aiqbench · insight feed
Cross-axis insights
Editorial sentences summarising the notable cross-axis findings computed elsewhere on this site — top divergent performers, frontier-setting models, weight-sensitive rankings, and the shape of axis correlations. Every claim is a computed fact; every phrasing is verifiable in one click.
Biggest finding
granite-4.1-8b ranks between #150 and #978 across 6 built-in weightings — best under "Agentic", worst under "All sources".
- rank swinggpt-4o-2024-05-13
gpt-4o-2024-05-13 ranks between #154 and #955 across 6 built-in weightings — best under "Agentic", worst under "All sources".
Agentic · All sources - rank swinggpt-4.1-nano-2025-04-14
gpt-4.1-nano-2025-04-14 ranks between #197 and #987 across 5 built-in weightings — best under "Coding-focused", worst under "All sources".
Coding-focused · All sources - rank swinggpt-4o-mini-2024-07-18
gpt-4o-mini-2024-07-18 ranks between #236 and #1,023 across 5 built-in weightings — best under "Coding-focused", worst under "All sources".
Coding-focused · All sources - rank swinggrok-4-fast-reasoning
grok-4-fast-reasoning ranks between #157 and #936 across 6 built-in weightings — best under "Agentic", worst under "All sources".
Agentic · All sources - divergenceGemini 3.7 Flash (High)
Gemini 3.7 Flash (High) shows a 93.4-point percentile gap across 13 covered axes: 97.7 on GPQA Diamond versus 4.3 on Terminal-Bench 4.0.
GPQA Diamond · Terminal-Bench 4.0 - divergencegpt-5.4-nano-high
gpt-5.4-nano-high shows a 93-point percentile gap across 14 covered axes: 99 on HHEM Hallucination (Vectara) versus 6.1 on SimpleQA Verified.
HHEM Hallucination (Vectara) · SimpleQA Verified - frontiergpt-5.4-pro-2026-03-05
On FrontierMath, gpt-5.4-pro-2026-03-05 raised the running best score by 9.3 percentage points to 50% on 2026-03-05.
FrontierMath - divergencegpt-5.1-high
gpt-5.1-high shows a 92.1-point percentile gap across 22 covered axes: 97.3 on LiveCodeBench (Artificial Analysis) versus 5.3 on Document Arena.
LiveCodeBench (Artificial Analysis) · Document Arena - divergenceGrok 4 Fast (Reasoning)
Grok 4 Fast (Reasoning) shows a 90.6-point percentile gap across 5 covered axes: 93.5 on LiveCodeBench (Artificial Analysis) versus 2.9 on WebDev Arena.
LiveCodeBench (Artificial Analysis) · WebDev Arena - divergenceo4-mini-2025-04-16
o4-mini-2025-04-16 shows a 90.6-point percentile gap across 20 covered axes: 97.6 on LiveBench versus 7 on Factuality Arena.
LiveBench · Factuality Arena - correlation
2 correlation-linked groups emerged; the largest spans 37 axes with median pairwise r of 0.89; 5 axes remain singletons with adequate coverage but no cluster-worthy correlation; 9 axes have insufficient overlap to cluster honestly.
- frontiergpt-6-astra
On ARC-AGI-2, gpt-6-astra raised the running best score by 4.6 percentage points to 95% on 2026-09-13.
ARC-AGI-2 - frontierclaude-opus-4-7
On SWE-bench Verified, claude-opus-4-7 raised the running best score by 4.1 percentage points to 83.5% on 2026-09-13.
SWE-bench Verified - frontiergpt-5.6-sol
On OTIS Mock AIME, gpt-5.6-sol raised the running best score by 0.8 percentage points to 100% on 2026-09-13.
OTIS Mock AIME - frontiergemini-4-argon
On GPQA Diamond, gemini-4-argon raised the running best score by 0.6 percentage points to 96.4% on 2026-10-13.
GPQA Diamond - frontiergemini-4-argon
On SimpleQA Verified, gemini-4-argon raised the running best score by 0.4 percentage points to 76% on 2026-10-13.
SimpleQA Verified
How these sentences are made
The FACTS are computed deterministically from the current database — numbers, model names, axis labels. A low-temperature LLM phrases each fact into one sentence, and a validation gate rejects any output that adds a number, superlative, or causal claim the struct didn't carry (falling back to a deterministic template phrasing on any failure). The prose is a convenience over the numbers, never a replacement — every insight ships with a 'verify' link to the underlying data.
How the leaderboards work →