MiniMax Wants to Make Long-Running AI Agents Economically Boring

The Macro: Open-Source Models Are Done Being the Runner-Up

For the past two years, the conventional wisdom went something like this: open-source models are great for tinkering, fine-tuning on niche datasets, and avoiding API bills. But when you actually need to ship something that works, you reach for Claude or GPT-4. That calculus is getting messier every month.

The frontier has been compressing fast. Meta’s Llama series normalized the idea that open weights could be genuinely useful. Mistral proved a smaller shop could punch above its weight. DeepSeek’s R1 dropped earlier this year and made a lot of people reconsider what a lab outside the Bay Area / London corridor could produce. The market underneath all of this is substantial. Estimates vary wildly. The open-source services market is pegged anywhere from $18B to $39B in 2025 depending on which analyst you ask, which tells you something about how coherent the category definition actually is. The directional consensus is clear regardless: 16–17% CAGR through the next decade, minimum.

What’s actually interesting right now isn’t the raw capability race.

It’s the agentic use case. Long-horizon agents that browse, code, call tools, and loop for minutes or hours are a genuinely different compute problem than a single-shot chat completion. The cost structure matters enormously here. If you’re running an agent that makes 200 tool calls to complete a task, the per-token price of your model isn’t an accounting rounding error. It’s the business model.

That’s the context in which MiniMax-M2.5 lands. Chinese AI lab MiniMax, whose CEO Junjie Yan reportedly delivered a keynote at WAIC 2025, has been building a broad multimodal stack covering text, speech, video, and music. M2.5 is their current flagship push into the open-source frontier tier. The timing isn’t random. Benchmark pressure from Kimi K2 and GLM-5 is real, and the agentic coding niche is crowded enough that differentiation requires actual numbers, not vibes.

The Micro: $1/Hour for an Agent That Can Actually Code

M2.5’s headline benchmark is SWE-Bench Verified at 80.2%. This is a coding evaluation built around resolving real GitHub issues, not toy problems. That reportedly edges out GLM-5’s 77.8% on the same benchmark, according to third-party comparisons. The reference to Claude Sonnet in the product tagline is about the open versus closed divide. M2.5 is claiming it’s the first open model to cross the threshold where you’d previously have just reached for Sonnet.

Beyond coding, the benchmark spread is deliberately broad. BrowseComp at 76.3% for web search and navigation tasks. BFCL at 76.8% for function and tool calling. Plus claims around office work tasks. The BrowseComp number matters specifically for agentic use. If a model can’t reliably navigate and extract information from the web, your agent is going to hallucinate its way into embarrassing errors at step 12 of a 20-step task.

The speed and cost claims are where MiniMax is making a real argument.

The product cites 37% faster on complex tasks and $1 per hour at 100 tokens per second. That 100 tps figure has practical weight. It’s the difference between an agent that wraps up a task in three minutes and one that takes eight. For anything running in a user-facing loop, that’s a UX problem, not just an infrastructure footnote.

The model is available via the MiniMax API and has already shown up on SiliconFlow and NetMind. The distribution play is happening through existing inference marketplaces rather than expecting developers to self-host. It’s on OpenRouter too. That’s a reasonable bet. Most developers building agents aren’t running their own GPU clusters, and meeting them where they already are is the move. It got solid traction on launch day with a respectable showing from builders who actually read benchmark tables.

The Verdict

MiniMax-M2.5 is doing something more specific than “we have a good model.” It’s making an economic argument for long-horizon agentic workloads, and that argument is coherent. The benchmark numbers are verifiable, the pricing is concrete, and the distribution is already live on platforms developers actually use.

What I’d want to know at 30 days: do the benchmark numbers translate to real task completion in messy, real-world codebases. The ones with undocumented dependencies and weird legacy patterns. Or does the gap with Claude Sonnet reappear the moment you leave benchmark-shaped problems? SWE-Bench is good, but it’s still a benchmark.

At 60 days: is the inference infrastructure actually holding at 100 tps under load, or is that a launch-day number that softens when real traffic hits?

At 90 days: does MiniMax’s multimodal stack become a coherent platform story, or does M2.5 float as an interesting model without enough gravity to stick?

My read is that this is genuinely useful for developers building cost-sensitive agentic pipelines who need open weights and can’t justify Sonnet pricing at scale. I’m more skeptical that $1/hour holds as a differentiator once competitors adjust their own pricing, and I’d want to see the real-world codebase performance before betting a production system on the SWE-Bench number alone. The economic argument is smart. Whether it’s durable is a separate question.