Alibaba leads a $300M round in UniPat AI, a model-testing lab it spawned
Two of the morning's stories come out of the same place: the bottleneck on agentic AI is no longer the model, it's knowing whether the model actually did the job.
Alibaba is set to lead a $300 million round in UniPat AI at a $2.5 billion valuation, Bloomberg reported, with Tencent and existing backer Sequoia China also taking part. UniPat was founded in late 2025 by Li Kuan, who worked on post-training analysis, data synthesis, and reinforcement learning at Alibaba's Tongyi lab before leaving, and its business is selling the thing every frontier lab is short of: realistic test scenarios and the high-grade training and evaluation data behind them. The deal is not closed and terms can still move, so treat the numbers as reported, not signed.
What makes it worth watching is the price tag on a company less than a year old. A $2.5 billion valuation for an evaluation-data shop — before a widely known product — is a bet that judging models is becoming its own industry, not a chore labs do in-house on the way to shipping. It also quietly undercuts the idea that eval is a neutral public good: when the same giant funds both the model and the referee, the referee's independence gets harder to assume.
Alibaba's Accio put a number on how bad today's agents are at real commerce work. At Alibaba International's CoCreate 2026 in Los Angeles on September 9, the Accio team open-sourced CommerceAgentBench: 107 end-to-end business tasks distilled from ten million active small-business users, 1.6 million real conversations, and 200,000 execution traces. The tasks are deliberately mundane and hard — dig the real lead out of 300 messy emails, catch an easy-to-miss payment scam, work out landed cost and book a route across three carriers — and each one runs in a fresh container graded on whether the agent changed the state of the system, not on what it said.
The reference results are the story: the strongest agent evaluated completes 61.68% of the tasks, leaving more than 40 failures. Accio argues the fix is routing — match each task to the model that suits it, then post-train on real commercial work — and says the approach cuts token cost 50% versus a general-purpose agent on core cases. Alibaba International president Zhang Kuo framed the split simply: agent equals model times harness times context. Roughly 15,000 US small businesses showed up in Los Angeles, up from 3,000 a year ago, which is the demand side of the same bet.
DeepSeek cut Flash-series API prices by up to 60%, effective midday Beijing time on September 10 — 27 days after raising V4 Pro rates by as much as elevenfold. The steepest cut is on cache-hit input, now RMB 0.02 per million tokens against RMB 1.00 for a cache miss, a 50x spread that makes cache reuse the single biggest lever on an agent's bill. Output pricing, the biggest line on most invoices, did not go back to pre-August levels: RMB 4.00 per million off-peak is still double the old RMB 2.00.
Read together with the day's other two stories, the pattern is coherent. DeepSeek is making repeated, long-running agent work cheap enough to build on while holding margin on premium generation, and the labs buying evaluation infrastructure are doing it because nobody can yet answer the question CommerceAgentBench asks: did the agent finish the job? Cheaper tokens make the attempt affordable; they don't make the answer easier.
If agents fail four in ten real commerce tasks today, would you let one place an order on your behalf? Tell us in the comments.
Sources: CommerceAgentBench (GitHub)
