Frontier AI Model Benchmarks

How to read this table

• Benchmark scores are as each provider published them, on the provider's announcement, model card or documentation. Providers choose their own settings (reasoning effort, tools, number of attempts), so a gap of a point or two between providers isn't meaningful. The small grey note under a score gives the setting or test version when the source states it.
• The hallucination rate comes from one independent test that runs every model the same way: Vectara's hallucination leaderboard, last updated October 22, 2026. Each model summarizes the same set of more than 7,700 articles using only the facts in each article, and Vectara's HHEM-2.3 model checks every summary against its source. Lower is better.
• Prices and context windows are from each provider's own pricing and model pages, for its own API, at standard rates unless the grey note says otherwise. Open-weight models run elsewhere are priced by whoever hosts them.
• Every figure was read on October 23, 2026. We don't estimate missing numbers; a dash means we found no published figure that matches the column. Newer models are often scored on newer tests, so the classic benchmarks have gaps. This update replaces our March 2026 snapshot.

What each benchmark measures

Intelligence Index

A combined score from 10 evaluations that Artificial Analysis runs on every model under identical conditions, including Terminal-Bench 4.0, SciCode and Humanity's Last Exam. Higher is better, and it's the most like-for-like comparison in this table.

Hallucination rate

The share of a model's summaries that contain something the source document doesn't support, on Vectara's test of more than 7,700 articles, judged by Vectara's HHEM-2.3 model. Lower is better.

SWE-bench Pro

Scale AI's set of 1,865 real coding tasks from 41 professional code repositories; 731 of the tasks are public. A task counts as solved only when the model's code change fixes the problem without breaking existing functionality.

Humanity's Last Exam

2,500 hard questions across more than a hundred subjects, from the Center for AI Safety and Scale AI. Scores here are without web search or code tools.

GPQA Diamond

198 multiple-choice questions in biology, physics and chemistry, written by domain experts and designed so that a web search doesn't give away the answer. Many models in this table score above 90%, so it no longer separates the top models well.

From benchmarks to production

Benchmark scores help you shortlist models; they don't predict how a model behaves on your data. The models in this table that Vectara has tested produced unsupported statements in 5.1% to 12.9% of their summaries on its test, and rates on your own documents and prompts can be higher or lower.

1. Shortlist two or three models on the scores, price and context window that matter for your task.
2. Test them on your own data and prompts, and measure hallucinations, speed and cost at realistic volume.
3. Check outputs at runtime before anyone relies on them, and keep monitoring as providers ship updates.