OpenRouter Ships Next-Gen Auto Router and Auto Code: Coding Router Benchmarks Near Opus 5.5 at ~74% Lower Cost per Task

Core Background & Industry Pain Points

Coding agents often swap model slugs to chase quality versus cost. OpenRouter's Auto Router has run about three years and processes over 2.5T tokens per month; the prior generation used coarse signals such as task type and underperformed in coding harnesses. On 2026-11-06 OpenRouter rebuilt Auto Router and shipped coding-focused Auto Code.

Architecture Highlights & Internals

openrouter/auto and openrouter/auto-code share one engine: narrow via guardrails, ZDR, and allow/deny lists; score eleven request factors; pick model and reasoning effort from live prices, cache, and benchmarks; then provider routing. Auto Code starts in a higher cost tier and weighs speed and failures more heavily; beta stayed on the same model for about nine of ten turns to protect prompt caches. Default cost_tier is less "cheap-drift" than recently; set cost_tier=low for prior behavior.

Authoritative Benchmarks & Measured Scores

Vendor-stated: Auto Code roughly matches Claude Opus 5.5 across five benchmarks at about 74% lower cost per task; routing-step p90 about 0.65s in a live test (prior about 1.96s). See Model Routing Benchmarks. OpenRouter notes no router beats the single best frontier model at every cost-quality point.

Developer Hands-on Guide

Set model to openrouter/auto or openrouter/auto-code; send session_id for agent sessions. Use allowed_models/excluded_models; enable X-OpenRouter-Metadata or the chatroom insights panel for per-request reasons. See external_links for docs and the announcement blog.