Jev Router cuts coding agent costs 67-85% by routing tasks to cheap open models
HowDevelop · x · 2026-10-19
The author extends TypeSafe's Jev decision model from one-shot routing to multi-step coding agents, letting it pick between open models on Nebius Token Factory instead of paying fixed-model prices for every file read and small edit. Running 20 hidden-test coding tasks via Claude Code + nconnect: fixed Kimi K3 cost $15.41 (18/20 passed), fixed Muse Spark $2.80 (19/20), fixed Reflection Beam $1.89 (20/20, $0.088 median/task, 72s, $0.25/$1.00 per M tokens). With Auto + Jev routing across 32 difficulty-graded tasks, savings vs fixed Kimi K3 were 81% on easy ($4.42→$0.84), 85% on medium ($12.75→$1.96), and 67% on hard ($8.13→$2.66), with pass rates held or improved. Task-difficulty-based routing to cheap open models can slash agent costs without hurting completion.
More from coding & agent
- LobsterAI report: Agent task Token size in Sept is 3.2x that of March — xiaohu · 2026-10-19
- Claude autonomously works through a data pipeline checklist, from schema to cost pilot — andrew_n_carr · 2026-10-19
- LobsterAI DAU up 83.1% in September as office agents become coding night school — lxfater · 2026-10-19
- LobsterAI report: DeepSeek cache hit rate hits 97.3%, cutting effective cost to ~6.2% of list price — lxfater · 2026-10-19
- China office agent report: coding surges over Golden Week, peak usage shifts to 10pm — lxfater · 2026-10-19
- Opus 'cheats' at cloud rendering — author calls in a second agent to audit its diagnosis — teortaxesTex · 2026-10-19