Kwaipilot publishes KAT-Coder-V2.5 technical report, an agentic coding model trailing only Opus 4.8

Kwaipilot, the AI coding team at Chinese short-video giant Kuaishou, published a technical report on July 6 for KAT-Coder-V2.5, an agentic coding model the team says ranks second only to Anthropic's frontier Opus 4.8 on repository-level software engineering benchmarks.

What's new

The paper's abstract frames the model's design goal plainly: "We present KAT-Coder-V2.5, a coding-focused agentic model trained to act autonomously inside real, executable repositories rather than as a single-turn code generator." Rather than treating coding as a one-shot generation task, the model is built to operate inside live, executable repos — planning, editing, running, and iterating the way a human engineer would inside an actual codebase.

Key claims from the report:

• Benchmark results: the team reports the model "delivers the best agentic tool-use result on PinchBench and ranks second only to the frontier Opus 4.8 on repository-level software engineering."
• Training approach: rather than leaning primarily on scale, Kwaipilot attributes the gains to "an end-to-end agentic post-training framework" combining automated repository reconstruction, reinforcement learning optimization tied to verifiable rewards, and multi-teacher distillation.
• Availability: a lighter variant, KAT-Coder-Air V2.5, is already listed on OpenRouter, ModelsLab, and Vercel AI Gateway, with a 256K-token context window, up to 80K output tokens, and pricing starting at $0.15 per million input tokens and $0.60 per million output tokens — aggressive pricing relative to Western frontier coding models.

Context

Kwaipilot sits inside Kuaishou, better known internationally for its short-video app and for the Kling video-generation model, but it has been building out a dedicated coding-agent line under the KAT-Coder name. The V2.5 report follows a wave of open and semi-open agentic coding models out of Chinese labs this year — alongside DeepSeek, Qwen, and Zhipu/GLM — all competing on the same repository-level agentic benchmarks that Anthropic's Claude Code and OpenAI's Codex line have made the industry's reference point for "real world" coding capability, as opposed to single-function code-completion tests.

Why it matters

A model that Kwaipilot's own benchmarking places second only to Opus 4.8 on repository-level engineering tasks — at a fraction of frontier pricing and available immediately through multiple API aggregators — is a meaningful data point in how fast the gap between frontier labs and fast-following coding-agent teams is closing. Independent verification of the PinchBench and repository-level claims hasn't happened yet, and self-reported benchmark numbers from any lab should be read with that caveat. But the broader trend — cheap, capable agentic coding models proliferating out of Chinese AI labs and immediately landing on OpenRouter and Vercel's model marketplaces — is squeezing margins across the entire agentic-coding tooling stack, from API pricing to the coding-assistant products built on top of it.