MiniMax M2.5 Puts a $1 an Hour Price on AI Work, With Strong Benchmarks and Uneven Results

A cheaper AI worker enters the competition

MiniMax, the Shanghai artificial intelligence startup, has released M2.5 with a striking proposition: an AI model that can write software, conduct research and create office documents while costing roughly $1 an hour at a generation speed of 100 tokens per second. As of February 14, 2026, its combination of low prices and strong test results is putting pressure on the economics of enterprise AI.

The company presents M2.5 as a digital employee capable of completing work, rather than simply answering questions. Its advertised prices are roughly one tenth to one twentieth those of competitors including Claude Opus 4.6 and GPT-5.2. At a slower operating speed, MiniMax puts the hourly cost at about $0.30.

M2.5 has been released on Hugging Face under a modified MIT License. That gives organizations access to the model's weights, the numerical settings learned during training, and allows them to run it on their own infrastructure. Commercial users must display the name "MiniMax M2.5" prominently in their user interfaces.

Investors responded strongly. MiniMax shares, traded in Hong Kong as MINIMAX-WP under stock code 00100.HK, rose between 11% and 16% following the announcement and closed at HK$680. The release came during a busy week of Chinese AI launches that shook global markets and intensified competition across software, robotics and video generation.

The attraction is straightforward: if capable AI becomes cheaper to run, businesses can assign it more work. MiniMax estimates that four agents operating continuously for a year would cost approximately $10,000, an amount described as barely enough for a month of comparable operation with Western models. Its benchmark results also show why businesses will need to assess performance against their own tasks.

How close is M2.5 to the leading coding models?

M2.5 scored 80.2% on SWE-Bench Verified, a widely used test of software engineering ability based on problems from real software projects. The result placed it 0.6 percentage points behind Anthropic's Claude Opus 4.6, ahead of OpenAI's GPT-5.2 at 80%, and ahead of Google's Gemini 3 Pro at 78%.

For a reader unfamiliar with these tests, the central idea is simple: the model receives a programming problem and must produce a working solution. A higher score means it successfully completes a larger share of the benchmark's tasks. The surrounding tools and testing setup also influence how effectively it can work.

M2.5 performed particularly well on Multi-SWE-Bench, which examines programming across languages and tasks requiring an understanding of multiple code repositories. Its score of 51.3% exceeded Claude Opus 4.6's 50.3% and Gemini 3 Pro's 42.7%. On VIBE-Pro, an internal test of complex development covering both the visible interface and underlying software, MiniMax reported performance on par with Claude Opus 4.5.

Fresh coding problems produced a different ranking. The SWE-rebench January 2026 leaderboard, using a different method and 48 fresh GitHub pull requests, recorded 39.6% for M2.5, compared with 52.9% for Claude Code. On that evaluation, Claude Opus 4.6 scored 51.7% and GPT-5.2 scored 51.0%.

A pull request is a proposed change to a software project. Testing recent proposals gives another view of how a coding agent handles new work. Here, the Western systems retained an advantage on the freshest code, while M2.5's stronger SWE-Bench Verified result showed that its competitive position varies with the tasks and evaluation setup.

Browsing, tool use and independent measurements

M2.5's results extend beyond programming. It scored 76.3% on BrowseComp, which tests difficult web browsing and information retrieval, and 76.8% on the Berkeley Function Calling Leaderboard, known as BFCL, for tool use across multiple conversational turns.

The BFCL result was more than 13 percentage points above Claude Opus 4.6. Tool use matters because an agent often needs to do several things in sequence: retrieve information, call a software function, inspect the result and choose its next action. A useful answer may depend on completing that chain successfully.

Artificial Analysis, an independent evaluation firm, assigned M2.5 a score of 42 on its Intelligence Index, compared with an average of 25 for comparable models. It also measured a generation speed of 80 tokens per second, above the comparison group's average of 55.

Tokens are the pieces of text a language model reads and produces. They can be words, parts of words or punctuation. More tokens per second generally means text arrives faster, while a charge per million tokens determines how much that processing costs.

MiniMax advertises 100 tokens per second for its faster API variant. The independent measurement of 80 tokens per second provides another observed speed figure. Together with the coding evaluations, these results give businesses several measures to consider: task success, tool handling, response speed and price.

A large model that activates a smaller part

M2.5 contains 230 billion parameters but activates only 10 billion for a given task. Its Mixture of Experts architecture, usually shortened to MoE, is central to the company's explanation of its efficiency.

Parameters are numerical settings that shape a model's responses. In a conventional explanation of MoE, different parts of the model act like specialists, with a smaller selection used for each piece of work. That reduces the computation needed at any moment while preserving access to a much larger collection of learned capabilities.

