Chinese artificial intelligence company MiniMax has launched M3.1 Flash, an updated AI model designed to make coding, enterprise automation and computer-use agents faster and cheaper to operate. Announced on October 19, 2026, the model builds on the company’s earlier M3 system and introduces improvements in inference speed, task execution and cybersecurity-related work. MiniMax claims M3.1 Flash can run nearly three times faster than its predecessor while reducing overall usage costs for AI-agent tasks by 70%.
The launch targets a growing challenge for businesses adopting AI agents: these systems often need multiple model calls, intermediate reasoning steps and repeated interactions with software to complete a single task. Lower token prices can help, but faster inference and fewer calls may also reduce the total cost of completing work. MiniMax says M3.1 Flash is available through its API, M Plan subscription and MiniMax Code desktop application. However, its claimed cost savings are not a guarantee for every workload, and independent evaluations indicate that factual reliability remains an important consideration.
Key Takeaways
- New model: MiniMax launched M3.1 Flash on October 19, 2026.
- Lower costs: The company claims a 70% reduction in overall AI-agent usage costs compared with its predecessor.
- Faster inference: Online inference is reportedly nearly three times faster, with everyday output speeds exceeding 200 tokens per second at peak.
- Developer pricing: The API costs $0.15 per million input tokens and $0.60 per million output tokens.
- Agent capabilities: The model targets coding, enterprise workflows, computer use and cybersecurity tasks.
- Multimodal support: It accepts text, images and video, with a context window of up to one million tokens.
- Availability: Developers can access it through the MiniMax API, M Plan and MiniMax Code; model weights are planned for a later release.
- Reliability caveat: Independent testing has raised concerns about the model’s tendency to give incorrect factual answers rather than acknowledge uncertainty.
What Is MiniMax M3.1 Flash?
M3.1 Flash is the latest addition to MiniMax’s M3 model family, which is designed for demanding reasoning and agentic tasks. Unlike a conventional chatbot that primarily answers individual questions, an AI agent can plan a sequence of actions, use tools, interact with software and work toward a larger goal.
For example, an AI agent might review a collection of documents, extract relevant information, update a spreadsheet and prepare a report. A coding agent might inspect a software repository, identify a bug, edit the relevant files and run tests. Each task can involve several model calls and substantial amounts of input and output.
This makes efficiency particularly important. A model that responds quickly and uses fewer tokens to complete a task can reduce operating costs for developers running agents at scale.
MiniMax says M3.1 Flash improves performance in coding, computer use, enterprise workflows and cybersecurity. It also retains a large context window, allowing developers to provide extensive instructions, documents or code within a single interaction.
The model has approximately 428 billion total parameters, according to its developer documentation. Its design uses increased sparsity during inference to improve efficiency. In practical terms, this approach aims to reduce the computation required to produce an answer without proportionately reducing the model’s overall capabilities.
How MiniMax Claims to Cut AI Agent Costs by 70%
The headline cost reduction concerns the overall usage cost of agentic tasks, not simply a 70% reduction in every API price.
AI agents can make repeated calls to a model while planning, checking their work, using tools and correcting mistakes. The cost of a completed task therefore depends on the number of calls, the tokens consumed, the model’s pricing and the amount of additional computation required.
MiniMax says M3.1 Flash reduces the total cost of agentic tasks by 70% compared with its predecessor. It also claims that inference speed has nearly tripled.
The company attributes the speed improvement partly to more aggressive sparsity at inference time. Its published information says everyday output speeds can exceed 200 tokens per second at peak, while median speeds can remain above 170 tokens per second in the stated usage conditions.
These figures could matter for applications that require frequent model interactions. Faster responses can reduce waiting time, while lower per-task costs can make it more practical to run agents across larger volumes of work.
However, developers should distinguish the advertised improvement from the savings they will achieve in production. A complex task may still require numerous calls, expensive external tools or human review. A model that is cheaper per token may not deliver the same percentage reduction in total project costs.
The actual benefit will depend on the workflow, model settings, context size, task-completion rate and reliability requirements.
