OpenClaw is a versatile, open-source AI agent platform that allows users to connect various large language models (LLMs) to automate a wide range of tasks. The choice of model is a critical decision that impacts the agent's performance, cost, and overall effectiveness. This document provides a comprehensive overview of the best paid and free models available for OpenClaw, based on recent community discussions and official documentation.
Model Tiers and Performance
The OpenClaw community has developed a practical rubric for classifying models into tiers based on their reliability for agentic tasks, particularly those involving tool-use and automation. The tiers are defined as follows:
•S-Tier (Agent Default): These models exhibit highly reliable tool-calling capabilities, with excellent JSON/schema correctness and the ability to recover from failures. They are ideal for complex automation and multi-step tasks.
•A-Tier: Models in this tier are mostly reliable for tool-use but may have occasional formatting slips. They are great for general chat, coding, and less critical automation.
•B-Tier: These models are suitable for chat and isolated coding tasks but can be flaky with more complex agentic workflows. They may require more supervision.
•C-Tier: C-Tier models are the most affordable (or free) and are best suited for simple, non-critical tasks like heartbeats or basic chat.
Paid Models
Paid models generally offer the highest performance and reliability, making them the preferred choice for production environments and critical tasks. They are typically accessed via API keys from their respective providers.
Model Provider | Model Name | Tier | Pricing (per 1M tokens) | Notes |
Anthropic | Claude Opus 4.6 | S | Varies | The top choice for tool-heavy agent work . |
OpenAI | GPT-5.3 Codex | S | Varies | Optimized for code generation and complex reasoning. |
Google | Gemini 3 Pro | A | Varies | A strong contender for general-purpose tasks. |
DeepSeek | DeepSeek V3.2 | B | ~$0.25 (input) / ~$0.38 (output) | A budget-friendly option with good performance. |
Google | Gemini Flash-Lite | C | ~$0.10 (input) / ~$0.40 (output) | A cost-effective choice for simpler tasks. |
Free Models
Free models have become increasingly powerful, with some now rivaling the performance of paid alternatives. These models are an excellent choice for developers, hobbyists, and those on a budget.
Model Provider | Model Name | Tier | Access Method | Notes |
Moonshot AI | Kimi K2.5 | S | OpenClaw Platform | The first open-source model to reach top-tier status, offered for free on OpenClaw . |
Qwen | Qwen Coder/Vision | A | Qwen Portal OAuth | Provides a generous free tier with 2,000 requests per day . |
MiniMax | MiniMax M2.1 | B | MiniMax OAuth | Offers a free 7-day coding plan for OpenClaw users . |
Local Models | Various | B/C | Ollama, vLLM, etc. | Self-hosted models that offer privacy and control, with performance varying by model and hardware. |
Recommendations
For most users, a two-model strategy is recommended :
1.Primary Model: A reliable S-Tier or A-Tier model for complex tasks, browser automation, and coding.
2.Secondary Model: A cheaper B-Tier or C-Tier model for simple chat, heartbeats, and less critical tasks.
This approach provides a balance of performance and cost-effectiveness. As the AI landscape is constantly evolving, it is recommended to stay updated with the latest community discussions and model releases.
Comments