AI model ranking

There is no single winner across every job. zerouter splits the LLM leaderboard into seven skill boards so you can pick a model for coding, math, reasoning, writing, research, tool calling, or long context. Rank 1 on each board is the current conservative TrueSkill lead from LLM Stats, not a blended intelligence score.

Rankings on this page are republished from LLM Stats (llm-stats.com). The hub covers coding on LiveCodeBench, math on AIME 2025, reasoning on GPQA, writing on WritingBench, research on MMLU-Pro, tool calling on MCP Atlas, and long context on LongBench v2. zerouter does not run the underlying evals. Scores typically refresh about hourly.

The number on each row is LLM Stats conservative TrueSkill, a lower-confidence bound rather than the raw mean. It is used so a model with a few noisy wins does not jump the board. Compare it as a skill rating on that category, not as a percent on one exam unless the board also shows a named benchmark such as AIME 2025 or GPQA.

zerouter refreshes the published LLM Stats boards about once an hour. New model announcements can land on the New models strip within hours. If a skill board is empty, LLM Stats had no live table for that category at the last pull.

Coding uses LiveCodeBench, math uses AIME 2025, reasoning uses GPQA, writing uses WritingBench, research uses MMLU-Pro, tool calling uses MCP Atlas, and long context uses LongBench v2. The table header also shows the ranking method LLM Stats published, often TrueSkill.

Open a skill board and read rank, organization, conservative rating, and the open or closed weights flag together. That is how this page compares Claude, GPT, Gemini, Grok, Kimi, Qwen, DeepSeek, and other labs without mixing chat popularity with a coding or math score.

No. zerouter republishes the public LLM Stats boards. LiveCodeBench, AIME 2025, GPQA, WritingBench, MMLU-Pro, MCP Atlas, and LongBench v2 are scored by their publishers and aggregated by LLM Stats. Use llm-stats.com for methodology detail.

Input prices are published USD per 1 million input tokens from the LLM Stats catalog when a rate exists. Switch the price board the same way you switch coding or writing. Missing prices mean LLM Stats did not publish an input rate, not that zerouter caller billing is free.