March 2026 Summaries

Auto Exacto is a routing enhancement for tool-calling requests that automatically selects the best-performing providers based on three key metrics: throughput, tool-call telemetry, and benchmark scores, which are re-evaluated every five minutes. This feature improves upon the previous Exacto system by dynamically adjusting provider rankings and addressing initial provider variance during model launches, as demonstrated by the significant reduction in error rates and improved TauBench scores for models like GLM-5 and gpt-oss-120b. The system categorizes providers into tiers based on data sufficiency and performance, without creating a composite score, allowing flexibility and easy debugging. While routing currently focuses on established metrics, future iterations may incorporate additional signals such as pricing and token usage. Auto Exacto is now live, offering default routing for tool-calling requests and optional quality-weighted routing for other requests by appending :exacto to model slugs. The initiative aims to eventually publish provider performance data to encourage continual improvements from providers.

Auto Exacto is an adaptive quality routing system designed to improve tool-calling accuracy by dynamically re-evaluating providers based on throughput, tool-call telemetry, and benchmark scores every five minutes. Initially launched as Exacto with static, manually updated provider lists, the new system addresses feedback by automating provider evaluation and is enabled by default for tool-calling requests. It helps stabilize provider performance during model launches when variance is highest, automatically deranking underperforming providers and promoting improved ones without manual updates. Extensive data collection since August 2025 has allowed for rigorous evaluation using controlled benchmarks like TauBench Airline and GPQA-Diamond, resulting in significant reductions in tool call error rates across tested models. Auto Exacto categorizes providers into tiers based on their statistical performance, maintaining the original routing order for price and latency while pushing underperforming providers to the back. The system is designed to be transparent, with plans to publicly share provider performance data, and allows for user customization of routing preferences to balance quality and cost considerations.