Provider Variance: Introducing Exacto (openrouter.ai)

OpenRouter announced "exacto" — curated model endpoints that route requests only to a sub‑set of providers that demonstrably deliver higher tool‑calling accuracy and appropriate tool‑use propensity. The move responds to observed provider variance: even when shipping the same open‑weight models, differences in production inference stacks, quantization choices and tuning can change whether an LLM emits valid JSON tool_call payloads, names the right tool, or matches the tool schema. Exacto aims to reduce surprises for agentic workflows by prioritizing providers that pass OpenRouter’s empirical checks and customer preference signals. Technically, OpenRouter aggregates billions of real-world tool_call events and benchmarks (Artificial Analysis, Groq OpenBench, tau2‑Bench, LiveMCPBench) and checks three failure modes per tool_call (JSON validity, tool name presence, schema conformance). Providers are selected if they rank highly on tool‑call accuracy, maintain normal propensity to call tools, and aren’t frequently ignored by users. Exacto endpoints are available now for Kimi K2, DeepSeek v3.1 Terminus, GLM 4.6, GPT‑OSS 120b and Qwen3 Coder. OpenRouter reports measurable reductions in tool‑calling failures and increased correct tool usage; it will continue benchmarking new models, rotate providers based on performance, and publish more underlying data later in the year.