There is a coding model near the top of the SWE-Bench Pro leaderboard right now that most Western founders could not name. It comes from Kuaishou — the short-video company — and on July 10 it quietly slotted into second place on the hardest public repository-level coding benchmark, one rung below Opus 4.8 and above both GLM-5.2 and GPT-5.5. It costs less than either of the open-weight Chinese models it beats.
That combination is the whole story, so here it is up front: KAT-Coder-Pro V2.5 scores 65.2 on SWE-Bench Pro versus Opus 4.8's 69.2, and it rents for roughly $0.74 per million input tokens and $2.96 per million output. If you route coding work by cost, that line just changed your routing table.
What it is#
KAT-Coder-Pro V2.5 is the flagship of Kwaipilot, the KwaiKAT team inside Kuaishou. It's a Mixture-of-Experts model with 72B active parameters and a 256K-token context window, and it was trained with large-scale agentic reinforcement learning inside reconstructed, executable repository environments — not just next-token prediction on code, but issue localization, edits, and test runs graded end to end. The technical report went up as arXiv 2607.05471, and the model landed on third-party provider listings the same week.
It is built to be driven, not prompted once. The design target is the agent loop: read the issue, find the file, make the change, run the tests, iterate. That shows up in the numbers.
The numbers that matter#
Three benchmarks, read in order of how much they should move you:
- SWE-Bench Pro: 65.2. This is the harder, repository-level cut, and it's where the surprise lives. Opus 4.8 leads at 69.2; KAT-Coder-Pro V2.5 is next; GLM-5.2 sits at 62.1 and GPT-5.5 at 58.6. A Kuaishou model is beating a frontier OpenAI model on the benchmark that best approximates real codebase work.
- SWE-Bench Verified: 73.4%. The friendlier, more-quoted benchmark. Here it's in the same band as the frontier closed models rather than ahead of them — useful as a sanity check, less as a differentiator.
- PinchBench agentic tool use: 94.9. KwaiKAT reports this as the best tool-use result among the models it tested. If your agent lives or dies on clean tool calls, this is the number to weigh.
The usual caveat applies: these are largely vendor-run harnesses, and harness choices move SWE-Bench scores by several points. But the shape — strong repository-level agentic coding, strong tool use — is consistent across the report and the independent provider listings, and it's the shape that matters for a founder shipping features, not winning a leaderboard.
A Kuaishou model is now the cheapest way to get within four SWE-Bench-Pro points of Opus — if you're willing to rent the weights instead of own them.
The price, and the catch#
At about $0.74 / $2.96 per million tokens, KAT-Coder-Pro V2.5 undercuts the two open-weight Chinese coders it's usually shelved next to: GLM-5.2 (~$1.40 / $4.40) and Kimi K2.7 Code (~$0.95 / $4.00). Against frontier closed models it isn't close — it's a fraction of the cost. For a solo founder burning tokens on an agentic coding loop all day, that delta compounds fast.
The catch is not the quality. The catch is the ownership. Pro V2.5 is closed-weights. You reach it through third-party providers — OpenRouter, Atlas Cloud, ZenMux — over an OpenAI-compatible API. You are renting, not hosting. If your constraint is data residency, air-gapping, or "the weights must live on my box," this model is off the table, and you should be looking at GLM-5.2 (MIT) or Kwaipilot's own open release, KAT-Dev-32B (Apache-2.0, 62.4% SWE-Bench Verified) — a different, smaller model that you can run yourself.
The routing call#
For the reader this site is built for — a founder or small team choosing where to send agentic coding work — the decision is narrow and clear:
- Route by cost, comfortable with a hosted API? KAT-Coder-Pro V2.5 is now the strongest price-to-SWE-Bench-Pro pick that isn't a frontier lab. Put it in the table for the bulk of your coding loop and reserve Opus 4.8 for the tasks that actually stall.
- Need the weights on your own hardware? It's not your model. GLM-5.2 gives you MIT and a million-token context; KAT-Dev-32B gives you Apache-2.0 in a size you can serve on one box.
- Multimodal loop (reading screenshots, verifying UI)? Neither KAT variant is built for that — that's still MiniMax M3 territory, which we covered in the open-weight coder routing breakdown.
The broader signal is the one worth sitting with. Three months ago the cheap-coding-model conversation was GLM versus Kimi versus DeepSeek. A short-video company just walked in with a model that beats all of them on the benchmark that counts, priced under all of them, and pointed it straight at Claude Code and Cline. The frontier is still Opus. The value frontier moved.
