SkyClaw-v1.0: High-Performance Agent Model at Unprecedented Value

SkyClaw-v1.0 outperforms Minimax 2.7, DeepSeek V4 Flash, and Qwen 3.6 35B A3B / 27B models across all major agent benchmarks. On OpenClaw-related tasks, its performance approaches that of much larger closed-source models, including DeepSeek V4 Pro, Claude Opus 4.6, and Qwen 3.6 Plus.

In addition to SkyClaw-v1.0, we also offer SkyClaw-v1.0-lite, a much faster and cheaper model that inherits strong agentic performance (e.g., better performance compared to Minimax 2.7). This model is suitable for basic agentic tasks that are more sensitive to costs.

Strong Performance on Agent Benchmarks

Across mainstream agent benchmarks and our internally developed Claw task evaluations — including PinchBench, Claw-Eval (with the ^3 stability test), and Skywork-Claw-Bench (Skywork's in-house agent evaluation suite built on the OpenClaw environment) — both the main and lite versions outperform Minimax 2.7, DeepSeek V4 Flash, and the Qwen 3.6 35B A3B and 27B models. Substantial improvements are also observed on related code-task evaluation metrics.

Model Training

This release underwent extensive large-scale and high-complexity mid-training, high-quality synthetic-task SFT, and systematic reinforcement learning optimization. These improvements not only significantly enhance performance on mainstream tasks in OpenClaw-like environments, but also strengthen the model's generalization ability across different agent harnesses.

Environment Construction

SkyClaw is trained on a high-quality and highly complex OpenClaw environment. The training simulation environment is equipped with frequently used, high-quality tools and skills for the model to explore and utilize. In addition, the team analyzed real user tasks in OpenClaw-like environments, combined with usage statistics and feedback from high-frequency online skills, to construct a relationship graph among tools. Based on this graph, we synthesized complex tasks that more closely resemble real-world user demands.

High-Quality Mid-Training and SFT

Leveraging a thoroughly validated high-quality OpenClaw environment, we synthesized a large amount of training data tailored for both mid-training and SFT. We also established strict data filtering and quality evaluation rules that assess not only final-answer correctness, but also the quality of trajectories throughout the task-solving process. Furthermore, we conducted a series of data-mixture experiments to determine the optimal composition of agent-task training data.

Reinforcement Learning (RL)

Our self-built OpenClaw environment naturally supports exploration and reinforcement learning. End-to-end agentic RL training further improves the model's generalization ability and stability on general-purpose agent tasks.

Pricing and Availability

SkyClaw-v1.0 offers exceptional cost efficiency — pricing is only half or even lower compared to Minimax 2.7 and the Qwen 3.6 series models. A free trial is now available.

API Usage

SkyClaw-v1.0 is available via apifree.ai with an OpenAI-compatible API interface. Get started in minutes:

curl https://api.apifree.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $APIFREE_API_KEY" \
  -d '{
    "model": "skywork-ai/skyclaw-v1",
    "messages": [
      {"role": "user", "content": "Hello, SkyClaw!"}
    ]
  }'

from openai import OpenAI

client = OpenAI(
    api_key="your-api-key",
    base_url="https://api.apifree.ai/v1"
)

response = client.chat.completions.create(
    model="skywork-ai/skyclaw-v1",
    messages=[
        {"role": "user", "content": "Hello, SkyClaw!"}
    ]
)
print(response.choices[0].message.content)

For full documentation, authentication setup, and advanced usage (tool calling, multi-turn conversations, streaming), visit the SkyClaw API page on apifree.ai →

Showcase

Below are examples of tasks completed entirely by SkyClaw-v1.0 in a single session. Each was generated from a single natural-language prompt, demonstrating the model's ability to deliver complete, polished results across diverse domains.

Interactive Web & Games

SkyClaw-v1.0 excels at generating fully functional interactive web applications — from physics simulations to complete game logic — with correct rendering, smooth animations, and proper user interaction handling.

Deep Research & Data Visualization

Beyond coding, SkyClaw-v1.0 can autonomously research real-world topics, collect data from multiple sources, and synthesize findings into interactive, publication-quality dashboards and reports.