Key Info
OpenRouter notes that while new models get most of the attention, the inference layer improved sharply in 2026: the median agentic cache rate rose to 89% from the mid-60s at the start of the year.
Highlights
- Median agentic cache rate reached 89% in 2026, up from the mid-60s early in the year.
- Better caching translates directly into lower cost on the same model.
- The point is framed as an often-overlooked inference-layer gain, not a model release.