Vibeleaderboard
← All Intel
Intel / article

DeepSeek V4 Pro 0813 vs Claude Fable 5 on DeepSWE: Cost, Coding, and Routing

Source
together.ai
Date

Context

A cheaper coding model can handle the first attempt while an expensive model handles rejected patches. Together AI, which sells access to DeepSeek, reports this approach on 113 DeepSWE software tasks, with four trials per model. Trying DeepSeek V4 Pro 0813 before Anthropic's Claude Fable 5 produced a reported 82.7% solve rate at $8.28 per task, compared with Fable's official single-attempt score of 69.7% at $21.63. Comparison and results.

The improvement combines selective spending with an additional attempt on failed tasks. It depends on a test suite recognizing when to escalate. These are costs and outcomes from the configured experiment; applying the approach elsewhere requires similarly useful checks.

Read the source www.together.ai
Recommended reads
Comments

Checking sign-in…

Loading comments…