$10,000 = 4 Agents Year-round, 24/7 MiniMax M2.5 Cost Reconstruction Model Is the critical point for AI scaling implementation here?

Spring Festival AI (Artificial Intelligence) Battle, MiniMax (HK00100, stock price 680.0 HKD, market cap 213.272 billion HKD) was not absent.

On February 12, MiniMax officially launched its latest flagship programming model, MiniMax M2.5. It is reported that, as the world’s first production-level model natively designed for Agent (intelligent agent) scenarios, its programming and agent performance (Coding & Agentic) rival top international models, directly comparable to Claude Opus 4.6, supporting full-stack programming development across PC, app, and cross-platform applications.

Influenced by this news, by the close of Hong Kong stocks on February 13, MiniMax’s stock price rose by 15.65%, with a total market value of 213.272 billion HKD.

Also noteworthy is that the M2.5-lightning version supports output speeds of over 100 TPS (transactions per second), roughly twice that of mainstream models; input cost is about $0.3 per million tokens, output cost about $2.4 per million tokens.

Calculating at 100 tokens per second output, the cost for continuous operation for one hour is about $1; at 50 tokens, about $0.3. This means that theoretically, $10,000 can support four agents working continuously for a year.

Before the Spring Festival, different AI companies had already shown divergent betting directions. Some bet on multimodal models, some rush into C-end (consumer side) entry points, but MiniMax has all its chips on “performance and cost breakthroughs” in the agent scenario. Is this a prelude to price wars, or a new path to accelerate AI commercialization?

Behind the Launch of the Text Model: Is MiniMax Trying to Reconstruct the Agent Economy?

Regarding the performance of MiniMax M2.5, Tian Feng, director of the Fast & Slow Thinking Institute, told the Daily Economic News that from M2.5’s positioning, it is an “native agent production-level model,” with core value in providing reliable task text understanding and long-term reasoning ability for complex problems in agent scenarios.

“In core capabilities such as programming, tool invocation, and complex task decomposition, M2.5 has already reached the global SOTA (State of the Art) level. These capabilities are the foundation for building efficient agents,” Tian Feng said.

All signs indicate that MiniMax’s layout for agents is already clear.

In the past 108 days, MiniMax has iterated from M2 to M2.1 to M2.5, with scores on the SWE-Bench Verified (software engineering benchmark test) rising from 69.4 to 80.2. The team attributes this leap to large-scale agent reinforcement learning (RL scaling).

It is understood that their self-developed Forge framework decouples the training engine from the agent, enabling generalized optimization for any agent scaffolding and tools, and accelerates training by about 40 times through asynchronous scheduling and tree-based merging strategies.

At the algorithm level, they adopt CISPO optimization and process reward mechanisms to alleviate credit assignment issues in long-context scenarios, incorporating “real task duration” into the reward function to balance effectiveness and response speed.

On February 12, M2.5 was launched on MiniMax Agent, and on the 13th, it was open-sourced globally to support local deployment. In less than a day, users worldwide have built over 10,000 experts on MiniMax Agent, with numbers still rapidly growing.

MiniMax states that it hopes to continuously improve model capabilities while building a sustainable, scalable agent ecosystem—Agent Universe.

It is worth noting that at this stage, AI companies’ bets are more focused on multimodal large models. Why does MiniMax choose to launch a text large model now?

Tian Feng explained that MiniMax has concentrated almost all resources on the continuous enhancement of foundational model capabilities. “The launch of M2.5 is a continuation of this strategy—first establish a strong base model, then radiate out to specific application scenarios.”

He also mentioned that MiniMax is one of the earliest domestic companies to adopt a full multimodal model technology route. Launching a pure text model does not mean abandoning multimodality; rather, it is an optimization tailored for agent scenarios based on existing multimodal capabilities.

iMedia Research CEO Zhang Yi told reporters that MiniMax’s push for low-cost large models reflects a clear differentiation strategy: avoiding the red ocean of multimodal competition and directly addressing the core pain points of high cost and low efficiency in agent deployment.

Beyond performance improvements, a major focus on M2.5 is cost control. MiniMax believes that when performance and cost are no longer constraints, the economic model for large-scale agent deployment will fundamentally change.

Wang Peng, associate researcher at Beijing Academy of Social Sciences, said that MiniMax and other vendors lowering agent usage costs to very low levels mark the transition of AI commercialization from “technology validation” to “scale replacement.”

He pointed out that in the past, high inference costs limited agent applications to high-value tasks, but now low costs enable enterprises to deploy AI in bulk for routine repetitive work (such as customer service, data entry), and even create new business models (like pay-per-result AI services).

Will the Industry Shift Toward Price Competition? Experts: More Likely to Trigger a “Value War”

Notably, before the Spring Festival, many AI companies had already made agent-related moves.

In products, various players are racing to seize opportunities. On February 11, Meituan LongCat released a native “deep research” agent. User blind tests showed an overall usability rate of 61.1%, better than ChatGPT’s 42.8%. This feature is now freely available on LongCat’s web platform.

