Author | Yu Jing
Editor | Ma Xiaoning, Liang Bingjian
"M3 points its sword at Claude Code. "
Can an open-source model simultaneously possess top-tier coding capabilities, ultra-long context understanding, and native multimodal abilities?
This is almost the entire implication of an Agent. We raise this question because, since the era of OpenClaw, a company can no longer prove it is betting on the future solely through its investments in models. The victory lies entirely in Agents.
MiniMax M3 seems to have realized this as well.
As MiniMax's latest flagship model, M3 focuses on enhancing Coding and Agent capabilities. Compared to the traditional code model approach of "just writing the code," it places greater emphasis on long-term planning, multi-turn collaboration, and the ability to autonomously execute complex tasks.
In simple terms, these capabilities collectively point to a single goal: enabling the model to independently study hundreds of thousands of words of materials, work continuously for hours, call tools, write code, and ultimately deliver a truly usable result. This has become the core technical foundation for the simultaneously launched MiniMax Code product.
The derived question is, now that Claude Code has become one of the most recognized Agent tools among developers, are M3's capabilities sufficient to support MiniMax in building its own, truly competitive Agent ecosystem?
01. 12 Hours of Autonomous Work: How Long is Your "Long Task"?
The evolution of coding capabilities is no longer just about writing code.
If MiniMax M3 is merely regarded as a model that is better at writing code, the focus of this release will be severely underestimated. What is more worth discussing about M3 is its capabilities in long tasks, long context, and Agentic workflows.
Two cases provided officially illustrate this well. One is M3 autonomously reproducing an ICLR paper in nearly 12 hours, and the other is completing CUDA Kernel optimization in about 24 hours and 147 iterations. Essentially, both examples are typical long-chain tasks where the model needs to understand the goal, break down the steps, continuously check intermediate results, and keep adjusting after failures.
From the perspective of model architecture, MiniMax M3's 1M token context and MSA sparse attention architecture are designed to serve such scenarios. The significance of long context is not just about fitting in more text; more importantly, it reduces information fragmentation in long tasks. For instance, a real code repository, a complex requirements document, or a set of historical modification records—these real-world requirements cannot be clearly explained in just a few thousand tokens. If the model can only see the local part each time, it is easy to encounter situations where "the first part is answered correctly, but the later part is messed up." A longer context window gives the model the possibility to understand tasks across files and stages.
02. Startup Simulator: A Direct Competition Between M3 and Sonnet 4.6
To test M3's code delivery capabilities, I designed a relatively complete small project, asking the model to implement a "Startup Simulator" mini-game from scratch. Claude Sonnet 4.6 also underwent this test.
| Please develop an AI Startup Simulator Web App from scratch.
Requirements: 1. Users can create a startup, inputting the company name, industry, initial funding, and target users. 2. The game is turn-based, with each round representing one month. 3. In each round, users can choose 3 business decisions, such as product development, marketing, recruitment, financing, cost reduction, and user research. 4. The AI generates a monthly report based on the current company status and user decisions. 5. The page needs to display funding, user count, revenue, team morale, product completion rate, market heat, and competitive pressure. 6. Update these metrics at the end of each round. 7. There must be success and failure endings. 8. Implement using React + Tailwind, and the interface should look like a modern startup management game. 9. The AI interface can use mock data first, but the code structure should be convenient for integrating real LLM APIs. 10. Please ensure the project can run and provide the startup method. |
The prompt is not complex, but this task is actually very suitable for testing the comprehensive capabilities of a Coding Agent. Because it simultaneously tests requirement understanding, state management, UI design, numerical systems, and balance. Users play the role of entrepreneurs in the game, needing to decide each round what product to make, who to hire, how to price, whether to raise funds, and how to market. The AI then provides feedback on user growth, cash flow, team morale, market reaction, and competitive pressure based on these decisions.
Specifically, the real difficulties mainly include three dimensions:
State management: Once the mini-game enters multi-round decision-making, it is easy to encounter issues such as data loss after page refresh, previous round data overwriting the next round, inability to review historical records, and progress bars exceeding 100%. In fact, games are just high-frequency scenarios for these problems; similar requirements can be seen in many software development tasks.
UI presentation: The "games" generated by many models are actually just a form plus a few buttons. The functions work, but at first glance, they have a "plastic feel."
Numerical balance: This is the most difficult part. Improper numerical design can easily lead to cash flow explosion, abnormal user growth, and the game quickly getting out of control after just one or two rounds, ultimately affecting playability. What kind of numerical design can be considered balanced? This requires the model to have an additional layer of aesthetic and taste for games, beyond complex task decomposition.
M3 completed the program writing and code checking in about 11 minutes. The final generated mini-game runs normally, has a clean interface, and includes certain animation effects. More importantly, it basically handled the core difficulties mentioned earlier well: company data is displayed clearly, historical records can be reviewed, and there is no obvious confusion in game progress and business metrics.
