KAT-Coder Air 2.5: Everything You Need to Know Before You Use It

Most developers chasing cheap, capable coding models are watching DeepSeek, GLM, and Kimi. There's a fourth name worth paying attention to, and it's coming from a company better known for short form video than AI research. Here's everything KAT-Coder Air 2.5 actually offers, what it costs, and where it genuinely fits in your stack.

What Is KAT-Coder Air 2.5?

KAT-Coder Air V2.5 is the lightweight, budget tier model in Kwaipilot's KAT-Coder family, released July 10, 2026. Kwaipilot is the AI research division of Kuaishou Technology, the company behind one of China's largest short video platforms, operating its coding model lineup through a dedicated arm called StreamLake.

The KAT-Coder family follows a clear three tier structure. Pro is the flagship, enterprise focused model, built to compete directly with frontier systems on the hardest repository level engineering tasks. Air is the lightweight sibling, positioned explicitly for less demanding, higher volume work at a fraction of Pro's price. A third branch, the Exp-72B academic open source line, releases select model weights publicly for research and local experimentation. Air 2.5 sits in the middle of that spectrum: not free and open weight like the Exp line, but meaningfully cheaper and lighter than Pro.

Specs at a Glance

What KAT-Coder Air 2.5 Is Actually Built For

Kwaipilot's broader V2.5 research, documented in the KAT-Coder-V2.5 technical report, centers on one core argument: further progress in agentic coding depends more on training infrastructure than on raw parameter count. The team built a system called AutoBuilder that reconstructs real multilingual repositories into reproducible, executable sandbox environments with verified pass and fail states, then used it to generate large volumes of high quality training trajectories rather than relying purely on scaling up model size. A companion system, KwaiClawEnv, synthesizes large scale tool use data from real, executable services rather than synthetic scenarios.

One specific engineering detail stands out because it explains a lot about how the whole family behaves in practice. During early reinforcement learning, Kwaipilot found that roughly 16 percent of training failures came from the sandbox environment itself misreporting results, not from the model actually failing the task. After hardening the sandbox, that error rate dropped below 2 percent, and training timeouts fell from 6 to 7 percent to under 1 percent. That kind of infrastructure fix doesn't show up in a benchmark chart, but it's a meaningful part of why the KAT-Coder family, Air included, tends to behave more reliably on multi step, tool calling tasks than its raw parameter count would suggest.

Benchmarks: The Full Picture, Tier by Tier

This is the part worth being precise about, because a lot of the impressive numbers attached to the KAT-Coder name in 2026 technically describe the Pro tier, not Air specifically. Here's every benchmark data point currently available across the family, organized so you can see exactly which model each score actually belongs to.

KAT-Coder-Pro V2.5 (the flagship, for reference)

KAT-Coder-Pro V1 (previous generation, direct benchmark precedent)

KAT-Dev-72B-Exp (open source sibling, closest public proxy for Air-class capability)

KAT-Dev-32B (smaller open source sibling, another reference point)

KAT-Coder-Air V2.5 (this model, specifically)

That last table is the one that matters most if you're deciding whether to trust Air's reputation. As of this writing, there is no confirmed, apples to apples benchmark score for KAT-Coder Air V2.5 itself on any of the standard tests Kwaipilot uses to showcase Pro. Everything positive said about Air's capability right now is inference from the rest of the family: the shared AutoBuilder training infrastructure, the shared agentic reinforcement learning pipeline, and the KAT-Dev-72B-Exp open source model's strong 74.6% SWE-Bench Verified score as the closest publicly available proxy at a comparable scale.

That's a reasonable basis for cautious optimism, not a confirmed benchmark claim. If you see a comparison anywhere quoting KAT-Coder's "second only to Opus 4.8" result and applying it directly to Air, treat that specific claim as unverified until Kwaipilot publishes an Air 2.5 scorecard using the same methodology as its Pro report.

Pricing and How the Credit Math Works

KAT-Coder Air V2.5 is priced at $0.15 per million input tokens and $0.60 per million output tokens, putting it firmly in the budget tier alongside models like DeepSeek V4 Flash and GLM-5.2 Air. For context, that output price is roughly one fifth of what many mid tier proprietary models charge, though it's worth noting that Air's 256,000 token context window and 80,000 token output cap are both smaller than several open weight competitors that now default to context windows in the 1 million token range.

Kwaipilot has also run limited time free access programs on StreamLake in the past, capping registration at a set number of new users per day, which is worth checking for if you want to test the model with zero upfront cost before committing to paid usage.

How to Actually Use KAT-Coder Air 2.5

Kwaipilot has built out genuinely broad integration support for the KAT-Coder family, and most of it applies to Air as well as Pro.

Claude Code. You can route Claude Code through a StreamLake proxy by setting two environment variables, ANTHROPIC_BASE_URL pointed at your StreamLake endpoint and ANTHROPIC_AUTH_TOKEN set to your API key, then launching Claude Code normally. Kwaipilot documents this as a three step process: get an API key and endpoint ID from the StreamLake Wanqing console, set the environment variables, and start coding.

VS Code extensions. Both Kilo Code and Roo Code, popular open source coding agents for VS Code, support KAT-Coder through a custom model configuration pointing at StreamLake's gateway URL.

OpenCode and Droid. Both support KAT-Coder through configuration files pointing at the same StreamLake gateway pattern, letting you add it as a custom model alongside other providers.

Cursor. Custom model configuration is supported, though Cursor restricts this feature to Pro plan subscribers or higher.

Aggregator platforms. KAT-Coder Air V2.5 is also available through OpenRouter and Vercel AI Gateway, which is the simplest path if you're already using either as a multi model routing layer and don't want to manage a separate StreamLake API key.

Where Air Fits Against the Rest of the Budget Tier

KAT-Coder Air's closest competitors are the other lightweight, cost optimized coding models that emerged through 2026: DeepSeek V4 Flash, GLM-5.2 Air, and Kimi's smaller tiers. Air's key differentiators are its direct lineage from Pro's genuinely strong research (the AutoBuilder sandbox infrastructure and agentic reinforcement learning pipeline apply across the whole family) and its broad first party integration support for coding specific tools like Claude Code and VS Code agents, which some competing budget models don't document as thoroughly. Its main limitations relative to the field are a smaller context window than context heavy competitors like DeepSeek V4 Flash's 1 million tokens, and the lack of an independently published, Air specific benchmark table to verify capability claims against.

Who Should Actually Use It

KAT-Coder Air 2.5 is a reasonable choice if you want a budget coding model backed by research infrastructure built for a genuinely frontier competitive flagship, you're comfortable working within a 256K context window, and you want first party support for coding agents like Claude Code or VS Code extensions without extra configuration work. It's a less obvious fit if your workload needs a very large context window for whole codebase analysis, or if you specifically need independently verified benchmark numbers before committing budget to a model, since Air's own scorecard isn't yet as publicly documented as its Pro sibling's.

Bottom Line

KAT-Coder Air 2.5 is a genuinely interesting entrant in the budget coding model space, not because of flashy independently verified benchmarks, since those largely don't exist yet for this specific tier, but because of what it inherits from a team that's clearly serious about the underlying research. The sandbox reliability work, the agentic reinforcement learning pipeline, and the broad tooling integration all suggest a lightweight model built on real engineering rather than a scaled down afterthought. Whether that translates into best in class performance for your specific workload is still worth testing directly, especially given how thin the public benchmark picture is for this exact model. But as a low cost option with strong lineage and genuinely broad IDE and agent support, it's earned a spot on the shortlist.