Results are ordered by score in the display harness, followed by models with no result. Other harnesses are listed separately below. Missing results remain unknown.
Table8.1.A and section8.12.1: full2500-question HLE, reasoning-only no-tools, adaptive/auto thinking, maximum effort,1M token cap, no compaction, Opus4.6 grader; trials averaged. Production safeguards may route flagged Fable tasks to Opus models; this is disclosed deployed-model behavior, not an unsafeguarded base-model score.
Table8.1.A and section8.12.1: full2500-question HLE, reasoning-only no-tools, adaptive/auto thinking, maximum effort,1M token cap, no compaction, Opus4.6 grader; trials averaged. Production safeguards may route flagged Fable tasks to Opus models; this is disclosed deployed-model behavior, not an unsafeguarded base-model score.
Cited Qwen page returned no readable content in this audit; exact model/version, value and protocol could not be independently verified from that source.
Fixture value corresponds to source BF16 column on the NVFP4 repository page; preserve checkpoint/precision and per-benchmark NeMo configuration before mapping to a preferred unit.
Same full HLE with web search, restricted web fetch, programmatic tools and code execution. Kept separate from no-tools model capability; production safeguard fallback applies.
Same full HLE with web search, restricted web fetch, programmatic tools and code execution. Kept separate from no-tools model capability; production safeguard fallback applies.