Model · Anthropic

Claude Fable 5.2

2 bench decks · 4/4 problems correct on canonical boards · 4 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.

Board summarybars = share of each board's best model · numbers = bench-native score
Mega
31.01x1/1
CUDA
34.1%3/3
Megaclaude
1/1 pass1 audited

RTX PRO 6000· canonical board

Kimi-Linear Decodepass
clean
31.01xfull-model speedup vs torch
5850 tok/s2048 ctx 32.04x8192 ctx 29.54x16384 ctx 31.52xptx

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.

RTX PRO 6000 Blackwell · Kimi-Linear Decodeclean
31.01xpublishable

Re-graded 2026-11-16 under grader v3 on tetra: 31.01x (31.41x before; see regrade_v3). Original audit: Fable 5.2 wrote one persistent 188-block kernel that runs the whole Kimi-Linear decode token: three KDA layers, absorbed MLA and the 64-expert MoE. It runs the int4 GEMVs on the integer tensor cores (mma m16n8k32 u8 x s8) by holding each activation as 16-bit fixed point with a scale per 64-element slab, split into two signed base-256 digit planes. The two products recombine exactly, so the math is W4A16 with exact integer accumulation, not an 8-bit activation shortcut. Attention weights are kept L2-resident and experts are streamed evict-first, with a speculative expert prefetch at short contexts. The agent itself warned that its 35x reading was inflated by GPU sharing and that the idle-GPU number is about 31x; the regrade agrees.

audited run
CUDAclaude · high
3/3 pass3 audited

RTX PRO 6000· canonical board

GLM-5.3 Fused MoEpass
10.8%clean
session 2h 12m
DeepSeek NSApass
0.117 msclean
session 4h 4m
MegaQwen Decodepass
6.7%interesting
session 4h 35m

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.

RTX_PRO_6000 · GLM-5.3 Fused MoEclean
10.85%publishable

Fable 5.2 treated the MoE layer as a weight-streaming problem. One prep CTA counting-sorts (token, slot) pairs by expert and builds balanced work lists; persistent CTAs then stream each expert's weights from DRAM once with warp-wide cp.async and bf16 mma.sync, fusing SiLU(gate) * up * route weight into the first GEMM's epilogue, and a combine kernel sums shared plus top-k outputs in fp32. The warp grid is picked per expert from its token count.

audited run
RTX_PRO_6000 · DeepSeek NSAclean
0.117 mspublishable

Fable 5.2 wrote three launches per forward in raw CUDA with inline-PTX bf16 mma. k_prep builds compensated 8-key prefix sums of K with a hi + lo bf16 split for tensor-core block scoring; k_select picks top-9 by tournament and re-evaluates every candidate near the 8th/9th boundary in error-free fp32, so the selected set equals the exact top-8 with the reference's tie rule. Long sequences use persistent workers with an in-kernel ticket scheduler (dense, remote far-field pooled per key block, local window plus LSE merge); S <= 2048 uses a fully dense masked kernel driven by the selection bitmask.

audited run
RTX_PRO_6000 · MegaQwen Decodeinteresting
6.71%publishable

Fable 5.2 wrote a persistent megakernel (188 blocks x 16 warps) that runs every decode step and all layers in one launch, handing work between phases through atomic-exchange publishes and phase counters instead of grid.sync. Weights and KV feed mma.sync directly in fragment order while each fp32 partner vector is split exactly into three bf16 pieces, so the math stays bf16 x fp32 with fp32 accumulation. For long contexts, completed 16-position KV chunks get a lossless 12-bit packed mirror with a per-fragment exponent base and raw fallback, cutting DRAM bytes on the attention read. During the run it found and fixed a real atomic-publish race in its own phase handoff.

audited run