You will get banned and it is against the terms.
The models the community keeps coming back to (tested in real agent loops, not just chat):
- Claude Sonnet 4 / 4.5 (via official API key) This is the cheat code. 90–95 % of Opus quality for agents, 5–10× cheaper, lightning fast, rock-solid tool calling. Most ex-Opus users say they literally can’t tell the difference in daily OpenClaw use.
- GPT-4o / GPT-5.2 (native OpenAI support) Sometimes even better at certain coding/agent styles. Zero drama.
- Gemini 1.5/3 Pro 1M+ context = god-tier when your agent needs to swallow whole codebases or long histories.
- Moonshot Kimi (k2/k2.5) via OpenRouter Stupidly cheap + surprisingly sharp. A lot of heavy users live here now.
- GLM models