How this request was routed

Code debugging · completed in 14.2 sIllustrative data

1 · UnderstandCode debugging, high complexityJev is 86% confident · needs tool calls
2 · Shortlist4 of 7 models qualifyFiltered on public benchmarks and capabilities
3 · Auction3 valid sealed bidsWithin 300 ms · 1 late bid rejected
4 · Routegemini-pro · GoogleBest score for the quality-first policy

Why gemini-pro won

claude-sonnet scored higher on quality, but gemini-pro bid cheaper and responds faster — enough to come out ahead with priorities set to quality 75, cost 15, speed 10.

gemini-pro0.775

claude-sonnet0.758

gpt-mini0.730

utility = 0.75 × quality + 0.15 × cost + 0.1 × speed − uncertainty penalty

Bid	Quality (× 0.75)	Cost (× 0.15)	Speed (× 0.1)	Utility
gemini-pro	0.78 → 0.585	0.80 → 0.120	0.70 → 0.070	0.775
claude-sonnet	0.82 → 0.615	0.55 → 0.083	0.60 → 0.060	0.758
gpt-mini	0.74 → 0.555	0.90 → 0.135	0.40 → 0.040	0.730

Quality is a normalized composite of public benchmark results — a routing proxy, not a success probability. Cost and speed are scored against fixed bounds saved with this run, so one bid never shifts another’s score.

Model · deployment	Reason	What would change it
qwen-coder-q4 · operator vLLM	Quantized build; its benchmark scores are inherited from full-precision weights, so they count only as a proxy	Turn on “allow missing coverage” — joins with an uncertainty penalty
small-flash · Google	Coding quality 0.51, below the policy minimum of 0.60	Lower the minimum quality in the policy
legacy-chat · OpenAI-compatible	Endpoint probe found no tool-call support; this request declares a tool	Hard requirement — can’t be relaxed

Deployment	Arrived	In / out per Mtok	Strategy	Queue delay	Status
gpt-mini	84 ms	— / —	fixed rate	—	valid
claude-sonnet	142 ms	— / —	capacity-adjusted	—	valid
gemini-pro	205 ms	— / —	bounded discount	—	valid · awarded
open-coder	341 ms	— / —	fixed rate	—	rejected · arrived after deadline

Bids were sealed. Invitations carried only an auction ID, each bidder’s eligible offering IDs, the deadline and protocol version — no prompt, classification or app weights. Rates are simulated demo offers, not vendor discounts.