Discera 1.0 · Router benchmark
The router is the entire product — one chat that picks the best AI for each message. A claim like that needs numbers, so here are ours: the test, the scores, and where it’s wrong.
The test set
147 real questions across 12 categories: coding, writing, math, research, documents, images, and more. Each question is labeled with the model that should handle it. The labels are our own judgment calls. The set grows over time: every time the router gets one wrong in the real world, that question joins the test.1
How we grade it
Every question gets one label: the AI that should handle it. The rule is simple — match the task to the model’s strengths:
- Writing, explanations, document help, code review, personal advice → Claude
- Math, science, writing code → GPT
- Anything needing live information → Perplexity
- Images → GPT Image
- Small talk with no clear task → general chat
- Things Discera doesn’t do (video, voice calls) → correctly refused
The labels are our judgment calls, made before the test runs — the router never sees them. Borderline questions are the most useful ones to keep.
The 147 questions were written by hand: 56 seed questions including real misses caught during development, plus 91 more written to cover all 12 categories. Nothing is scraped and nothing comes from user traffic — there isn’t public traffic yet. The set is built for coverage, deliberately including the boundary cases where two categories overlap.
The scores
Tested against the live production site on October 10, 2026 — the same routing code that serves real users, called through the classify endpoint (routing decision only, no answers generated, so the test cost pennies and touched nobody’s chats). Every question ran three times.
73.7% across all three runs (run 1: 73.5%, run 2: 73.5%, run 3: 74.1%)
146 of 147 questions routed identically every single run — the router is deterministic where it counts.2
What the numbers mean
The router gets roughly 3 out of 4 questions right on the live site. Explanations and writing are essentially solved. The weak spots are coding review (50%) and documents (54%) — categories where the question wording overlaps heavily with general explanations. Those are exactly what the 1.5 router rebuild is targeting, and its scores against this same test set will be published here before launch. Our target: 80%. The rerun goes up here before 1.5 launches — whatever the number is.
Where it misses
Two real misses from the live test, identical across all three runs:
- “What is the latest news about electric vehicles?” went to the explainer instead of the research model — the words “what is” and “latest news” pull in opposite directions, and it guessed wrong.
- “Explain this contract clause in plain English” went to the explainer instead of the document model.
And one the fallback fixed: “What’s 15% of 240?” has no keywords for the classifier to grab, but the production fallback sent it straight to the math model all three runs. That’s the system working as designed.
How it gets better
When we catch a miss — in our own testing or from someone reporting a bad route — that question joins the test set, labeled with the model that should have handled it. We re-run the full set before every release: if a change drops the score, it doesn’t ship. Over time the test set becomes a record of every routing mistake we’ve ever found, and beating it is what each new version has to do.
1 This is our own test set, not an industry-standard benchmark.
2 Tested against a snapshot of the keyword classifier alone, the same set scores 62.6%. The live number is higher because the production fallback — a small LLM that handles questions the keywords can’t — rescues about 11 points.