MiniMax M3.1 Flash API Pricing
MiniMax has published the following API prices for M3.1 Flash:
| API usage | Price per million tokens |
|---|---|
| Input tokens | $0.15 |
| Output tokens | $0.60 |
| Cached input tokens | $0.015 |
Input tokens represent the text and other supported content supplied to the model, while output tokens represent the content it generates. Cached input pricing applies when eligible input can be reused under the API’s caching rules.
For illustration, a request involving one million input tokens and one million output tokens would cost $0.75 at the listed uncached rates. This is a simplified calculation rather than a typical single request, and it excludes external tools, infrastructure and other application expenses.
| Example workload | Input cost | Output cost | Total |
|---|---|---|---|
| 100,000 input + 100,000 output tokens | $0.015 | $0.060 | $0.075 |
| 1 million input + 1 million output tokens | $0.15 | $0.60 | $0.75 |
| 10 million input + 10 million output tokens | $1.50 | $6.00 | $7.50 |
Calculations use the published API rates and exclude any other fees.
Output tokens are more expensive than input tokens, which means applications that generate long responses or substantial intermediate reasoning may incur higher costs. Developers should estimate both input and output usage before comparing models.
MiniMax’s pricing is particularly relevant to businesses running high-volume agents, where small differences in the cost of each completed task can accumulate across thousands of operations.
Coding and Enterprise Automation Capabilities
MiniMax is positioning M3.1 Flash for tasks that require more than generating short pieces of text. These include software development, navigating business applications and completing multistep workflows.
The company reports improvements on several benchmarks:
| Benchmark | Reported M3.1 Flash result | What it evaluates |
|---|---|---|
| OSWorld V2 | 61.7 | Operating a computer through software interfaces |
| BankerToolBench | 76.2 | Financial workflows and tool use |
| DeepSWE v1.1 | 73.7% | Software engineering tasks |
| CyberGym | About 82% | Software vulnerability discovery and validation |
These are reported benchmark results. Scores may not be directly comparable across different models unless the evaluation conditions, versions and settings are equivalent.
Computer-use benchmarks assess whether an AI system can navigate interfaces and perform actions rather than simply describe what a person should do. This is an important capability for agents designed to work with business software.
Coding benchmarks, meanwhile, evaluate how effectively a model can handle software engineering problems. The results can provide useful signals about performance, but they do not guarantee that an agent will safely complete every real-world coding task.
MiniMax also highlights enterprise workflows involving sales, marketing, operations, customer service, finance and human resources. These tasks can require coordinating information across multiple applications, keeping documents consistent and following a sequence of actions.
For businesses, the potential value is reduced manual work. The practical challenge is ensuring that agents operate within permissions, respect confidential data and ask for human approval when an action could have significant consequences.
Cybersecurity Is Another Target
M3.1 Flash is also being marketed for cybersecurity work, including code auditing, vulnerability analysis and security testing.
MiniMax reports a score of approximately 82% on CyberGym, a benchmark focused on whether AI systems can reproduce and validate software vulnerabilities. Such capabilities can help security teams examine code, investigate potential weaknesses and prioritise remediation work.
AI-assisted security analysis can be useful because large software projects contain extensive codebases that are difficult to inspect manually. An agent may help identify suspicious patterns, explain a potential vulnerability or suggest a patch for human review.
But cybersecurity is a dual-use field: the same capabilities that help defenders identify weaknesses can also be misused. Automated testing must therefore be performed within authorised environments and with appropriate safeguards.
Benchmark performance should not be interpreted as proof that the model can independently secure an organisation or reliably identify every vulnerability. Real-world effectiveness depends on the quality of the code, the tools available, the testing environment and the expertise of the people supervising the system.
The Space Bunny Connection
MiniMax also revealed that an earlier version of M3.1 Flash had operated anonymously under the name “Space Bunny” on the OpenRouter model-routing platform.
According to the company, the model processed more than 200 trillion tokens while its identity remained undisclosed. Community testers had linked the anonymous model to MiniMax before the company confirmed the connection during the M3.1 Flash launch.
The episode illustrates how model-routing platforms and developer communities can help new AI systems gain exposure before their developers formally announce them.