MiniMax's design therefore seeks the reasoning capacity of a large model with the speed and operating costs of a smaller active system. The full model still contains 230 billion parameters, an important consideration for organizations planning to host the weights themselves.

The training effort also reached across many programming languages. MiniMax lists Go, C, C++, TypeScript, Rust, Kotlin, Python, Java, JavaScript, PHP, Lua, Dart and Ruby, more than 10 languages in total. Training covered more than 200,000 environments, supporting its ambition to handle varied software projects rather than a narrow type of coding task.

How Forge trains the model to finish work

MiniMax developed Forge, its own reinforcement learning framework, specifically for agentic AI. Reinforcement learning means improving a system through feedback about its actions. In Forge, the reward focuses on completing tasks rather than producing answers that merely sound pleasing.

The company describes training across hundreds of thousands of simulated working environments, including code repositories, web browsers and office applications. Those settings let the model practice taking actions and checking results in situations resembling actual work.

Forge achieves approximately 40 times the training speed of standard reinforcement learning approaches, according to MiniMax. Its methods include asynchronous scheduling, which lets work proceed without every process waiting for the others, and tree structured merging of training samples, which organizes and combines the examples used for learning.

MiniMax engineer Olive Song said on the ThursdAI podcast that the team spent two months on intensive training across varied practical scenarios. She described the opportunity and the difficulty of teaching a model with a relatively small active component through extensive agent practice.

What we realized is that there's a lot of potential with a small model like this if we train reinforcement learning on it with a large amount of environments and agents. But it's not a very easy thing to do,

Song added that the team spent substantial time refining the process. To keep training stable, MiniMax used CISPO, short for Clipping Importance Sampling Policy Optimization. The mathematical method limits excessive changes during training iterations so that the model does not overcorrect as it learns.

MiniMax calls one resulting behavior an "Architect Mindset." Before generating code, M2.5 plans the project's structure, features and interface design, then divides the work into manageable components. The aim is to organize a solution before executing it.

Documents, spreadsheets and work inside MiniMax

MiniMax worked with senior professionals in finance, law and the social sciences to train M2.5 on knowledge specific to those fields. The objective was to produce finished work products, including the practical judgment professionals use, rather than leave users with unformatted blocks of text.

M2.5 can create and edit Microsoft Word documents, build PowerPoint presentations with layouts and charts, and perform complex financial modeling in Excel. These tasks move the model into familiar office workflows where the result must be usable in a particular application.

On MEWC, a benchmark drawn from actual Microsoft Excel World Championship competition problems, M2.5 scored 74.4%. On MiniMax's internal GDPval-MM evaluation, it achieved an average win rate of 59% against mainstream competitors. GDPval-MM compares outputs in pairs and assesses both the quality of the deliverable and the professional process used to reach it.

MiniMax says M2.5 already autonomously completes 30% of all tasks inside the company. Its deployment spans research and development, product management, sales, human resources and finance. In coding work, the company says the model generates 80% of newly committed code at its headquarters.

Those figures describe different parts of the company's operation: the 30% figure covers tasks across departments, while the 80% figure covers new code committed in its coding environments. Both are company statements about its own use of the model.

MiniMax Agent packages the capabilities into a platform where users have created more than 10,000 specialized Experts. These configurations combine Office Skills with knowledge specific to an industry. An Expert can follow standard operating procedures for industry research or apply particular risk control rules in financial modeling.

M2.5 also supports software development for Web, Android, iOS and Windows. MiniMax describes work covering server APIs, business logic and databases, as well as the visible interface. Its intended role spans system design, feature changes, code review and testing across the software lifecycle.

What the two API prices actually buy

MiniMax offers two API variants with the same capabilities but different generation speeds. An API is the connection through which another application sends work to a model and receives its output.

• Standard M2.5 runs at 50 tokens per second and costs $0.15 per million input tokens and $1.20 per million output tokens.
• M2.5-Lightning runs at 100 tokens per second and costs $0.30 per million input tokens and $2.40 per million output tokens.

Input tokens are the instructions, documents and other text sent to the model. Output tokens are the text it generates. Both variants support caching, which can reduce costs when the same material is used repeatedly.

At 100 tokens per second, an hour of continuous generation produces 360,000 output tokens. At Lightning's output price, that costs about $0.86, before adding input charges. This calculation explains the company's headline figure of roughly $1 an hour.

At the standard variant's 50 tokens per second, continuous generation produces 180,000 output tokens an hour. Its lower output rate and lower token price help explain MiniMax's advertised operating cost of about $0.30 an hour once input costs are included.

The estimate of approximately $10,000 to run four agents continuously for a year follows the cheaper hourly proposition. Four agents running at $0.30 an hour for 8,760 hours would cost $10,512. Actual spending depends on how much input and output each workflow uses, as well as the benefit from caching.

