Discera · 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. Our earlier benchmarks (the 1.0 benchmark and the 1.5 benchmark) graded the router on sets of 147 questions. This time we wanted a test the router had never had the chance to be tuned against: 1,000 completely new questions, written for this test, each one run twice.
The results — two scores, both published
We report this benchmark two ways, because the gap between them is itself part of the result.
98.5% on the questions the router can answer (98.39% and 98.61% across the two runs)
Of the 1,000 questions, 934 ask for something the router has a category for — writing, research, math, documents, images, and the rest. Those are the “returnable” questions: ones the router is allowed to answer with a route. Across both runs — 1,868 returnable decisions — it sent 98.5% of them to the right AI.
92.0% across all 1,000 questions (91.9% and 92.1% across the two runs)
The other 66 questions asked for things the router has no category for at all — video and audio generation, mostly. In this test they score zero by construction: the router’s classifier has no “unsupported” answer it can return. In the live product those requests are caught and handled outside the router, but we count them here rather than quietly removing them from the test — so the overall score, 92.0%, is the harder, more honest denominator.
Some context for both numbers: this test scores the router’s classification step — the decision itself — using the exact router prompt that shipped in production on the morning of November 5, 2026 (the router is GPT-4o mini). The extra layers around it in the live product aren’t scored here. The two runs agreed with each other 99.3% of the time, all 2,000 decisions completed with zero errors, and the median routing decision took about 0.8 seconds.
How fresh is “fresh”?
Every question in this set was written for this test. Before a single one was run, the set was audited against all 462 questions from every earlier Discera test — the 1.0 and 1.5 benchmark sets and every tuning set since: zero exact repeats, zero near-duplicates. The router couldn’t pass this one by having seen the questions before.
By category
Across both runs combined:
Six categories were perfect on both runs: research, documents, explanations, code review, personal advice, and artistic images. Photo-style image requests and writing each missed one question per run.
Where it still misses
There were 28 returnable misses across 2,000 decisions, and they aren’t random — they concentrate in one pattern:
- “Explain…” gravity on named concepts (14 of the 28 misses). When a named science or programming concept is phrased as “explain…” or “what is…”, the router sometimes sends it to the general explainer instead of the specialist. “What is the periodic table organised by?” and a question about plate tectonics went to the explainer instead of math & science; “compiler vs interpreter” went to the explainer instead of the coding model. It’s the same residual pattern our earlier benchmarks found — narrower than it was, and still the router’s number one error.
- The general-chat boundary. General chat — the fallback for casual one-offs — scored 92–93%, the weakest returnable category. Trivial requests (a random fun fact, “say something encouraging”, a quick translation) get pulled toward a substantive category that sounds close.
- Math & science (95%) and agentic coding (97%). The next-weakest returnable categories — and most of their misses are the “explain…” pattern above, arriving from the other direction.
How this compares to the 94.1%
The 1.5 benchmark published 94.1% — and this post’s 92.0% overall is not a fall from it, because the two numbers aren’t measured the same way. The 94.1% was a full-pipeline grade — classifier, keyword fallback, and backstops together — over 147-question sets the router had effectively been developed against. This test is 1,000 fresh questions the router had never seen, scored on the classification step alone, with the unreturnable slice left in the denominator. It is a much bigger, harder, fresher test — and on the questions it can actually answer, the router held 98.5% on questions it had never seen, against 99.3% on the set it was tuned on. That small gap is what an honest fresh test is supposed to look like.
How it keeps getting better
The same discipline as always: the misses from this test — starting with the “explain…” pattern — join the labeled test sets, and routing changes get re-graded against the full sets before they ship; if a change drops the score, it doesn’t ship. The earlier results are in the Discera 1.0 router benchmark and the Discera 1.5 router benchmark.
1,000 fresh questions. Two full runs. No repeats from earlier benchmarks. That’s the number as it stands.
