Route Anything: one router across LLMs, media generators and robot policies, 8.82x efficiency
_akhaliq · x · 2026-11-05
The "Route Anything" paper presents a single router that assigns each input to the best model across language models, image/video/speech generators, coding agents, and robot policies, trained and evaluated on 374,401 scored runs across 25 benchmarks.
Key results:
- 8.82x average efficiency gain over a reference model at near-zero quality loss
- Per-input oracle headroom over the best fixed model: video generation +18.9, robots +18.4, code & agents +8.1, image generation +5.0, text QA +4.3
- Price-based routing: 25.5x cost savings on code and 7.1x on text QA (ΔV ≈ −0.4 pts); 1.35x on image generation, 1.33x on robot latency
Takeaway: modalities with the widest routing space (video, robots) gain most, and price routing is especially compelling for code and QA workloads.
Related event: Route Anything: A Unified Router for Cross-Modal Model Dispatching(2 posts)→
More from coding & agent
- Conductor, the praised agent orchestrator, joins Meta to build a new interface with MSL — pzakin · 2026-11-05
- Cutting LLM-as-judge eval costs ~70% with prompt caching and a decision-model cascade — Longjumping-Piano550 · 2026-11-05
- particle-1 tested on real agent work: near-Opus quality for under $2 a day — heyneighbor · 2026-11-05
- Aether runs Claude Code and Codex on your existing AI subscription, charging only for the cloud machine — Scobleizer · 2026-11-05
- Claude Code desktop adds /storage to view and clear local disk usage — lydiahallie · 2026-11-05
- AI code is unmaintainable? Human code wasn't a masterpiece for 30 years either — feregri_no · 2026-11-05