In contrast, Sonnet 4.6 took about 19 minutes to complete the same task. It also got the game running normally and added a little cleverness in content design. For example, it added random events to make the game difficulty and uncertainty stronger, making it indeed more playable.
This is a very interesting difference.
MiniMax Code, based on M3, is more like a highly executable engineer Agent. It will deliver very faithfully around your prompt. The advantage lies here: it acts fast, has high completion rates, and once the instructions are given, it will build the pages, logic, state, and basic interactions together around the final product.
On the other hand, Claude Code, based on Sonnet 4.6, is more like a collaborator who proactively supplements product ideas. It might add some extra designs beyond the requirements.
There is no absolute good or bad between these two styles. If your requirements are very clear and you want the model to strictly follow instructions and complete the task quickly, M3's performance will be very comfortable. After all, who wouldn't want an employee who executes exactly where pointed. But if you expect the model to proactively complete product creativity, enhance gameplay, and propose more possibilities, Sonnet 4.6 currently still has the advantage in creative expansion.
03. Front-End from Images: Hands-on Review of Native Multimodal Capabilities
Compared to long tasks and coding capabilities, multimodality might be the most underestimated capability of MiniMax M3.
Many models promote that they support image input, but in actual experience, they often stay at the stage of "describing what they see," able to describe what elements are on the page, but struggling to further translate this visual information into runnable code. The biggest surprise M3 gave me this time is precisely that it demonstrated a complete chain of capabilities from visual understanding to engineering delivery.
To test this, I chose a very direct scenario, using MiniMax's own official website homepage as the test subject. I provided M3 with two screenshots of the homepage and asked it to replicate the page using React and Tailwind CSS.
| Please fully replicate the page based on this web page screenshot using React + Tailwind CSS.
Requirements: 1. Restore the original page as much as possible: Overall layout Font hierarchy Card design Color scheme Spacing and whitespace Button styles 2. The page must be responsive, adapting to: Desktop Tablet Mobile 3. Identify and restore: Hero Section Navigation bar Feature Cards CTA Button Banner Footer 4. Use a componentized structure: Navbar.tsx Hero.tsx FeatureCard.tsx Footer.tsx 5. Do not use placeholder code. 6. Output complete runnable code. Make the visual similarity between the generated page and the screenshot reach over 90%. |
The reason for choosing the official website homepage is that such marketing pages often contain a large number of visual design details: navigation bars, card modules, gradient backgrounds, button styles, information hierarchy, and complex page layouts. For the model, this is not just about recognizing the text in the image, but also understanding the design logic behind the entire page.
The final result was somewhat unexpected.
First is the restoration of the page structure.
With just two screenshots, M3's replication of the overall layout of the homepage has reached an extremely high level. The navigation bar, Hero area, feature introduction modules, and the hierarchical relationships between various content sections were all accurately identified, and the overall page structure is almost consistent with the original web page.
Looking only at the macro layout, it has almost reached the point of being indistinguishable from the real thing. The remaining differences are mainly concentrated at the detail level, such as some font spacing and element alignment. But even if you take separate screenshots of these different local parts, you would have to recall whether the genuine MiniMax official website actually looks like this.
What is even more interesting is that M3 did not mechanically "copy the screenshot."
Since I only provided part of the homepage content during the test, theoretically, the model could not know how the lower part of the page should be designed. However, in the actual generation process, M3 did not simply leave it blank. Instead, it proactively analyzed the overall visual style and color characteristics of the official website and independently supplemented several content modules with a consistent style for the subsequent pages. Although these contents do not completely correspond to the real official website, both the color scheme and design language maintain a high degree of consistency with the original page, and the overall look does not give people a sense of obvious fragmentation.
This point is actually very important. Because it shows that the model is not just doing OCR or screenshot replication, but is trying to understand the design rules behind the page and use this understanding to make reasonable inferences.
In addition to the visual restoration, M3's recognition of interactive elements also left a good impression on me.
In the generated results, the model correctly identified interactive components such as the navigation bar and buttons, and endowed these elements with actual functions. For example, menu items in the navigation bar can directly jump to the corresponding content areas, and button components are correctly implemented as clickable elements.
Going a step further, M3 also proactively added interactive animations to the page. When the mouse hovers over a button, transition animations and visual feedback will appear on the page. These effects were not in my prompt, but are details added by the model itself based on the design habits of modern web products.
In summary, M3 has demonstrated quite strong competitiveness. It can not only understand the structural information in web page screenshots, but also identify interaction logic, infer missing content, and ultimately generate a front-end page that is runnable, interactive, and highly consistent in visual style.
Of course, it is not without shortcomings. There are still some deviations in layout details on the page, but considering that the entire process was almost completely completed autonomously by the model, and the input was merely two screenshots, such a result has far exceeded initial expectations.
04. Price is Also Productivity
Price is the most realistic topic in large model competition. Over the past year, the AI industry has almost experienced a comprehensive price war. DeepSeek overturned the market with extremely low API costs, while OpenAI, Anthropic, and Google have been continuously adjusting their pricing strategies while steadily improving model capabilities.