On January 20, MiniMax released Agent 2.0, positioning it as an “AI native workbench,” supporting desktop versions on Mac and Windows, and launching specialized “Expert Agents” for vertical scenarios. On January 19, Ziejue Xingchen officially announced the upgraded desktop agent product “Ziejue AI Desktop Partner,” with a Windows version available for free.

In terms of models, on the evening of February 3, Alibaba open-sourced the new generation intelligent agent programming model Qwen3-Coder-Next, which, with only 3B parameters activated, already rivals top open-source models like DeepSeek-V3.2 and GLM-4.7 in agent programming performance.

It is understood that based on technological breakthroughs, Qwen3-Coder-Next significantly reduces reasoning costs, only 5% to 10% of similar-performance models, making it especially suitable for low-cost agent deployment on home computers and lightweight servers. It is currently the most capable small open-source programming model for agent programming.

Alibaba stated that facing real-world challenges such as long-context reasoning, tool use, and recovery from execution failures, the new model Qwen3-Coder-Next can handle them with ease.

It is evident that reducing costs is the current main direction for large models related to agents. Does this mean that with M2.5’s entry, the industry might move toward price competition?

Zhang Yi believes, “This is not necessarily a money-burning race for market share.” He explained that MiniMax’s low prices are mainly achieved through technical optimization, not through subsidies or reckless spending. “Whether it triggers a price war depends on future market developments. But what’s certain is that it will accelerate the淘汰 of inefficient competitors and shift the industry toward performance and cost dual competition.”

Tian Feng also said that M2.5’s low cost is an inevitable result of architectural optimization and engineering improvements, which will push agents from “proof of concept” toward “large-scale commercial use,” more likely sparking a “value war” rather than a traditional “price war.”

He predicted, “We expect that reducing inference costs by tenfold annually in the large model industry is a likely ongoing trend. The emergence of M2.5 will accelerate this trend, driving the entire industry toward higher efficiency and lower costs.”

He also mentioned that previously, agent products were generally priced high, mainly targeting enterprise clients. The cost advantages brought by M2.5 could enable agents to reach a broader market, including small and medium-sized enterprises, developers, and even individual users. “This could expand the overall agent market size, rather than just lead to price competition.”

Wang Peng believes that the explosion of agents and falling costs mark the transition of AI from a “technological singularity” to a “product singularity.” “Just as smartphones replaced feature phones, future AI competition will no longer be about parameter size, but about how well it integrates into workflows and creates real value.”

Spring Festival AI Battle in Full Swing, Is the Critical Point for Large-Scale AI Deployment Here?

Regarding the reason for the collective bet on agents across the industry, Wang Peng believes it fundamentally reflects a paradigm shift from “passive response” to “active execution.”

“Traditional large models are like ‘knowledge bases,’ while agents are more like ‘digital employees’—able to decompose tasks, invoke tools, handle exceptions, and even self-optimize. This shift stems from enterprise demand upgrades: users no longer just want information, but expect AI to complete work in a closed loop (such as automatic order processing, financial report generation),” Wang Peng said.

With new models and products launching collectively, have companies already begun to widen the gap in the agent track?

Tian Feng believes that the gap in the large model agent track is indeed widening, but more in terms of engineering capability, scenario implementation, and cost-effectiveness rather than just parameter scale or basic ability.

It’s worth noting that the current deployment of agents by various companies is driven by an urgent need for AI commercialization.

Tian Feng pointed out that M2.5 is explicitly positioned as an “native agent production-level model,” with all core capabilities focused on programming, tool invocation, and high-value economic tasks like office productivity. This specialized path reflects MiniMax’s deep understanding of commercialization.

It’s also notable that during this Spring Festival, AI has become a key focus for major tech companies. Despite different strategic focuses, the industry shows a gradually converging direction.

Tian Feng mentioned that the industry is shifting from “parameter competition” to “revenue” and “profit” competition. Pure technological lead is no longer enough to win the market; those who can quickly convert technological advantages into measurable commercial revenue will ultimately succeed.

He also noted that whether it’s ByteDance’s “traffic + scenario” model, Alibaba’s “e-commerce platform + ecosystem,” or MiniMax’s “specialization + deployment,” fundamentally they are building different ecosystem barriers.

Zhang Yi added that the AI battle during the Spring Festival shows a move from homogeneous “involution” to differentiation. “Manufacturers are clearly diverging in their strategic directions, mainly focusing on multimodal, agent efficiency, and ecosystem implementation, driven by differences in technological endowments, commercialization stages, and scenario demands.”

He sees this as a transition from a parameter-scale “arms race” to a more scenario-based, practical, and segmented new phase of competition.

From “showcasing skills” to “commercial use,” from “trying out” to “popularization,” this Spring Festival, the AI track was lively with many players. As the cost barriers of AI gradually lower and more users engage with AI products, this intensive competition during the holiday may be seen in the future as a historic turning point—AI moving from “festival fireworks” to “everyday lighting.”