Skywork released SkyClaw-v1.0, a model tuned for tool use and multi-step agent tasks, with a cheaper lite variant alongside it. Both versions are live on apifree.ai through the new model page.
The pitch: pick a tool, build the call, parse the response, keep going across long action chains without falling apart. Skywork says it trained the model inside its OpenClaw environment using synthetic tasks built from real user patterns. That positions it as purpose-trained for agentic work rather than a chat model with tool calling stapled on.
Benchmark claims are thinner. Skywork reports both variants beat Minimax 2.7, DeepSeek V4 Flash, and two Qwen 3.6 sizes on PinchBench, Claw-Eval, and its in-house Skywork-Claw-Bench. The third one carries Skywork's own name. No independent runs have surfaced yet, and the listing skips pricing detail entirely.
Bottom Line
Two of the three benchmarks cited are Skywork-affiliated, and the company hasn't published independent results or pricing.
Quick Facts
- Two versions released: SkyClaw-v1.0 and SkyClaw-v1.0-lite
- Trained inside Skywork's OpenClaw environment on synthetic agent tasks
- Company-reported wins on PinchBench, Claw-Eval, and Skywork-Claw-Bench
- Listed competitors: Minimax 2.7, DeepSeek V4 Flash, Qwen 3.6 35B A3B, Qwen 3.6 27B
- API access via apifree.ai; pricing not disclosed on the model page
