Best LLM for agents — tools, loops and reliability
Agents multiply failures: each tool call is a chance to pass a bad argument. The best LLM for agents is the one that follows tool schemas and knows when to stop — not the wittiest chat model.
Quality index is a rough prior. You still need to measure valid JSON, correct tool names, and loop length. Cost matters because agents burn tokens on traces. A cheaper model that finishes in three steps can beat a flagship that ramble-calls ten tools.
Highest quality index right now
Live OpenRouter pricing and Arena quality, cached about an hour. Not a static blog table.
| Model | Provider | Blend $ / 1M | Quality | Context |
|---|---|---|---|---|
| Qwen3.8 Max (0902) | Qwen | 4.00 | 95.1 | 1,000,000 |
| Claude Opus 5 | Anthropic | 15.00 | 93.4 | 1,000,000 |
| Claude Fable 5 | Anthropic | 30.00 | 91.1 | 1,000,000 |
| GLM 5.3 Flash | Z.ai | 0.16 | 89.7 | 1,048,576 |
| Qwen3.8 27B | Qwen | 1.71 | 89.0 | 1,000,000 |
Tool calling vs text
If your framework sends JSON schemas, pick models known to respect them. If you parse free-form text, you will fight edge cases. Compare models in the arena with a prompt that says “call only these tools” and see who complies.
Latency in a loop
Eight sequential calls at 2s TTFT is a 16s wait before work happens. Fast, cheaper models often feel better in an agent UI even if single-shot quality is lower. Read the tokens-per-second guide.
Reasoning models in agents
Thinking SKUs can plan better and stall more. Use them for the planner node, not for every observation. Mixed graphs (cheap executor, expensive planner) usually win on cost.
Context growth
Traces accumulate. A 128k window fills up with junk. Summarize observations; do not buy a million tokens to hoard logs.
Evaluation
Define success as “task complete, no extra side effects.” Run 30 tasks. Rank models by success/dollar, not Elo. Use this site to read $/1M so the denominator is honest.
FAQ
Is GPT the best agent model?
- It is a common default because of tool ecosystems. Claude, Gemini, and others compete. Measure schema validity on your tools.
Can open models run agents?
- Yes, especially if you host a fine-tune for your tool set. Start from a strong open instruct row on the leaderboard.
Why is my agent bill so high?
- Retries, traces, and reasoning tokens. Cap steps and use a cheap model for routine tools.
Do I need long context for agents?
- You need enough for tools + a short memory. Huge windows without memory hygiene still fail.
Related guides
Also try: LLM leaderboard, compare tool, live arena.