From the official positioning, M3 focuses on Frontier Coding, Agent, multimodal, and million-level context capabilities. This first determines that its competitors are not those mid-tier models used for intelligent customer service and meeting minutes, but a batch of the most cutting-edge flagship models in the current industry, such as Claude Opus 4.8, GPT-5.5, Gemini 2.5 Pro, GLM-5.2, and DeepSeek V4-Pro.
Looking directly at the price, the current API price for Claude Opus 4.8 is $5/million tokens for input and $25/million tokens for output. GPT-5.5 is $5 for input and $30 for output. After the latest price cut, DeepSeek V4-Pro is $0.435 for input and $0.87 for output. In comparison, the official price for MiniMax M3 is $0.6 for input and $2.4 for output.
If Claude Opus 4.8 is used as the baseline, M3's input cost is only about 12% of it, and the output cost is less than 10%. Even when facing OpenAI's latest GPT-5.5, M3's invocation cost is only about one-tenth of it. In other words, under the same consumption of 1 million input tokens and 1 million output tokens, using GPT-5.5 requires $35, using Claude Opus 4.8 requires $30, while M3 only requires $3.
For ordinary users with low usage, this difference is not yet obvious. But if you are a developer who is already used to relying on a large number of Agents every day to process long documents, batch generate code, or build AI applications, the cost gap will be rapidly amplified. Assuming a project consumes 10 million input tokens and 10 million output tokens per month, the cost of using Claude Opus 4.7 is about $300, while using M3 only requires about $30. Under the premise of maintaining capabilities close to flagship models, a cost advantage of about ten times is already sufficient to influence technology selection.
Of course, price can never be discussed in isolation from capabilities.
If M3 only has cheapness to commend it, then such a comparison would be meaningless. But interestingly, in multiple evaluations published by MiniMax, M3 is targeting the competitive zone where Claude Opus 4.7, GPT-5.5, and Gemini flagship models are located. In code capability tests such as SWE-Bench Pro, M3 has surpassed GPT-5.5 and Gemini flagship models, approaching Claude Opus 4.7. In long-task Agent scenarios, the officially displayed cases of paper reproduction and CUDA optimization can even run autonomously for more than a dozen hours continuously.
More importantly, M3 does not rely on castrating capabilities to exchange for low prices. As mentioned earlier, behind this price is a 1 million token context window, native multimodal architecture, and Agent workflow capabilities, and Claude, GPT, and Gemini are evolving in exactly the same direction. In other words, the market M3 is trying to compete in is not the "cheap model market," but the most expensive and technologically highest flagship model market.
Therefore, if looking only at absolute prices, DeepSeek V4-Pro is still the most aggressive price killer currently. But if simultaneously considering flagship capabilities such as Coding, Agent, multimodality, and ultra-long context, then M3 might be one of the most impactful cost-performance players in the entire market currently.
05. The Ambition of MiniMax Code
After several tests, my feeling about MiniMax M3 is that it can already be ranked as one of the most noteworthy Coding/Agentic foundation models among domestic models, especially demonstrating strong competitiveness in long tasks, long context, multimodal input, and code delivery.
Pay attention to this positioning of a foundation model. Previously, the constant changing of the guard in various Benchmarks once made performance level the only angle to measure large models. But when we discuss Agents and discuss implementation, a more realistic dimension is usability. MiniMax M3 seems to have no intention of challenging the status of the "strongest model" again, but is trying to become the most cost-effective infrastructure in the Agent era.
This is a clearer path. As Agent products such as Claude Code and OpenAI’s code agent products gradually become the main entry points for developers, models are increasingly returning to their original role, which is a foundational capability. For developers, whether a model can complete long tasks, call tools, understand images, and control costs is far more important than simply running scores.
From this perspective, MiniMax's strategy is quite clear. M3 has steadily entered the first tier in long context, multimodal, and coding capabilities, while bringing these capabilities into more real workflows at a price far lower than GPT and Claude.
Real workflows, which is also the reason why MiniMax Code is launched at this moment.
The business model of selling tokens is short-lived. The fact is that developers vote with their feet without hesitation, and API suppliers are increasingly easy to replace. In this situation, a 3-month lead in model capabilities does not mean a 3-month commercial advantage. This forces model manufacturers to ask: why hand over the most bargaining-powerful entry points to others?
In addition, today's production-grade Agents have become a highly integrated system engineering capability. The delivery level of a complex task depends only partially on the model, and partially on the Agent Runtime. If the right evaluation angle is found, every company has a "SOTA" model, then the battle for the execution layer will become one of the new focus points of competition.
So is MiniMax Code a software for writing code?
This is still the perspective of an IDE. It is actually the model, codebase context, tool invocation, execution environment, and workflow orchestration—everything that determines the final effect of the Agent is here. With these, MiniMax has the qualification to compete for the entry point of developers' workflows.