{"slug": "show-hn-taskrouter-one-api-learning-to-choose-models-per-ai-task", "title": "Show HN: Taskrouter – One API Learning to Choose Models per AI Task", "summary": "Neurometric launched Task Router, an API that selects AI models per task based on published benchmark evidence and a user-set quality floor, keeping the OpenAI SDK and changing only the base URL and model alias. On a measured 100,000-email task, the company reports routing cost of $15.21 plus $14.85 in inference versus $666.00 for its baseline, a $635.94 saving at 95% quality, with Qwen3 4B Instruct 2507 selected at $0.4906 per 100,000 tickets and fallbacks Gemma 4 E4B IT, Ministral 8B Instruct 2410, and Granite 4.1 8B. Pricing is free for 1,000 routing decisions per month, then $0.15 per 1,000 decisions, plus inference at the provider's price with no markup.", "body_md": "Neurometric Task Router\n\n# One API.\n\n*Intelligent Routing*\n\nOn All Your Tasks\n\nOther routers guess from price tags and leaderboards. Task Router learns from your workloads so your inference is always optimized for your requirements.\n\nKeep your OpenAI SDK. Change the base URL and model alias.\n\nSee routing in action\n\n## Your requirements choose the model.\n\nChoose a task and set your quality target. See which models qualify, which one is selected, and the fallback order—all from published benchmark evidence.\n\n1. 1 · Your task Quick-Reply\n2. 2 · Your requirements 90% minimum quality · lowest cost\n3. 3 · Model selection Qwen3 4B Instruct 2507\n4. 4 · Stable API endpoint taskrouter/<your-route>\n\nLowest-cost policy. Quality is measured on the published evaluation set; it is not a guarantee for every request.\n\nSign in to save your route. Your task and requirements come with you.\n\nSelected model\n\n### Qwen3 4B Instruct 2507\n\nSelected because it meets your 90% quality target at the lowest measured cost.\n\nFallback order: Gemma 4 E4B IT → Ministral 8B Instruct 2410 → Granite 4.1 8B\n\n| Measured inference cost for 100,000 ticket. Routing fees are additional. |  |  |  |  | \n|---|---|---|---|---|\n| Model | Quality | p50 | Cost | Selection | \n|---|---|---|---|---|\n| Nemotron Nano 9B v2 | 100.0% | 2.33s | $9.402 | Qualifies | \n| Qwen3 235B A22B | 99.0% | 1.12s | $11.198 | Qualifies | \n| DeepSeek V3 | 100.0% | 1.07s | $24.696 | Qualifies | \n| Mistral Large 2407 | 100.0% | 1.47s | $153.30 | Qualifies | \n| Arcee Trinity Large Thinking | 99.0% | 1.41s | $42.08 | Qualifies | \n| GPT-5.4 | 100.0% | 1.31s | $145.50 | Qualifies | \n| Claude Opus 4.7 | 100.0% | 2.04s | $490.60 | Qualifies | \n| Gemini 3.1 Pro Preview | 100.0% | 3.97s | $455.60 | Qualifies | \n| Gemma 4 E4B IT | 95.0% | 0.47s | $1.0234 | Fallback 1 | \n| Granite 4.1 8B | 99.0% | 0.72s | $2.13 | Fallback 3 | \n| Ministral 8B Instruct 2410 | 100.0% | 1.39s | $1.3061 | Fallback 2 | \n| Qwen3 4B Instruct 2507 | 100.0% | 1.80s | $0.4906 | Selected | \n\nEvidence preview, not a live request. Route creation checks current availability; your API alias stays stable as qualified models change.\n\nWhy task-aware routing\n\n## Model choice is a task question, *not a model question.*\n\nA model can excel at one workload and fall short on another. General leaderboards cannot tell you which. Task-level measurements can.\n\n### Choosing a model yourself\n\n- Compare general benchmarks and price tags\n- Build and run your own task evaluations\n- Review each new model and update your integration\n- Maintain your own fallback chain\n\n### With Task Router\n\n- Inspect published evidence for the task you are running\n- Set a quality floor and an optional latency cap\n- Keep a stable route alias as qualified models change\n- Use an ordered fallback chain of qualified models\n\nFor developers\n\n## Two changes to your code.