A Kuaishou team just priced 94 percent of Opus 4.8's coding score at a seventh of the cost. The two things holding KAT-Coder v2.5 back are that you can't download it and it stumbles in the terminal.
Kwaipilot, the AI team inside Kuaishou, shipped KAT-Coder-Pro v2.5 and a cheaper Air tier on July 10 at $0.74 input and $2.96 output per million tokens. On SWE-Bench Pro it lands second only to Claude Opus 4.8, and on the money it undercuts Opus by roughly 7.5 times. That combination is rare enough to take seriously. It also comes with two honest asterisks: the model is proprietary and API-only, and its Terminal-Bench score is nowhere near the leaders. Here is what the rate card actually buys.
Photo by Juanjo Jaramillo on Unsplash
Two tiers, both priced to fit under everyone
| Line | Pro v2.5 | Air v2.5 |
|---|---|---|
| Input, per 1M | $0.74 | $0.15 |
| Cached input, per 1M | $0.15 | - |
| Output, per 1M | $2.96 | $0.60 |
| Context window | 256K | 256K |
| Max output | 80K | 80K |
USD per million tokens, from the OpenRouter and CloudPrice listings for the Kwaipilot models. Both tiers share the 256K window and 80K output ceiling. Air is the distilled, roughly 5x-cheaper sibling aimed at high-volume work where you do not need Pro's top score.
What a coding month actually costs
Rate cards flatter budget models, so put a real workload behind the numbers. Take a mid-size coding-agent month at 30M input tokens and 6M output, no cache. KAT-Coder-Pro v2.5 comes to about $40. The frontier coder it is chasing, Opus 4.8, runs $300 on the same traffic, and GPT-5.6 Sol runs $330. That is not a discount, it is a different order of magnitude: a $260 monthly gap against Opus, held open by a model that scores within four points of it on the hardest public SWE benchmark.
| Model | Input / output | 30M in / 6M out |
|---|---|---|
| KAT-Coder-Air v2.5 | $0.15 / $0.60 | $8.10 |
| KAT-Coder-Pro v2.5 | $0.74 / $2.96 | $39.96 |
| Kimi K2.7 Code | $0.95 / $4.00 | $52.50 |
| GPT-5.6 Luna | $1 / $6 | $66.00 |
| GLM-5.2 | $1.40 / $4.40 | $68.40 |
| Claude Opus 4.8 | $5 / $25 | $300.00 |
| GPT-5.6 Sol | $5 / $30 | $330.00 |
Notice Pro is not the cheapest row. Its own Air tier lands at $8, and if all you want is the floor, DeepSeek V4-Flash gets you a coding month under $6. Pro sits in the interesting middle: pricier than the bargain bin, but the only model in that price band that trades blows with Opus on repo-level work. The question is not whether it is cheap. It is whether the score justifies paying five times Air.
Priced by the benchmark point, it is the best deal on the board
Sticker price only tells you what a token costs, not what a working patch costs. Divide that same $40 month by the model's SWE-Bench Pro score and you get a rough dollars-per-point figure that folds capability and price into one line. Kwaipilot's own tech report is the source for the scores, and it uses SWE-Bench Pro, the harder variant, rather than the friendlier Verified set that older KAT versions quoted.
| Model | SWE-Bench Pro | Cost / point |
|---|---|---|
| KAT-Coder-Pro v2.5 | 65.2 | $0.61 |
| GLM-5.2 | 62.1 | $1.10 |
| Claude Opus 4.8 | 69.2 | $4.34 |
Opus buys you the highest score, and if your work genuinely needs those top four points you pay for them at $4.34 a point. But KAT-Coder-Pro delivers 94 percent of that score at 61 cents a point, roughly a seventh of the cost per unit of measured coding ability. Even GLM-5.2, itself a cheap model, costs nearly double per point. On this axis Pro is not just competitive, it is the outright winner, which is exactly the pitch a Kuaishou team needs to make to pull traffic off the US frontier labs.