Large volumes of usage may provide valuable feedback about how a model performs across different prompts and applications. However, the reported token total alone does not establish how many users adopted the model, how satisfied they were or whether the workload represented paid production usage.
MiniMax has not published a detailed breakdown of the anonymous model’s usage in the available launch material.
Model Weights and Availability
M3.1 Flash is available through three main routes: the MiniMax API, the company’s M Plan subscription and MiniMax Code desktop application.
The company says it plans to release the model weights later. Until the weights and their licensing terms are available, developers should not assume they can download, modify or self-host the model.
Open weights can give developers greater control over deployment and customisation, but the degree of freedom depends on the licence. The release of weights is also separate from API access, which allows applications to use a hosted model without managing the underlying infrastructure.
MiniMax has also said that a more capable Pro model is planned for a later date, but it has not confirmed a launch date in the announcement.
Reliability Remains a Critical Question
Lower cost and faster inference are valuable only if a model can complete tasks reliably enough for the intended application.
Independent testing reported by VentureBeat found that M3.1 Flash showed a strong tendency to produce incorrect factual answers instead of expressing uncertainty. That finding highlights a limitation that matters when agents are expected to make decisions or execute actions with minimal supervision.
For low-risk tasks, an incorrect answer may be relatively easy to detect and correct. In financial workflows, cybersecurity, legal analysis or automated changes to production software, the consequences can be much more serious.
Businesses considering M3.1 Flash should test it on their own workloads, track task-completion rates and measure how often human intervention is needed. They should also implement permissions, verification steps and approval gates for consequential actions.
The most useful comparison is not simply price per million tokens. It is the total cost of achieving a correct, verified outcome.
The Bigger Picture
MiniMax M3.1 Flash reflects the growing competition to make capable AI agents affordable enough for everyday business use. Its claimed combination of faster inference, low API prices and improved coding and computer-use performance could appeal to developers building software assistants and enterprise automation systems.
Yet the launch also demonstrates why cost alone is an incomplete measure of AI performance. Benchmark results, factual reliability, security safeguards and the ability to finish real tasks successfully all influence whether a cheaper model delivers genuine value. M3.1 Flash’s long-term position will depend on how its reported improvements translate into practical deployments.
Looking Ahead
Developers can evaluate M3.1 Flash through the MiniMax API, M Plan and MiniMax Code. The next milestones include the promised model-weight release and the company’s planned Pro model. Independent testing and real-world deployments should help clarify whether the claimed 70% reduction in agent costs holds across different workloads and whether the model’s capabilities are reliable enough for sensitive applications.
For businesses, the opportunity is to run more automated tasks at a lower cost without compromising accuracy or security. The strongest use cases will likely be those where the model’s work can be checked, permissions can be limited and failures can be corrected before they cause damage. As competition among AI developers intensifies, those operational measures may matter as much as headline speed and pricing.
Frequently Asked Questions
1. What is MiniMax M3.1 Flash?
M3.1 Flash is a MiniMax AI model designed for coding, enterprise automation, computer use and cybersecurity tasks. It is intended to make AI-agent workflows faster and less expensive.
2. How much cheaper is M3.1 Flash?
MiniMax claims that overall usage costs for agentic tasks are 70% lower than with its predecessor. Actual savings depend on token consumption, task complexity and how many calls an agent needs to complete its work.
3. What does the MiniMax M3.1 Flash API cost?
The published prices are $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens.
4. Can developers run M3.1 Flash locally?
The model is currently available through the MiniMax API, M Plan and MiniMax Code. MiniMax says the model weights will be released later, but developers should wait for the release and licensing terms before assuming local deployment is possible.
5. Is M3.1 Flash suitable for coding agents?
Its reported benchmark results suggest it can be useful for software engineering tasks. Developers should still test the model against their own repositories and require code review and testing before applying changes to important systems.
6. Is M3.1 Flash reliable enough for autonomous business tasks?
That depends on the task and the safeguards in place. Independent evaluations have raised concerns about factual reliability, so consequential actions should use verification, restricted permissions and human oversight.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
