BREAKING
DeepSeek V4 matches Sonnet 4.6
0%
before
0%
after
0
criteria
Same score, one-seventh the cost
DeepSeek V4 ProOpen-weight
5% all-pass
~1/7 cost
Claude Sonnet 4.6Closed
~5% all-pass
Baseline cost
Evolve the harness, not the model
1Propose mechanism
2Verify effect
3Adopt or reject
File handling was the blind spot
Harness, Loop, Memory, Evals
AI NEWS BLITZ
An open model reached frontier scores on a legal benchmark with no weight changes.