KwaiKAT releases KAT-Coder-Pro V2 with AAI 44 and 109 OTPS

• Artificial Analysis says KwaiKAT's launch thread put KAT-Coder-Pro V2 at 44 on its Intelligence Index, up 8 points from V1 and roughly level with Claude Sonnet 4.6 among non-reasoning models.
• The same benchmark thread says V2 made its biggest jump on agentic work: Terminal-Bench Hard rose to 49%, up 40 points from V1, while GDPval-AA reached 1123 Elo.
• On efficiency, Artificial Analysis' token breakdown puts the run at about 8.7M-9M output tokens, well below similarly capable reasoning models, while its pricing post estimates $73 to run the full index at a public price of $0.30/$1.20 per 1M input/output tokens.
• Speed is part of the pitch too: according to speed post, KAT-Coder-Pro V2 runs at about 109 output tokens per second, and omniscience post says its hallucination-oriented Omniscience score improved by 15 points, driven more by fewer hallucinations than higher accuracy.

KAT-Coder-Pro V2 is a text-in/text-out coding model with a 256K context window, available through StreamLake and AtlasCloud, and Artificial Analysis ranks it just behind Claude Opus 4.6 among non-reasoning models launch thread. The headline number is the 44 Intelligence Index score, but the more implementation-relevant change is where the gains came from: Artificial Analysis says V2 now posts 90% on Tau2-Telecom tool use, 49% on Terminal-Bench Hard, and a 1123 GDPval-AA score after a large jump in agentic evaluations full breakdown.

The tradeoff is that V2 did not improve uniformly. Artificial Analysis says the model regressed on long-context reasoning and knowledge recall versus V1, falling 8 percentage points on AA-LCR to 66% and 17 points on HLE to 16% launch thread. Its hallucination profile improved anyway: the omniscience post reports a -22 Omniscience score, up 15 points from V1, with the gain coming primarily from reduced hallucinations rather than a higher share of correct answers.

The release matters because V2 is posting frontier-adjacent coding numbers without a reasoning pass. Artificial Analysis says that design keeps output-token use to about 8.7M for the full index, versus roughly 11M for Claude Opus 4.6, 14M for Claude Sonnet 4.6, and far more for reasoning models such as DeepSeek V3.2 and Qwen3.5 397B A17B token breakdown. That lower token budget helps explain why the model can hit about 109 output tokens per second and, in Artificial Analysis' phrasing, deliver "one of the fastest end-to-end response times" because there is no extra reasoning delay speed post.

Cost lands in the same pattern. Artificial Analysis prices the model at $0.30 per 1M input tokens and $1.20 per 1M output tokens, and estimates $73 to run the Intelligence Index, down slightly from V1's $76 because V2 needed fewer turns in agentic evals pricing post. That is still not the absolute cheapest option at this capability tier, but it undercuts Sonnet 4.6 by a wide margin in Artificial Analysis' estimate and comes close to DeepSeek V3.2 while staying in a non-reasoning latency envelope; the model page has the full benchmark breakdown.