Agents burn 95x more on yes/no decisions: routing them to cheap decision models
Arindam_1729 · x · 2026-10-13
Developer Arindam shows how agents overpay for trivial decisions: 10,000 yes/no decisions at 1,000 input tokens each cost $0.42 on Jev vs $40 on Opus 5.5 (input only). At 100k decisions/day that's $126/month vs $12,000.
- Most agent calls aren't hard problems — they're condition checks, classification, and next-step routing that don't need frontier intelligence.
- Jev prices at $0.042 per million input tokens with free output.
- Its decision style is novel: instead of "looks good to ship," it returns odds — Ship 52% / Return 46% with confidence 0.18 — behaving differently when confidence is low.
He also published a tutorial on building a Jev-like decision model in 15 minutes; two weeks after Jev's launch, OpenAI and AWS shipped their own decision models.
More from coding & agent
- Cursor now ships real Supabase backend changes via MCP, no SQL needed — XFreeze · 2026-10-14
- Developer: superhuman agents finished the last 20%, so vibe coding is dead — josh_bickett · 2026-10-14
- MIRA-AC trains research agents on proxy feedback, lifting symbolic regression from 50 to 63 solved — anirudhg9119 · 2026-10-14
- MIRA-AC behavior shift: hypothesis-distinguishing tests up 16.8% to 28%, stalled tweaks halved — anirudhg9119 · 2026-10-14
- MIRA-AC RL details: stitching decisions into one logical episode, updating only the meta-reasoner — anirudhg9119 · 2026-10-14
- MCP Event spec debuts: agents that listen to your system and act proactively — irvinebroque · 2026-10-14