Model evidence

DeepSeek V4

Keep Flash, Pro, and their effort levels separate while comparing them inside one published batch.

Current batch

DeepSeek V4 Flash / High is both the highest-scoring measured configuration at 66/100 and the fastest in this family at 21m 51s in the selected batch.

Open the complete Radar Read the evaluation method

Selected evidence

Published: Aug 26, 2026, 11:00 PM UTC
Question pack: coding-fast-v4.10
Coverage: 2 measured configurations
Batch: snapshot-2026-08-26T23-00-00Z-r3

Measured configurations

Configuration	Rank	Score	Elapsed	Reference cost	Route
DeepSeek V4 Flash / High	#11	66/100	21m 51s	$0.12	Custom endpoint
DeepSeek V4 Pro / High	#14	64/100	23m 34s	$0.32	Custom endpoint

Five-question profile

Question	DeepSeek V4 Flash / High	DeepSeek V4 Pro / High
Black-box Regression Audit	10/20	12/20
Retry Planner Counterexamples	14/20	13/20
CI Adversarial Audit	13/20	12/20
Transaction Regression Design	14/20	14/20
Cache Regression Test Design	15/20	13/20

Flash and Pro are not interchangeable rows

DeepSeek V4 Flash and Pro are distinct model identities. Each effort level is another configuration. Combining them into one DeepSeek score would hide which variant produced the result.

Use the table to find whether quality, elapsed time, and cost point to the same configuration. When they do not, apply a task-specific quality floor before optimizing the other two.