MiniMax is using its new M2.5 release to make a direct argument about the next phase of artificial intelligence: frontier AI should become cheaper to run, not just more capable. The Shanghai company has released the open-weights model under the MIT license and is presenting it as a fast, low-cost option for coding, search, agent workflows and office tasks.
The company’s message is clear. If its performance and pricing claims hold up in real-world use, M2.5 could add another layer of pressure on Western AI labs already competing on model quality, developer adoption and operating cost.
What MiniMax released
MiniMax M2.5 is an open-weights model from Chinese AI company MiniMax out of Shanghai. It is available under the MIT license, which makes the release notable for developers and companies that want more control than they typically get from closed model access.
MiniMax says the model was trained with reinforcement learning across hundreds of thousands of complex environments. According to the company, that training approach is meant to support strong results while keeping token use low, especially during long-running tasks.
The model is also described as more proactive in how it plans. MiniMax says M2.5 learned to optimize its actions through its own planning, rather than only planning when directly prompted. That distinction matters for AI agents, where a model may need to break down a task, choose steps, use tools and continue working over time.
MiniMax also says M2.5 can handle Word, Excel and PowerPoint files quickly and smoothly. That places the model squarely in a market where everyday productivity work is becoming a major test of practical AI usefulness.
Where the benchmark claims are strongest
The published benchmark claims put M2.5 in a competitive position on programming tasks. MiniMax reports 80.2 percent on SWE-Bench Verified, 79.7 percent on the Droid platform and 76.1 percent on OpenCode. In some cases, the source article says those results edge out Anthropic's Claude Opus 4.6.
The company also claims that M2.5 beats GPT-5.2, Gemini 3 Pro and Claude in several other areas. Those include web search, with BrowseComp at 76.3 percent, AI agent tool use, with BFCL at 76.8 percent, and office tasks such as Excel work.
For readers, the practical takeaway is not just that MiniMax is claiming strong test scores. It is that the company is positioning M2.5 as a general work model: able to write code, search the web, operate tools and work with common office formats.
- Programming: SWE-Bench Verified, Droid platform and OpenCode are central to the company’s benchmark pitch.
- Agent work: BFCL is used to support the claim that M2.5 can use tools effectively.
- Search: BrowseComp is part of the company’s case for web-search strength.
- Office work: Word, Excel and PowerPoint handling are presented as core practical abilities.
The pricing claim is the real challenge
MiniMax is not only selling M2.5 as capable. It is selling it as cheap to operate.
The company ships two variants, including M2.5-Lightning. MiniMax says M2.5-Lightning reaches 100 tokens per second, making it twice as fast as other top models. It also says one hour of continuous operation costs just one dollar.
According to MiniMax, that price makes the model ten to twenty times cheaper than comparable offerings from Anthropic, Google and OpenAI. The company also claims four M2.5 instances could run continuously for a full year at $10,000.
MiniMax summed up its pitch with a direct statement:
"M2.5 is the first frontier model where users do not need to worry about cost, delivering on the promise of intelligence too cheap to meter,"
That phrase is doing a lot of work. It frames cost as the next barrier to adoption. If AI becomes cheaper to run for long periods, developers can build systems that stay active, process more work and support more users without treating every model call as a major expense.
Access and distribution
M2.5 is freely available on Hugging Face and GitHub. A detailed technical report is available at minimax.io, while API access and a dedicated coding plan are available at platform.minimax.io.
That distribution strategy gives MiniMax several routes into the market at once. Open-weights availability can attract researchers and developers who want to inspect or adapt the model. API access can serve users who want hosted infrastructure instead of managing deployments themselves. A dedicated coding plan points directly at one of the model’s strongest claimed benchmark areas.
For an AI model competing with offerings from Anthropic, Google and OpenAI, that mix matters. The battle is not only about the model card or a benchmark table. It is about whether developers can try the model easily, compare it against their current tools and decide whether the economics are compelling enough to switch.
The market picture remains mixed
The M2.5 release lands in a market where Chinese labs are increasingly part of the pricing conversation. The source article notes that pricing like this could put more pressure on Western AI labs, which are growing fast but still losing money.
It also says some US startups are reportedly turning to Chinese models because of lower prices. At the same time, US enterprise AI remains firmly in the hands of Microsoft, Google, OpenAI and Anthropic. Even Chinese labs acknowledge that US labs are still ahead, and chip scarcity is not helping matters.
MiniMax itself is also more than a language-model company. It builds video AI tools as well. In June 2025, it released Hailuo 02, the second generation of its video AI model.
The company also faces legal pressure. Since September 2025, Disney, Universal and Warner Bros. have been suing MiniMax for copyright infringement tied to its models. That did not prevent the company from going public in early January 2026. Its stock surged 109 percent to HK$345 on its first day of trading, and the M2.5 release gave the stock another boost, with shares closing 15.7 percent higher at HK$680.
MiniMax was founded in 2021 by former SenseTime employees. With M2.5, the company is trying to turn low operating cost into a strategic advantage. The unanswered question is how much of that advantage will show up outside benchmarks, once developers test the model on messy, sustained, real-world work.