The pricing comparison accompanying the release puts Claude Opus 4.6 at roughly $75 per million output tokens and describes M2.5's output costs as one tenth to one twentieth those of Opus, Gemini 3 Pro and GPT-5. GPT-5.2 also features in the broader comparison of M2.5 with expensive Western systems.

Alex Volkov, host of the ThursdAI podcast, gave another measure: approximately $0.15 in tokens for M2.5 to complete a SWE-Bench task, compared with about $3.00 for Claude Opus 4.6. Those figures amount to a 95% cost reduction per task.

Lower spending could give developers more room to supply extensive context and allow longer reasoning, reducing pressure to shorten prompts and limit token use. It also makes a capable model financially practical for routine work that might otherwise go to a cheaper, less capable system.

Faster completion changes the operating calculation

M2.5 is described as completing tasks from start to finish 37% faster than its predecessor, M2.1. On SWE-Bench Verified, its runtime fell from 31.3 minutes to 22.8 minutes, almost matching Claude Opus 4.6's 22.9 minutes.

MiniMax also puts M2.5's cost per task at 10% of Opus's in that comparison. A business evaluating agents therefore has two separate savings to examine: the amount paid for the work and the time needed to receive a result.

The speed improvement matters for workflows in which one model hands work to another. Faster completion can reduce waiting between steps in an agent pipeline, allowing connected processes to operate closer to real time.

Automated code audits are another potential use. Lower prices allow organizations to examine more code, while deployment on their own infrastructure gives them control over where private data is processed. Hosting an open weights model also brings a separate infrastructure decision, since the published API prices describe MiniMax's service.

No date was given for a further release or a next formal evaluation. For technical leaders considering M2.5 as of February 14, the immediate choice is whether its price, speed and performance suit the actual workload, particularly where recent coding tests showed lower success rates than competing systems.

A crowded week for Chinese AI

M2.5 arrived alongside several other Chinese model launches. Zhipu AI released GLM-5, which surpassed Google's Gemini 3 Pro on the Artificial Analysis Intelligence Index. Alibaba introduced RynnBrain for robotics, ByteDance released Seedance 2.0 for video generation, and Kuaishou debuted Kling 3.0.

MiniMax and Zhipu AI both recorded stock gains of more than 10%, with Zhipu rising 30% in Hong Kong trading. The concentration of announcements placed fresh attention on Chinese companies competing across several branches of AI.

The sequence around M2.5's launch helps explain the comparisons being made:

• January 2026: the SWE-rebench evaluation used 48 fresh GitHub pull requests, producing a tougher result for M2.5 than its SWE-Bench Verified score.
• One week before M2.5's release: Anthropic released Claude Opus 4.6, the model used in several of MiniMax's performance and price comparisons.
• The week of M2.5's announcement: Zhipu AI, Alibaba, ByteDance and Kuaishou also introduced new systems.
• By February 14, 2026: M2.5's release, published pricing and market reaction had sharpened the debate over the cost of deploying capable AI.

Google DeepMind chief Demis Hassabis had previously estimated that Chinese models trailed Western offerings by only months. M2.5's near parity with Opus 4.6 on SWE-Bench Verified, so soon after Anthropic's release, suggested a gap measured in days on that particular comparison. The fresh pull request results placed Western models ahead, while the competitive gap was described as narrowing.

MiniMax's broader business argument is that low prices can change how AI is purchased. The company describes intelligence as "too cheap to meter," pointing toward a market built around large volumes of use rather than expensive access to scarce capability.

Western AI laboratories charging premium API rates face pressure when a competitor offers strong results for substantially less money. M2.5's launch brings that pressure into a practical calculation: how many tasks an organization can afford to automate, how quickly those tasks finish and how often the results succeed.

The Essentials

• M2.5 scored 80.2% on SWE-Bench Verified, 0.6 percentage points behind Claude Opus 4.6, while scoring 39.6% on the fresh SWE-rebench tasks.
• Standard API pricing is $0.15 per million input tokens and $1.20 per million output tokens at 50 tokens per second.
• Lightning pricing is $0.30 per million input tokens and $2.40 per million output tokens at 100 tokens per second.
• The model contains 230 billion parameters and activates 10 billion for a given task.
• MiniMax trained it using Forge and CISPO across more than 200,000 environments and more than 10 programming languages.
• Its intended work includes software development, browsing, Word documents, PowerPoint presentations and Excel financial modeling.
• MiniMax says it completes 30% of internal tasks and generates 80% of newly committed code.
• The Hugging Face release uses a modified MIT License requiring prominent commercial attribution to MiniMax M2.5.
• MiniMax shares rose between 11% and 16% after the announcement and closed at HK$680.