Domestic Agent Model Breaks into Global Top Tier! Limited-Time Free Access

Kunlun Tech has quietly broken the "impossible triangle" of Agent models: high performance, low cost, and ease of use. On May 26, 2026, the company released SkyClaw-v1.0 and its lightweight version SkyClaw-v1.0-lite, two native Agent models that stand shoulder to shoulder with the world's top closed-source and open-source models.

Unlike most models that add Agent capabilities as an afterthought, SkyClaw was designed from the ground up for task completion, not just language generation. From day one of training, the focus was on tool calling, multi-step task decomposition, and execution. This fundamental difference directly leads to higher performance ceilings.

SkyClaw-v1.0, the flagship model, is optimized for OpenClaw and can compete head-to-head with Claude Opus 4.6 on related tasks. In demos, it generated a playable Mario game and a fully functional financial terminal with global indices, news feeds, watchlists, and even breaking news popups. The lightweight version, SkyClaw-v1.0-lite, maintains core Agent capabilities while offering faster speed and lower cost, targeting high-frequency, cost-sensitive scenarios.

Pricing is aggressive: SkyClaw-v1.0 costs 0.5 RMB per million input tokens and 4 RMB per million output tokens; the lite version is 0.3 RMB input and 2 RMB output. This is half or less than mainstream top models. During the launch period, both models are free. After the trial, Kunlun Tech plans to open-source each model version progressively.

Integration is straightforward. The models deeply adapt to OpenClaw, Claude Code, Hermes, and Nanobot frameworks, and are compatible with the OpenAI API interface. Developers can often switch by simply changing the base URL and API key. For quick testing, users can access SkyClaw-v1.0 on the Tiangong Skywork platform (tiangong.cn) without any setup. For deep integration, API access is available via APIFree (currently free), supporting streaming output, tool calls, and multi-turn conversations.

Real-world tests demonstrate the model's capabilities. In one test, SkyClaw-v1.0 created a desktop pet cat with random walking, click interactions, work/break modes, a Pomodoro timer, and health reminders. It generated all source code, including SVG graphics, and even suggested IM connection panels for Skywork App, Feishu, Slack, Discord, and Telegram. In another test, it built a complete AI industry weekly report generation system that automatically fetches data from RSS, GitHub, HuggingFace, and web sources, cleans and classifies it, analyzes trends, and generates an interactive report page with SQLite storage. The system can be set as a scheduled task, running automatically every Monday.

The key to SkyClaw's success lies in its training paradigm: native Agent-oriented training from the start, with mid-training exposure to complex Agent tasks and reinforcement learning in real Agent environments. This approach ensures stable tool calling and long-context handling. The models are trained to work across multiple frameworks, avoiding the lock-in issue. By focusing on efficient training rather than massive parameter counts, Kunlun Tech achieved top-tier performance with lower computational cost.

In essence, SkyClaw-v1.0 represents a shift from the traditional "general model + tool wrapper" approach to a purpose-built Agent model. With its combination of high performance, low cost, and easy integration, it is poised to accelerate the adoption of Agent technology across industries.