A benchmark that tests whether AI models challenge nonsensical, ill-posed, or logically flawed prompts instead of confidently generating incorrect answers. Measures the critical ability to push back on bad input.
BenchLM mirrors the published BullshitBench v2 leaderboard using the official snapshot generated on June 1, 2026 at 5:53 AM UTC. The public view reports per-model clear-pushback rates across 100 nonsense prompts, scored by a 3-judge panel.
BullshitBench is a useful reasoning sanity check, but BenchLM currently keeps it display only rather than weighted. The public leaderboard is highly variant-specific and exposes reasoning-effort settings directly, so BenchLM treats it as a mirrored external benchmark instead of a canonical ranking input.
BenchLM mirrors the published clear pushback rate view for BullshitBench v2. Claude Opus 4.8 (none) leads the public snapshot at 95% , followed by Claude Opus 4.8 (xhigh) (94%) and Claude Sonnet 4.6 (high) (91%). BenchLM does not use these results to rank models overall.
Claude Opus 4.8 (none)
Anthropic
anthropic/claude-opus-4.8@reasoning=none
Claude Opus 4.8 (xhigh)
Anthropic
anthropic/claude-opus-4.8@reasoning=xhigh
Claude Sonnet 4.6 (high)
Anthropic
anthropic/claude-sonnet-4.6@reasoning=high
The published BullshitBench v2 snapshot is tightly clustered at the top: Claude Opus 4.8 (none) sits at 95%, while the third row is only 4.0 points behind. The broader top-10 spread is 16.0 points, so the benchmark still separates strong models even when the leaders cluster.
164 models have been evaluated on BullshitBench v2. The benchmark falls in the Reasoning category. This category carries a 17% weight in BenchLM.ai's overall scoring system. BullshitBench v2 is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
Year
2025
Tasks
Nonsensical and flawed prompts across multiple domains
Format
Prompt challenge and refusal evaluation
Difficulty
Robustness and critical reasoning
BullshitBench evaluates a crucial real-world capability: knowing when NOT to answer. Models that score highly recognize flawed premises, impossible physics scenarios, and logical contradictions rather than hallucinating plausible-sounding responses. V2 includes harder and more diverse challenge categories.
Version
BullshitBench v2 2025
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
A benchmark that tests whether AI models challenge nonsensical, ill-posed, or logically flawed prompts instead of confidently generating incorrect answers. Measures the critical ability to push back on bad input.
Claude Opus 4.8 (none) currently leads the published BullshitBench v2 snapshot with 95% clear pushback rate. BenchLM shows this benchmark for display only and does not use it in overall rankings.
164 AI models are included in BenchLM's mirrored BullshitBench v2 snapshot, based on the public leaderboard captured on June 1, 2026 at 5:53 AM UTC.
For engineers, researchers, and the plain curious — a weekly brief on new models, ranking shifts, and pricing changes.
Free. No spam. Unsubscribe anytime.