Baseten Router cuts coding-task costs 72% with only 0.8pt pass-rate drop vs Opus 5.5
baseten · x · 2026-11-05
Baseten launched Baseten Router, an intelligent model router for coding tasks on Baseten Grid that picks models by expected performance and cost per task. Key results:
- 72% cheaper than always using Opus 5.5 ($0.38 vs $1.34 per task) while staying within 0.8 percentage points on pass rate (69.6% vs 70.4%), measured across DeepSWE, SWE Atlas QnA, and Terminal-Bench
- Training mix: 92.7% software engineering (SWE-REBENCH 35.3%, MiMo-v2.6-RL code tasks 24.0%, SWE-GYM 17.1%, Multi-SWE-RL 14.6%), plus Q&A and terminal-use data
- Intermediate layers of GLM 5.3 Flash (layers 23, 31) predict agent step count with 0.57-0.59 Spearman correlation
- Routing threshold is adjustable for different cost/performance trade-offs; benchmarks show it beating OpenRouter Auto, FireRouter, and other routers on the Pareto curve
More from coding & agent
- Max out Claude thinking and ultracode on every prompt, says dev — curious_vii · 2026-11-05
- After 100+ team interviews, Swytchcode argues MCP is for building agents, not running them — curiousblack99 · 2026-11-05
- Yuchen declares 'Codex won the agent war' — Yuchenj_UW · 2026-11-05
- Interactive home-emergency trainer built with Opus 5.5, Blender and Three.js — techartist_ · 2026-11-05
- Developer proposes a 12-level AI coding autonomy scale based on who does the work — nimareq · 2026-11-05
- Claude Cookbook shows how to build a nightly "dreaming" memory service for agents — brada · 2026-11-05