Model · Anthropic
Claude Opus 5.5
Claude Opus 5.52 bench decks · 5/5 problems correct on canonical boards · 5 audited cells.
methodology + notes
How to read. Cell scores are peak fraction of the board roofline (Hard / most CUDA), geomean milliseconds for CUDA Native Sparse Attention, or speedup vs the torch baseline (Mega). Each cell has an unlimited agent budget; interrupted or overlapping resumes are documented in the audit. Audit chips come from the human/subagent reward-hack review of every published cell; scores from flagged sessions render dimmed — they don't count toward the charts.
Board summary bars are each score relative to the best published model on that board (1.00 = board leader); the printed number is the bench-native score.
Methodology. Rank per bench: valid passes (audited-clean correct cells / problems) desc, Cells with no audit annotation are shown but never scored. then mean normalized performance over the FULL problem deck (cell score / board best per problem; fail/invalid/missing cells count as 0) desc. Hack badge = flagged audited sessions / total audited sessions for that model; flagged = annotation verdict reward_hack | contamination | rubric_leak | suspect, 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.
RTX PRO 6000· canonical board
all audited claude attempts
Every attempted cell stays visible, including correctness failures, hardware mismatches, contaminated runs, and audit rejects. Only publishable results contribute to the board above.
Claude Opus 5.5 built a genuine single-launch cooperative megakernel for the Kimi-Linear decode block. torch.profiler shows 1.00 launches per step. The 35.46x isolated grade is real and physically plausible. At 0.159-0.187 ms per token it streams the 243-259 MB each token needs at an effective 1.39-1.53 TB/s against the card's 1.79 TB/s peak. The speed comes from memory-pipeline engineering, not a shortcut. The same-buffer token and cache overwrite probes, a fresh-state step, a 6-step chain, and a full-state comparison all track the reference. Templates are untouched and nothing outside the run was read. Correctness caveat: the fp32-atomic GEMVs are not bit-identical to the reference, so near-tie router picks flip. That gives 0.9964 on graded seed 2 and 0.9607 (a fail) on ungraded seed 3. baseline.py has the same class of failure, and the agent disclosed the rate itself.
audited runRTX PRO 6000· canonical board
all audited claude attempts
Every attempted cell stays visible, including correctness failures, hardware mismatches, contaminated runs, and audit rejects. Only publishable results contribute to the board above.
Opus 5.5 wrote a CUDA grouped-MoE path: count and scatter routed tokens, run two hand-written bf16 tensor-core GEMMs, fuse SiLU into the first epilogue, and accumulate the routed and shared expert outputs in fp32. The final kernel passed the corrected full-shape check and an isolated regrade on tetra.
audited runOpus 5.5 built block selection and sparse attention in CUDA, using tensor cores for block scores and attention and a separate window merge. The final kernel passes the revised graded-shape check and isolated regrade. An operator launch mistake ran two copies of its resumed session in one workspace, so its optimization trajectory is not a single-agent trace.
audited runOpus 5.5 wrote a cooperative CUDA decode kernel that runs every step and all four layers inside one launch. It packs weights once per model version, streams new noise inputs during the launch, and reads the full growing KV cache. The final source passes the long-context check and isolated regrade; two operator-resumed copies collaborated in the same workspace.
audited runOpus 5.5 fused the three-layer MinGRU policy and grid rollout in one cooperative CUDA kernel. It folded the linear encoder into the first gate projection, uses three fp16 tensor-core products to retain fp32 accuracy, and reproduces the reference's globally coupled food RNG. The final source passes the exact-position and numeric checks; a second resumed session touched scratch work during development.
audited run