aiqbench · LLM benchmarks

Benchmarks

Per-axis LLM benchmarks. One leaderboard, one number, one source — no fabricated composites.

Sources: 8
Entries tracked: 22,318
Last sync: 44 min44 minutes ago

All sources syncing normally

Arena · Text· 1 hour ago
Arena · WebDev· 1 hour ago
Arena · Vision· 1 hour ago
Arena · Agent· 1 hour ago
GPQA Diamond· 44 minutes ago
SWE-bench· 44 minutes ago
Terminal-Bench· 44 minutes ago
ARC-AGI-2· 44 minutes ago

Arena · Text· 1 hour ago
Arena · WebDev· 1 hour ago
Arena · Vision· 1 hour ago
Arena · Agent· 1 hour ago
GPQA Diamond· 44 minutes ago
SWE-bench· 44 minutes ago
Terminal-Bench· 44 minutes ago
ARC-AGI-2· 44 minutes ago

Filter this table

Jump to a model's profile

Filter by org:

Released:AfterBefore

Opus 5.5

64.8%±3.1%

61.7%CI68.0%

Anthropicswing #97–#97

Opus 5

51.8%±3.4%

48.4%CI55.2%

Anthropicswing #218–#218

One grid, every column. Pick axes in the Axes panel — one axis ranks like a single board, several blend into a weighted composite, all of them read as the index. Rank is assigned before any filter and never renumbers.

Axes1 on

Ranking on 1 axis — single-board order, no blending.

same = ranks models almost identically to an axis you already selected · close = nearly so · ? = too few shared models to tell.

Overall

Ability

Industry

Arena boards

Measured benchmarks

Agentic

Safety & factuality

Language

	Model	Org				± CI	Swing	Add to compare
1	Opus 5.5	Anthropic	—	—	64.8%	61.7%–68.0%	—
6	Opus 5	Anthropic	—	—	51.8%	48.4%–55.2%	—

Showing Terminal-Bench 4.0 exactly as the source published it — its own score and confidence interval, no blending. Select a second axis to build a weighted composite.

Opus 5.5

Opus 5

The few things in AI and quantum that actually mattered this week — honestly summarised.