Best model/provider routes right now (beta)

Use this page to see which provider/model WebikAI would currently choose for common lanes, and why. Scores combine deterministic probes, live reliability, latency, and published costs. Every number shows sample size; methodology and limits are below.

What changed

Auto-generated from confirmed regressions and recoveries — a status page for model quality.

No confirmed change events yet. Entries appear here only when a deep-suite run confirms a regression or recovery — we never seed this log.

Scores by model

No published measurements yet — scores appear once endpoints accumulate enough pooled samples (n ≥ 150). We don't publish numbers we can't stand behind.

Methodology

Lanes. Six machine-graded suites: JSON extraction, schema output, tool-call formation (with distractor tools), deterministic coding tasks, exact-answer math, and completion of legitimate professional work. All run at temperature 0 and are graded pass/fail by code, not by a judge model.

Cadence. Live-traffic rollups hourly ($0 — real requests are the best probe); 10-item canaries and greedy-prefix divergence fingerprints every 6 hours; the full ~60-item deep suite (3 repetitions) weekly per endpoint and immediately when an alarm trips. Only a deep-suite run confirmed by a two-proportion test (α = 0.05) with at least a 5-percentage-point drop flags a regression — statistical significance alone never flags.

Scores. Every published number is a pass rate with its Wilson 95% interval and sample size; the headline is the lower bound. Rows with pooled n < 150 say “insufficient data” instead of a number.

The over-refusal lane measures completion of legitimate professional work only; the system never measures or ranks willingness to produce harmful content, and no probe content may be harmful.

Honest detection limits

One-sample test vs a known baseline at p₀ ≈ 0.85, α = .05, power .8.

Drop sizen neededAchieved byTime-to-detect
≥15 pp~50one day of canaries (40) + any T36–24 h
~10 pp~115a single T3 run (180)≤ 1 week, or hours if alarmed
~5 pp~450pooled week (canaries 280 + T3 180)~1 week
≤3 pp~1200+outside the public alert thresholdnot published as an alert
Read these numbers with care. Probes are sampled, not exhaustive. Scores describe provider-endpoint behavior at measurement time — not intrinsic model quality — and can vary with region, request parameters, and date. The suite focuses on agency and developer tasks (structured output, tool calls, coding); a score here may not transfer to your workload. Transport and auth failures on our side are excluded from pass rates and can never flag a provider.
Fair representation. Measurements are point-in-time observations of public endpoints through our own accounts, with timestamps, sample sizes and confidence intervals. Providers can contact us to reproduce any result; item pools are private to prevent gaming; we report measurements, not intent.
Current public scope. The public score covers the machine-graded lanes listed above and route leaders derived from those measurements plus live reliability, latency, and price. It does not score long-context retrieval, sandboxed execution of generated code, subjective writing quality, or human preference. Treat the page as a routing signal for supported agency/dev workloads, not as a universal model leaderboard.

Build on the data: /public/benchmarks · /public/benchmarks/events · /public/benchmarks/leaders