KAT-Coder-Pro V2.5 Lands Second on SWE-Bench Pro

Kwaipilot, the AI research group inside Kuaishou, shipped KAT-Coder-Pro V2.5 this week, and it scores 65.2 on SWE-Bench Pro: second place, behind only Opus 4.8 at 69.2. That ranking is worth noting, but it is not the interesting part. The interesting part is buried in the technical report, where the team explains that they raised the success rate of automatically building runnable real-world repository environments from 16.5% to 57.2%, and used that to accumulate over 100,000 executable repos to train against. The score is the receipt. The infrastructure is the product.

• KAT-Coder-Pro V2.5 is a Mixture-of-Experts model with 72B active parameters, a 256K context window and up to 80K output tokens, available via API on the StreamLake platform.
• It scores 65.2 on SWE-Bench Pro (second to Opus 4.8's 69.2) and 94.9 on PinchBench, the best agentic tool-use result among models tested, all evaluated under a unified Claude Code harness.
• The cheaper KAT-Coder-Air V2.5 runs $0.15 per million input tokens and $0.60 per million output, roughly a fifth of Pro's $0.74 / $2.96, on the same 256K context.
• The training story is AutoBuilder: reconstructing verifiable repo environments with executable fail-to-pass and pass-to-pass signals, lifting build success from 16.5% to 57.2%.

What did Kwaipilot actually ship?

Two models, both API-only through StreamLake. KAT-Coder-Pro V2.5 is the flagship: a Mixture-of-Experts architecture with 72B active parameters, 256K of context and up to 80K output tokens in a single response. KAT-Coder-Air V2.5 is the same context window and output ceiling at roughly a fifth of the price, aimed at interactive development loops rather than one-shot generation. Both support function calling, structured JSON output and prompt caching, and both slot into existing toolchains including Claude Code and OpenHands.

RelatedCursor Builds 'Sand' Agent to Rival Claude Cowork

The positioning is explicit and unusual for a frontier-adjacent release: Kwaipilot is not claiming the top of the table. The report says KAT-Coder V2.5 ranks second only to Opus 4.8 on repository-level software engineering and takes first on PinchBench for agentic tool use. That combination, near-frontier on the hard benchmark plus best-in-class on tool orchestration, describes a model tuned for long-horizon agent work rather than single-file code completion.

Why does AutoBuilder matter more than the score?

Here is the problem nobody markets. To train a coding agent with reinforcement learning, you need environments where an action can be graded objectively: the test suite passes or it does not. That means you need real repositories that actually build and actually run their tests, at scale, unattended. Anyone who has cloned a random GitHub project and tried to get its test suite green knows the failure rate. Kwaipilot quantified it: their baseline success rate for constructing a reproducible repo environment was 16.5%.

AutoBuilder is the system that fixed that, pushing it to 57.2% and yielding a corpus of more than 100,000 runnable repositories. Each one is grounded with executable fail-to-pass and pass-to-pass verification signals, which is the distinction that matters. A fail-to-pass test proves the agent fixed the thing it was asked to fix. A pass-to-pass test proves it did not break something else on the way. Grade on the first alone and you train a model that happily bulldozes the codebase to make one test green.

The rest of the stack follows the same logic. A process-aware trajectory pipeline filters agent behavior on more than final pass rates, so a model that stumbles into the right answer by accident is not rewarded as though it reasoned. The post-training framework combines harness randomization, a reliability-hardened sandbox, asymmetric PPO, reward modeling and multi-teacher on-policy distillation. Harness randomization deserves the callout: train against one agent scaffold and you get a model that overfits to that scaffold's quirks. Randomize it and the model has to learn the task rather than the harness.

What does it mean for the market?

The signal for anyone tracking AI infrastructure spend is the Air tier, not the Pro tier. At $0.15 per million input tokens with a 256K window, Air undercuts frontier pricing by roughly an order of magnitude while keeping the context budget that agentic workflows actually consume. Agent loops are token furnaces: every tool call replays context, and a long-horizon task can burn millions of tokens before it converges. That is the cost line that decides whether an autonomous coding agent is economically viable in production or a demo you run twice.

Kuaishou is publicly traded in Hong Kong, and the honest read is that a coding model is not going to move that stock, which lives on short video and e-commerce. The exposure that matters is indirect and sits with the incumbents. Every credible near-frontier model priced at a fifth of frontier rates compresses the margin available to whoever is charging frontier rates for the same job. Watch whether Pro V2.5's SWE-Bench Pro number holds up under independent evaluation, and watch whether Air's pricing survives contact with real inference demand. Cheap launch pricing that quietly climbs after adoption is a familiar pattern. This is factual analysis rather than investment advice.

RelatedMicrosoft Coaches Sales Team to Undercut OpenAI, Anthropic

Who should actually care?

If you are running agentic coding in production, Air is the number to look at, and the caveat is that both models are API-only on StreamLake. There are no open weights, so self-hosting, air-gapped deployment and weight-level auditing are all off the table. For teams with data-residency constraints, that ends the conversation regardless of the benchmark.

If you are building evaluation infrastructure, AutoBuilder is the more valuable artifact. The 16.5% to 57.2% jump in reproducible environment construction is a reusable result, and the two internal benchmarks the team built alongside it, KAT Code Bench and KAT Claw Bench, point at where this is heading: everyone training coding agents is bottlenecked on verifiable environments, not on model architecture.

Our take

Second place on a benchmark is a marketing line. A 3.5x improvement in your ability to automatically construct verifiable training environments is a compounding advantage, and Kwaipilot published the second while everyone reads the first. The lesson of the last two years of coding agents is that the binding constraint moved from model capability to grading capability: you cannot reinforcement-learn your way to a better engineer without a machine that can tell you, cheaply and at scale, whether the engineer was right. AutoBuilder is that machine, and 100,000 runnable repositories is a real asset.

The API-only distribution is the strategic weakness. Kwaipilot built infrastructure worth more than the model it produced, then gated the model behind a single platform with no weights. That caps adoption at whoever is willing to route production code through StreamLake, which in most Western enterprises is nobody. The benchmark says this model belongs in the conversation. The distribution says it will not be in it.

Original analysis by GenZTech. Benchmark figures per the KAT-Coder-V2.5 technical report; scores are vendor-reported under a unified Claude Code harness and await independent confirmation.