Releases·Tool Release
Jev Codex Router: per-turn model and reasoning routing for Codex via TypeSafe System One, cutting costs ~60% vs full-frontier baseline
A new GitHub project, jev-codex-router, adds per-turn model and reasoning routing for Codex using Jev (TypeSafe System One). Each turn is classified by Jev, which selects the cheapest capable model, thinking depth, and speed mode, instead of defaulting to a frontier model for everything. The author reports roughly 60% cost savings versus a full-frontier baseline on a 7-day replay of 237 real turns.
58 engagement1 sourceThu, Sep 17, 2026, 04:14 AM
AI Overview
- jev-codex-router classifies each turn with Jev and routes to the cheapest capable model, adapting thinking depth and speed mode per turn
- Routing decision costs ≈$0.00003 and ≈0.6s per turn; measured savings ≈−60% vs full-frontier baseline on a 237-turn 7-day replay
- Plugs into an existing local Codex Router via a generic provider plus curated model, relaying SSE streams verbatim with fail-open fallback and a kill-switch sentinel file
Related
Products·launched
jev: CLI router model that routes questions to cheap or expensive LLMs to cut costs
jev is a CLI built around TypeSafe AI's JEV, a small router model that reads a question and decides which LLM should answer it. It routes cheap/simple questions to cheaper models and complex ones to expensive models, cutting spend while keeping quality. On failure it falls back to a default tier so the request always completes.
jev is a CLI built around TypeSafe AI's JEV, a small router model that reads a question and decides which LLM should answer it. It routes cheap/simple questions to cheaper models and complex ones to expensive models, cutting spend while keeping quality. On failure it falls back to a default tier so the request always completes.
1 engagement1 sourceredditMon, Sep 21, 2026, 05:29 AM
Social·Community
Reddit user routes local 3B/4B/12B/26B MoE models with Jev decision model, beating keyword rules 16/16
A Reddit user detailed replacing a keyword-rule router with Jev, a decision model, to route prompts across four local model tiers (3B to 26B MoE) served via LM Studio on an RTX 4060 Ti setup. The classifier picked the correct tier 16/16 on labeled prompts and caught refusals 15/15, costing roughly $0.0002 per routing decision. The writeup offers a practical template for multi-model local inference where a misroute forces a 15-45 second model swap.
A Reddit user detailed replacing a keyword-rule router with Jev, a decision model, to route prompts across four local model tiers (3B to 26B MoE) served via LM Studio on an RTX 4060 Ti setup. The classifier picked the correct tier 16/16 on labeled prompts and caught refusals 15/15, costing roughly $0.0002 per routing decision. The writeup offers a practical template for multi-model local inference where a misroute forces a 15-45 second model swap.
27 engagement1 sourceredditFri, Sep 18, 2026, 08:02 PM
Social·Community
Practitioners argue LLM routing misses the point; token efficiency is the real problem
A practitioner argues that the wave of LLM routers from Ramp, Martian, Coinbase, Devin, and Cursor is misguided, claiming the real problem is token efficiency, not model routing. They contend benchmark-based routers are fundamentally flawed because benchmarks are maps, not routing tables.
A practitioner argues that the wave of LLM routers from Ramp, Martian, Coinbase, Devin, and Cursor is misguided, claiming the real problem is token efficiency, not model routing. They contend benchmark-based routers are fundamentally flawed because benchmarks are maps, not routing tables.
9 engagement1 sourceredditSun, Aug 2, 2026, 06:04 PM