Benchmarks
Six benchmarks, six ways of asking a question. Each one lists its source, its last update and what it actually tells you — including what it cannot tell you.
AA Intelligence Index v5
Intelligence index · source Artificial Analysis · 2026-10-27
A weighted composite of 10 independent evaluations covering knowledge, reasoning, coding, agents and long-horizon tasks. v5 went live on 27 October with a stricter scale than v4.3.2; scores from the two versions are not comparable.
GPQA Diamond
Graduate-level science reasoning · source Epoch AI / Metatext · 2026-11
448 multiple-choice questions written and cross-verified by PhDs in physics, chemistry and biology. The answers cannot be found by searching, which makes it the most contamination-resistant knowledge benchmark available.
Terminal-Bench 4.0
Terminal-agent tasks · source Stanford / Laude Institute · 2026-10
Software-engineering tasks inside a real container: install the environment, run the tests, read the errors, fix the code. Every model runs the same mini-SWE-agent scaffold so scaffolding is not a variable.
LiveBench Design
Document generation (DOCX/PPT/PDF) · source LiveBench · 2026-10
The model must produce Word, PowerPoint and PDF files that actually open, scored on structural integrity, formatting correctness and content organisation.
Vals Index
Vals composite index · source Vals AI · 2026-10
A weighted score across vertical professional tasks — legal, finance, tax, biology, Excel modelling — all run by Vals itself, with no vendor-supplied numbers.
AA Intelligence Index v4.3.2
Intelligence index (legacy) · source Artificial Analysis · 2026-09
The pre-v5 version, which covers more models including discontinued ones. Historical reference only; never mixed into v5 columns.
Why not a single leaderboard
Put the per-benchmark rankings side by side and you find that Almost no model leads on every dimension. Claude Opus 5.5 tops the composite, yet Gemini 4 Argon beats it on GPQA and Gemini 4 Carbon matches it on Terminal-Bench 4.0; DeepSeek V4.1 Flash scores half of the leader on the composite while matching flagships on document generation.
The awkward part is that The benchmarks are not strongly correlated. A high GPQA does not mean it can change code — Gemini 3.8 Flash scores 95.4, level with GPT-6.1 Sol, but nobody would give it a long-horizon agent task. Conversely MiniMax M3 scores 90.9 on GPQA and 1.0% on Terminal-Bench.
So this site does the following:use the composite to read the overall level; for an actual decision, go back to the per-benchmark tables. If your workload is support Q&A, GPQA is barely relevant; if it is fixing bugs in CI, the composite is useless — look only at Terminal-Bench.
How to use this
Decide the use case first, then read a ranking
Split the work into knowledge Q&A, structured generation and multi-step tool use — GPQA, LiveBench Design and Terminal-Bench respectively. Reading the wrong ranking is worse than reading none.
Watch the sampling bias
New models and open models have far more missing data. A composite built from 3 results is not comparable with one built from 10.
Look at cost per task, not price per token
Two models at the same price can differ in output tokens by two orders of magnitude. Opus 5.5 costs twice as much per token as Sonnet 5.5 but only 10% more per task.