\n\n*Evidence already built.*\n\n1. 1### Create a route for your taskPick a published task, set your quality floor, and choose a routing preset.\n2. 2### Point your OpenAI SDK at itUse your organization API key and the stable taskrouter/<slug> alias.\n3. 3### Ship, then track the routeWith auto-optimize on, the route can adapt as qualified models and prices change.\n\n``` python\nimport os\nfrom openai import OpenAI\n\nclient = OpenAI(\n    base_url=\"https://api.neurometric.ai/v1\",\n    api_key=os.environ[\"NEUROMETRIC_API_KEY\"],\n)\n\nresponse = client.chat.completions.create(\n    model=\"taskrouter/privilege-triage\",\n    messages=[{\"role\": \"user\", \"content\": email_thread}],\n)\n```\n\nReplace `taskrouter/privilege-triage` with the alias shown on your route.\n\nPricing\n\n## Free\n\nfor 1,000 routing decisions per month\n\nThen **$0.15** per 1,000 decisions\n\nPlus inference at the provider's price, with no markup. Routing only pays for itself if it saves you money, so here's the math on one measured task.\n\n**100K emails**\n\n**$666.00**\n\n**$15.21**\n\n**$14.85**\n\n**$635.94 · 95%**\n\n## Not ready to route yet?\n\nTask Explorer is free. Describe a task, compare measured quality, cost, and speed, and inspect the evidence before you commit.\n\n[Open Task Explorer →](https://taskrouter.com/explorer)\n\n## Questions developers ask\n\n## What if my task is not supported?\n\nDescribe it in Task Explorer to find related published evidence. If nothing fits, submit a benchmark request for review.\n\n## Is it OpenAI-compatible?\n\nUse the OpenAI SDK, set the base URL to api.neurometric.ai/v1, and use the model alias shown on your route.\n\n## What if no model clears my quality floor?\n\nYou cannot activate a new route without a qualified model. The optimizer keeps an existing route’s last-good plan when it cannot qualify a replacement.\n\n## How does this work with OpenRouter?\n\nOpenRouter is a serving provider. Task Router adds task-specific benchmark evidence and a policy that orders qualified model candidates.\n\n## What can I control?\n\nChoose lowest cost, balanced, or highest quality, set a quality floor and p50 latency cap, and turn automatic optimization on or off.\n\n## Do I need to sign in to explore?\n\nNo. Explorer, Tasks, Models, and Providers are public. Sign in to run Playground requests, save a watchlist, or create production routes.", "url": "https://wpnews.pro/news/show-hn-taskrouter-one-api-learning-to-choose-models-per-ai-task", "canonical_source": "https://taskrouter.com", "published_at": "2026-10-02 16:21:18+00:00", "updated_at": "2026-10-02 16:36:34.677312+00:00", "lang": "en", "topics": ["ai-products", "ai-tools", "large-language-models", "ai-infrastructure", "developer-tools"], "entities": ["Neurometric", "Task Router", "Qwen3 4B Instruct 2507", "Gemma 4 E4B IT", "Ministral 8B Instruct 2410", "Granite 4.1 8B", "OpenAI SDK", "GPT-5.4"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/show-hn-taskrouter-one-api-learning-to-choose-models-per-ai-task", "markdown": "https://wpnews.pro/news/show-hn-taskrouter-one-api-learning-to-choose-models-per-ai-task.md", "text": "https://wpnews.pro/news/show-hn-taskrouter-one-api-learning-to-choose-models-per-ai-task.txt", "jsonld": "https://wpnews.pro/news/show-hn-taskrouter-one-api-learning-to-choose-models-per-ai-task.jsonld"}}