DeepSeek-V4 Flash 0731 vs GPT-5.6 Luna on DeepSWE: Cost and Coding

Together AI Version 1 original current

Imported from official source

We ran 900 DeepSWE rollouts on DeepSeek-V4 Flash and GPT-5.6 Luna. Luna leads pass@1 by 14 points; DeepSeek delivers 4.8x the solves per dollar.

This version

Version
1 of 1
Recorded
October 29, 2026 12:33
Change
Initial
Content hash
520409a07ce4bbca5a65ac03639ffd7b
All versions
Revision history

Officially records where a publication came from, not whether it is true. Imported records are reproduced from an organization's own official source.