Where the score falls apart: the terminal
One benchmark keeps this from being an easy call. On Terminal-Bench 2.1, which measures how well a model drives a shell to get real work done, KAT-Coder v2.5 scores 60.7. Opus 4.8 sits at 84.6 and even GLM-5.2 reaches 77.9. That is a wide, specific weakness. A model can top repo-level SWE tickets and still flail when the job is chaining commands, reading tool output, and recovering from a failed step, which is most of what a real coding agent spends its day doing.
It is not all soft, though. On PinchBench, the agentic tool-use test in Kwaipilot's report, Pro scores 94.9 and actually edges Opus 4.8's 93.5, the one benchmark where it takes the lead. So the profile is jagged rather than uniformly second-tier: excellent at structured tool calls and repo patches, shaky at open-ended terminal control. Which half matters depends entirely on how your agent is wired.
| Benchmark | KAT-Coder v2.5 | Field |
|---|---|---|
| SWE-Bench Pro | 65.2 | #2, behind Opus 4.8 (69.2), ahead of GLM-5.2 (62.1) |
| PinchBench (tool use) | 94.9 | Leads Opus 4.8 (93.5), the one win |
| Terminal-Bench 2.1 | 60.7 | Well behind Opus 4.8 (84.6), GLM-5.2 (77.9) |
All of these come from Kwaipilot's own arXiv report, run through a Claude Code harness, so treat them as vendor numbers until an independent lab like Artificial Analysis reruns them. The Terminal-Bench gap is large enough that it would take a lot of measurement error to close, which is why we would not hand this model a shell-heavy autonomous workload without piloting it first.
The other catch: you can only rent it
Most of the cheap Chinese coders that made this year interesting, DeepSeek, GLM, Kimi, Qwen, ship open weights you can self-host. KAT-Coder-Pro and Air v2.5 do not. They are proprietary and API-only, served through Kwaipilot's StreamLake platform and resellers like OpenRouter. Kwaipilot does publish an open-weight line called KAT-Dev on Hugging Face under Apache 2.0, but that is a separate, smaller family, not the model scoring 65 on SWE-Bench Pro. If your reason for shopping below the frontier was avoiding vendor lock-in or running on your own hardware, this one does not qualify, and the low price does not change that.
The Air tier is the part worth a second look. At $0.15 input and $0.60 output it is priced against DeepSeek V4-Flash and GLM's cheapest options, roughly $8 for that same 30M/6M month. Kwaipilot has not published separate Air benchmark numbers, so you are buying on the family reputation and the price, not a verified score. For high-volume, well-scoped tasks where you would reach for a budget model anyway, it is a reasonable thing to A/B against whatever you run now. For anything where the output has to be right the first time, pay up for Pro or stay on the model you trust.
When it is worth switching
KAT-Coder-Pro v2.5 is the strongest cost-per-point coder we have priced this month, so if you are paying Opus or Sol money for repo-level patch generation and structured tool calls, it deserves a pilot. Route a slice of non-critical traffic to it, diff a week of pull requests against your incumbent, and watch the two failure modes specifically: does it hold up when your agent has to run and interpret shell commands, and are you comfortable with a proprietary model from a vendor you may not have heard of a month ago. If both answers hold, the invoice math is hard to argue with.
Skip it if you self-host on principle, if your agent is terminal-driven, or if you were already below this price on an open-weight coder that does the job. The savings only count when the model actually finishes your work, so drop your real input-output mix into the cost calculator and line it up against Opus 4.8 and GLM-5.2 on the pricing page before you move anything that matters.
Sources
- - KAT-Coder-Pro v2.5 on OpenRouter: openrouter.ai
- - KAT-Coder-Air v2.5 on OpenRouter: openrouter.ai
- - KwaiKAT tech report (SWE-Bench Pro, PinchBench, Terminal-Bench): arxiv.org/abs/2607.05471
- - Kwaipilot KAT-Dev open weights: huggingface.co
- - Related: GLM-5.2 vs Kimi K2.7 Code
- - TokenCost pricing page: tokencost.app/pricing