AI News

Neurometric has introduced Task Router, an API service that selects AI models for individual workloads instead of relying only on general leaderboards or provider price lists. Users choose a published task, set a minimum quality target and optionally a p50 latency cap, then select a policy such as lowest cost, balanced, or highest quality. The router uses published task-level evaluation results to determine which models qualify, selects an eligible model, and creates an ordered fallback chain. Its evidence preview shows quality, p50 latency, measured inference cost, and selection status; the site says the catalog covers 263 published tasks and 3,120 measured model results. In one example for a Quick-Reply task with a 90% quality floor and lowest-cost policy, Qwen3 4B Instruct 2507 is selected, while Gemma 4 E4B IT, Ministral 8B Instruct 2410, and Granite 4.1 8B form the fallback order. The service is OpenAI SDK-compatible: developers change the base URL and use a stable `taskrouter/<slug>` model alias. Automatic optimization can adapt the route as qualified models and prices change, while an existing route retains its last-good plan if no replacement qualifies. Task Router charges nothing for the first 1,000 routing decisions each month, then $0.15 per 1,000 decisions, in addition to provider inference costs with no stated markup. The site illustrates a measured attorney-client privilege triage scenario in which routing to Qwen3 235B is presented as reducing the total cost from $666 to $30.06, but it describes the figures as task-specific measurements rather than a guarantee for every request. Users can explore tasks and evidence without signing in; creating production routes and saving them requires an account.