that combines multiple steps to complete work. The company says it achieved a top-tier result on AutomationBench-AA, which evaluates enterprise workflow automation, citing tasks in sales, marketing, customer support, finance, and HR that span multiple SaaS products. It also reported a score of 76.2 on BankerToolBench, which relates to financial work. That benchmark evaluates multi-step planning, consistency across documents and files, compliance with financial regulations, and sequential tool use.
In software development, it reports 73.7% on DeepSWE v1.1. According to MiniMax, this benchmark covers complex tasks that a model carries out autonomously, from understanding existing code to building new features, fixing bugs, and verifying through testing. It assesses whether the model can work through a long process while checking its own results, rather than generating code once and stopping.
MiniMax has also published benchmark results for computer operation, professional work, and cybersecurity. The main figures are as follows.
| What is evaluated | Benchmark | MiniMax's published figure |
|---|---|---|
| Computer operation and cross-app tasks | OSWorld V2 | 61.7 |
| Real-world professional and complex work | Agents’ Last Exam (ALE) | 29.6 |
| Vulnerability discovery and verification | CyberGym | 82% |
OSWorld V2 evaluates computer operation and tasks across apps, while ALE deals with complex tasks close to real professional work. Because the two differ in what they evaluate and how they score, you cannot simply compare 61.7 and 29.6 to judge the model's strengths and weaknesses. What matters is choosing tests that resemble your own tasks and checking how well the model performs on them. Likewise, the 82% result on CyberGym measures the ability to discover and verify vulnerabilities and does not guarantee the security of any individual system.
MiniMax also says that in evaluations by its in-house security experts, the model showed strong performance in code auditing, vulnerability remediation, penetration testing, malware analysis, and reverse engineering. However, these are measurements or self-assessments by the provider and should be distinguished from results reproduced by third parties under the same conditions. For companies considering adoption, it is important to use these benchmarks as a reference while measuring success rates on tasks close to their own work.
API now available; model weights to be released later
M3.1 Flash is already available through the MiniMax API, as well as through M Plan and MiniMax Code. The company says the model weights will be released at a later date, and a Pro version aimed at even higher performance is also planned. Being able to use a model through an API is different from being able to deploy and run it on your own infrastructure. The current official announcement does not specify a release date or licensing terms for the weights, so anyone considering self-hosting will need to check the contents and terms of use once they are published.
When comparing API usage in the meantime, it is useful to measure not just the per-token price but the actual cost of completing real tasks. For example, dividing the total API charges spent on the same task by the number of successes gives the "API cost per success." This calculation must include charges for failed runs and retries in the total, and it cannot be computed if the number of successes is zero. Note that this is a suggested evaluation method for comparing cost-effectiveness in actual operation, not a metric for reproducing the 70% cost reduction MiniMax claims.
What matters in judging M3.1 Flash's competitiveness is how much its low token prices and high output speed actually help in real work. If it can cut API cost per success and completion time while maintaining the same success rate on the same tasks, more work could potentially be entrusted to AI agents. Going forward, attention will focus on real-world cost-effectiveness as well as the terms under which the model weights are released.