Which benchmark should you trust for what?
A capability × status map of LLM benchmarks. Green is still discriminative; gray has saturated. Every score on every page is sourced — vendor-reported and independent numbers are kept distinct.
80 / 80 shown
| capability ↓ / status → | Active | Nearing saturation | Saturated | Deprecated |
|---|---|---|---|---|
| knowledge | ||||
| reasoning | ||||
| maths | ||||
| coding | ||||
| instruction-following | ||||
| long-context | ||||
| agentic-tool-use | ||||
| multimodal | ||||
| safety-alignment |