Discera 1.5 · Router benchmark
Discera is one chat that routes every message to the AI best suited to it — so the router’s accuracy is the product’s accuracy. When we published the 1.0 router benchmark, we set the rebuilt 1.5 router a public target of 80% and promised the rerun would be published whatever the number was. This is the rerun.
The result
This time we graded harder. Instead of one question set, the rebuilt 1.5 router faced three: the original 147 questions Discera 1.0 was graded on, plus two completely new sets of 147 written after the rebuild — no question repeated across sets, so the router couldn’t pass by memorizing the test. Each set ran three times: 1,323 routing decisions in total, graded with the same rubric published in the 1.0 benchmark.
94.1% average across the three sets (original set: 93.2%, new sets: 93.9% and 95.2%)
On the original set, the router scored 78.2% right before the final round of routing fixes — and Discera 1.0’s published score on that set was 73.7%. The target we set in the 1.0 benchmark was 80%. The rebuilt router cleared it by 14.1 points.
By category, on the newest set (October 26, 2026 — every question routed identically in all three runs):
Where 1.5 still misses
Nine of twelve categories scored 100% on the newest set. The remaining misses concentrate in two places:
- Audio generation. Requests for things Discera doesn’t do are mostly refused correctly — video generation was caught every time. But audio generation (“record a birthday song in a robot voice”, “produce a ringtone from this melody”) has no category to land in: Discera 1.5 doesn’t generate audio, and those requests make up most of the remaining misses. We count them as misses rather than quietly removing them from the test.
- Concept definitions at category boundaries. “Explain what a power of attorney is” went to the general explainer instead of the document model, and “Explain the difference between mass and weight” went to the explainer instead of the math and science model. Same shape as 1.0’s worst misses — definitions phrased as “explain…” — now down to a couple of questions per set instead of a pattern.
How it keeps getting better
The same discipline as 1.0: when we catch a miss, that question joins the test set, labeled with the model that should have handled it, and we re-run the full set before every release — if a change drops the score, it doesn’t ship. The full test set, the grading rubric, and the 1.0 results are in the Discera 1.0 router benchmark.