About rank

Rank brings publicly reported model performance together across reasoning, coding, tool use, research, and everyday tasks. Results come from model provider reports, original benchmark pages, and independent evaluation pages such as Artificial Analysis. Rank collects published findings; it does not run these benchmarks locally.

Each result keeps its source, benchmark version, and available evaluation settings, including tool access, model effort, and question subsets. Different versions and agent setups remain separate in the dataset.

Explore the models, providers, or benchmarks to see the reported data.

I update the leaderboard with every new model release and whenever there’s a significant change to published benchmark results, pricing, or model specifications. Updates are manually reviewed and added to the dataset. Page update dates track changes to the published content. “Latest data recorded” refers to the newest stored data entry used by that page. Source review dates are shown only when an explicit review is recorded; a new deployment does not refresh those dates.

Summaries are generated from the stored data. Benchmark leaders and comparison wins describe reported scores within the displayed version and setup, rather than a guarantee of performance on every task. The Agent Index is recalculated after relevant reviewed data updates. Each calculation stores a snapshot of the model scores and the exact benchmark reports used.

Agent Index measures how well AI models can complete practical work as agents. It blends reported results for reasoning, document understanding, professional deliverables, multi-step workflows, coding, terminal use, and browser interaction into a score from 0 to 100.

Cost, latency, and generation speed are separate. This preview measures capability across the selected tasks; it is not a measure of value for money or a complete assessment of every type of agent work.

Calculated November 2, 2026. 26 of 86 models currently have all 7 required results. Models without complete coverage are marked “Insufficient coverage” and receive no overall score or rank. An unranked model has not been shown to perform worse.

These are editorial weights, chosen to emphasize practical work. Benchmarks can measure overlapping skills, and three components come from LiveBench. The index is a transparent starting point, not a statistically validated predictor of success on your own tasks.

Each benchmark is normalized using fixed low and high reference scores from the 12 models with complete coverage on November 2, 2026. The low reference maps to 0 and the high reference maps to 100. Scores outside those bounds are capped at 0 or 100.

The original reference points are retained, and the browser reference points are frozen in v1.2, so adding a model does not rescale existing scores. A change to benchmarks, weights, or reference points requires a new index version. This is a relative capability score; 90 points does not mean a model completes 90% of all tasks.

Every component pins its preferred benchmark version and evaluation group, with explicitly listed fallback tracks for LiveBench. HLE-Diamond uses the revised 1,000-question exam without tools; GDP.pdf uses All-pass across five attempts; Briefcase uses its v1.1 combined Elo. LiveBench Agentic, Coding and Browser Use prefer the October 25, 2026 release. If a model has no result in that release, v1.2 uses its verified October 14 result. The score breakdown labels older-release inputs; question sets differ, so these are broader reported comparisons rather than same-release measurements. Browser Use tests screenshot-based interaction with web apps; a published zero for a model that cannot accept screenshots counts as zero rather than missing data. This is distinct from unrestricted text or DOM-based browser agents. Terminal-Bench 4.0 uses 66 tasks with the shared mini-swe-agent harness and three repeats.

We prioritize evaluations of agents working with tools when comparable results are available. The current v1.2 still uses HLE-Diamond without tools for its reasoning component. We do not substitute a different exam or mix tool-enabled and tool-free scores within that component; changing its protocol requires a new index version.

We always prefer the primary release, even when an older-release score is higher. Within each selected track, we use the latest published report date, falling back to its retrieval date when no publication date is available, then the best reported score on that date. Exact report IDs and settings are stored with every component. Effort can differ between models, and we do not mix native-agent terminal results into the shared-harness component.

LiveBench uses one run per suite, and some suites have relatively few tasks. The preview does not estimate uncertainty or adjust for differences in tool budgets and reasoning effort. Personal-assistance and broader computer-use performance are not yet represented by dedicated components.

Scores are stored to six decimal places and displayed to one. Ranking uses the stored score, so displayed values can tie while ranks differ; exact stored ties share a rank.

View the complete index data and model score breakdowns ↗

I curate rank to make reported model performance easier to explore and compare across reasoning, coding, tool use, research, and everyday tasks.

My work at Vellum has spanned growth, AI education, and helping people understand how to evaluate AI systems. I’ve published practical model comparisons and guides to LLM benchmarks and evaluation frameworks, and spoken at the AI Engineer Summit about building reliable agents with test-driven development.

I bring that focus on practical evaluation to this leaderboard: understanding what a benchmark measures and how a reported result translates to the work you want a model to do.

Follow @anitakirkovska on X / Twitter for my work on AI agents, evaluations, and how we use these models.

Want to add a model or provider, share a new result, or flag something that isn’t correct? Send me a DM on X. A link to the original report helps me review and update it.