OpenRouter launches Pareto Code with min_coding_score and 1B routed tokens per day

• OpenRouter's announcement post pitched Pareto Code as a free, experimental coding router, and the May 8 changelog describes it as routing to the cheapest code-capable model that clears a user-set min_coding_score.
• According to OpenRouter's usage post, Pareto Router is already processing almost 1 billion tokens per day, while the same post says Auto Router is at 12 billion.
• The Pareto Router docs map min_coding_score into three tiers, low, medium, and high, with omitted scores defaulting to high.
• OpenRouter's follow-up reply says both routers can be customized with guardrails and routing settings, while the Auto Router docs add a cost_quality_tradeoff knob from 0 to 10.

You can read the Pareto Router docs, inspect the live model page, and compare it with Auto Router, which uses NotDiamond for prompt-based selection. The odd bit is timing: OpenRouter's June 3 post framed this as an introduction, but the changelog entry from May 8 and the router page both show the feature had already been live earlier.

Pareto Code is OpenRouter's coding-specific router, exposed as openrouter/pareto-code. The docs say it keeps a curated shortlist of coding models ranked by Artificial Analysis coding percentiles, then picks a model based on the quality bar you set.

The pricing model is simple: the docs say the router itself adds no fee, and you pay the standard rate for whichever underlying model handled the request. The model page also says omitting min_coding_score routes to the High tier by default.

The Pareto Router docs split min_coding_score into three percentile bands:

• >= 0.66: high, the top of Artificial Analysis' coding field
• >= 0.33 and < 0.66: medium, strong flagships below the top tier
• < 0.33: low, still above the AA coding median
• omitted: high, the default

Inside a tier, the same docs say OpenRouter picks the cheapest currently available model. The :nitro variant flips that sort order to p50 throughput, so it chooses the fastest model in the tier instead. The router also keeps up to two same-tier fallbacks, and only steps into a neighboring tier if the whole chosen tier is unavailable.

The other part of the June 3 push is that Pareto is being positioned beside OpenRouter's broader routing stack, not as a one-off model alias. OpenRouter's follow-up reply says both Pareto and Auto can be customized through guardrails and routing settings.

The Auto Router docs add two concrete controls that Pareto does not have:

• allowed_models, for restricting the router to specific model families or providers
• cost_quality_tradeoff, an integer from 0 to 10, where 0 means pure quality and 10 means maximum cost savings

That makes the split pretty clean in the docs. Pareto gives one coding-specific quality bar, while Auto keeps the cross-task prompt analysis and exposes a direct cost versus quality dial. The June 3 numbers from OpenRouter's usage post also show where usage sits today, about 1 billion daily tokens for Pareto versus 12 billion for Auto.

The most useful caveat is that Pareto Code was not actually first published on June 3. The May 8 changelog already introduced pareto-code as an experimental coding router, and the router page lists it as released on April 21, 2026 with a 200,000 token context window.

So the June 3 story is really two things at once: a renewed public rollout, and the first disclosed usage number. OpenRouter's usage post is where the almost-1B-tokens-per-day figure appears, and neither the changelog nor the router docs surfaced that scale earlier.