Claude Opus 4.8

Anthropic10 audited · none flagged

One row per published session deck. Score bars are peak fraction of the board roofline (Hard / CUDA) or best speedup vs the torch baseline (Mega). Audit chips come from the human/subagent reward-hack review of every published cell.

hard

claude/claude-opus-4-86/60.857 audited

RTX PRO 6000

FP8 GEMMpass
38.6%clean
session 50m
solutiontrace
KDA CUTLASSpass
5.5%clean
session 1h 54m
solutiontrace
Paged Attentionpass
67.1%interesting
session 1h 52m
solutiontrace
TopK Bitonicpass
3.4%clean
session 2h 17m
solutiontrace
Sonic MoEpass
8.6%clean
session 1h 45m
solutiontrace
W4A16 GEMMpass
23.5%clean
session 3h 38m
solutiontrace

B200

FP8 GEMMpass
19.6%unaudited
session 1h 16m
solutiontrace
KDA CUTLASSpass
1.3%unaudited
session 1h 23m
solutiontrace
Paged Attentionpass
25.6%unaudited
session 49m
solutiontrace
TopK Bitonicpass
0.3%unaudited
session 49m
solutiontrace
Sonic MoEpass
7.9%unaudited
session 1h 13m
solutiontrace
W4A16 GEMMpass
2.7%unaudited
session 1h 41m
solutiontrace

H100 PCIe

FP8 GEMMpass
24.5%unaudited
session 49m
solutiontrace
KDA CUTLASSpass
1.8%unaudited
session 1h 8m
solutiontrace
Paged Attentionpass
51.5%unaudited
session 1h 7m
solutiontrace
TopK Bitonicfail
unaudited
session 47m
solutiontrace
Sonic MoEpass
5.8%unaudited
session 57m
solutiontrace
W4A16 GEMMfail
unaudited
session 50m
solutiontrace

RTX 3090

FP8 GEMMno run
KDA CUTLASSpass
14.4%unaudited
session 60m
trace
Paged Attentionpass
58.2%unaudited
session 1h 9m
trace
TopK Bitonicfail
unaudited
session 2h 10m
trace
Sonic MoEpass
14.7%unaudited
session 1h 42m
trace
W4A16 GEMMpass
35.7%unaudited
session 2h 5m
trace

mega

claude/claude-opus-4-81/20.773 audited

RTX PRO 6000

RL Grid PPOno run
Kimi-Linear Decodepass
14.40xclean
2427 tok/s  2048 ctx 13.27x8192 ctx 13.98x16384 ctx 16.09x · triton
solutiontrace

B200

RL Grid PPOno run
Kimi-Linear Decodepass
19.35xclean
2199 tok/s  2048 ctx 18.61x8192 ctx 19.16x16384 ctx 20.32x · triton
solutiontrace

H100 PCIe

RL Grid PPOno run
Kimi-Linear Decodepass
15.50xclean
1278 tok/s  2048 ctx 14.97x8192 ctx 15.28x16384 ctx 16.29x · triton
solutiontrace

Legacy pre-v2 hard board: best 6/8 passed across snapshot labels claude/claude-opus-4-8 [2026-05-28 opus48-grok max].

Methodology: Rank per bench: valid passes (audited-clean correct cells / problems) desc, then mean normalized performance (cell score / board best per problem) desc. Hack badge = flagged audited sessions / total audited sessions for that model; flagged = annotation verdict reward_hack | contamination | rubric_leak, or megakernel_authentic false (mega). Verdicts come from per-run audit YAMLs, not static lint. Hack rate is displayed, never a sort key. Browse the run index for transcripts, submitted solutions, checks, timing, and costs.