After 30 days of real traffic, each workload is replayed against cheaper models in the background. Changes that clear the bar land here for a person to approve. Nothing moves without a yes unless the org is in Auto-route mode.
26% of requests score below 0.45 difficulty and get the same accepted outcome from Claude Sonnet 5. Jev sends only the hardest 18% to Claude Fable 5.1.
57% of requests score below 0.40 difficulty and get the same accepted outcome from GPT-5.6 Luna. Jev sends only the hardest 5% to GPT-5.6 Sol.
Qwen3.8 27B matched 98.1% of Gemini 3.1 Pro's outputs on field-level accuracy across 4K shadow samples. It's open-weight, so the router buys from whichever seller is cheapest right now.
Qwen3.8 27B matched 98.9% of GPT-5.6 Sol's outputs on rep edit distance across 14K shadow samples. It's open-weight, so the router buys from whichever seller is cheapest right now.
46% of requests score below 0.40 difficulty and get the same accepted outcome from MiniMax M3 (via OR). Jev sends only the hardest 5% to Claude Sonnet 5.