Same model name, different answers.
DeepSeek V4 Pro: outside the reference
On DeepSeek V4 Pro, 4 of 16 endpoints sit outside the reference: gmicloud/fp8, novita/fp8, crusoe/fp8, cloudflare. Ranked endpoints run 50 to 65% on string match.
String match, mean over every night listed.
Cite: Bitdepth · bitdepth.dev/m/deepseek-v4-pro · as of Nov 11, 2026
venice/int4, night by night
On gpt-oss-120b, venice/int4 matched the reference on 95% of prompts over its first three nights and 83% over its last three. It now sits outside the reference.
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Price against agreement
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Agreement night by night
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Price against agreement
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Agreement night by night
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Price against agreement
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Agreement night by night
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Price against agreement
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Agreement night by night
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Price against agreement
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Agreement night by night
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Price against agreement
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Agreement night by night
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Price against agreement
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Agreement night by night
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Declared quantization against measured state
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Moved this period
Mean exact match of the last three nights against the first three.
- venice/int4 on gpt-oss-120b: -11.4 points
Cite: Bitdepth · bitdepth.dev/charts · as of Nov 11, 2026
Speak the language.
Endpoint
One provider serving one model at one declared precision on OpenRouter.
Declared quantization
The precision the provider lists: bf16 or fp16 is full precision; fp8, fp4 and int4 are smaller and cheaper to run.
Reference
The answer the full-precision endpoints give. What every endpoint is graded against.
Agreement
The share of prompts where an endpoint gives the reference answer.
Quality-adjusted price
Blended price divided by agreement. What a correct answer costs.
Drift
An endpoint whose answers move over the period, measured against its own first nights.