Home / LLM leaderboard
LLM leaderboard
Which AI model fits the job, and what does it cost?
A searchable database of current large language models from 22 developers. Compare them on independent benchmarks, find the strongest model for each kind of AI agent, and see which models give the most capability for the money. Every score and price links to the page it came from.
Best by agent type
Each card ranks models on the independent benchmark closest to that job. A high score is a starting point for your own trial, not a guarantee on your tasks.
Agents that read a codebase, write and test changes, and work in the terminal, as in Claude Code, Codex or Cursor.
Ranked by Artificial Analysis Coding Agent Index (Artificial Analysis; board shows no update date, retrieved Nov 11, 2026). 21 of the listed models have a result. Also see: Terminal-Bench, LiveBench Coding.
Tools that turn a description into a working web page or small app, judged by the people who asked for it.
Ranked by Arena WebDev (Arena (arena.ai); results as of Nov 10, 2026). 35 of the listed models have a result. Also see: LiveBench Coding.
Agents that operate a browser for you: navigating sites, filling forms and pulling information out.
Ranked by LiveBench Browser (LiveBench; release Nov 6, 2026). 26 of the listed models have a result. Also see: Arena Agent.
Agents that search, read many sources and pull the findings together into a cited answer.
Ranked by LiveBench Research (LiveBench; release Nov 6, 2026). 26 of the listed models have a result. Also see: LiveBench Long Agentic.
Agents trusted with a goal over a long period, making many decisions without a person checking each one.
Ranked by Vending-Bench 2 (Andon Labs; board shows no update date, retrieved Nov 11, 2026). 15 of the listed models have a result. Also see: LiveBench Long Agentic, Arena Agent.
Finance, legal, tax and analysis tasks of the kind office teams hand to an AI assistant.
Ranked by Vals Index (Vals AI; results as of Nov 9, 2026). 28 of the listed models have a result. Also see: AA Intelligence Index.
Working with spreadsheets, tables, reports and images of documents, and staying faithful to the source.
Ranked by LiveBench Data Analysis (LiveBench; release Nov 6, 2026). 26 of the listed models have a result. Also see: Hallucination rate, Arena Vision.
Drafting, explaining and answering questions in conversation: what most staff use AI for day to day.
Ranked by Arena Text (Arena (arena.ai); results as of Nov 9, 2026). 32 of the listed models have a result. Also see: AA Intelligence Index.
Problems that need long, careful reasoning: engineering calculations, proofs and technical analysis.
Ranked by FrontierMath (Tiers 1-3) (Epoch AI; results as of Nov 6, 2026). 23 of the listed models have a result. Also see: GPQA Diamond, ARC-AGI-2.
Model explorer
Pick a benchmark to rank by. Switch to the chart to see the Pareto frontier: the models that no cheaper model beats.
| # | |||||
|---|---|---|---|---|---|
| Turn on JavaScript to search and rank the models. | |||||
Price is the blended list price per million tokens, weighted 3 input to 1 output. It comes from the developer's own price list; for open-weight models, and where no developer price list could be read, it is the price a named host lists on OpenRouter (shown as a secondary source in the details). Select a row for specifications and every benchmark result with its source.
Try removing a filter or searching for a broader term.
Methodology
Data as of November 12, 2026. 473 benchmark results and 48 list prices, each linked to the page it came from.
Four research agents read each benchmark's own leaderboard or data file and each developer's model and price pages. Every value was copied as published, with the address it came from.
Rankings use the results each benchmark publishes on its own board. Scores that developers report about their own models in announcements or model cards are not used. Some boards accept submitted runs; their notes say so.
Benchmarks spell model names differently. Each name is matched to one entry. Earlier snapshots, previews, quantised builds and products not sold through the API are kept apart. When a model appears at several reasoning-effort settings, it is ranked on its best setting, which is shown with the score.
Two further agents that did not collect the data re-checked values against the source pages and challenged the name matches, prices, costs and wording. Errors they found are fixed, or the value is removed, before publication.
| Benchmark | What it measures | Source |
|---|---|---|
| Artificial Analysis Intelligence Index | A combined score from nine evaluations covering knowledge work, coding, science and general knowledge, run by Artificial Analysis on each model. Cost per task is Artificial Analysis's weighted average API cost per Index task. | Artificial AnalysisIndex v5.0.1; board shows no update date, retrieved Nov 11, 2026 |
| Arena Text | Human preference: people compare two anonymous answers to their own prompts and pick the better one. Ratings on an Elo-like scale; style control adjusts for answer length and formatting. Neighbouring models often have overlapping confidence intervals; see the ± range in each model's details. | Arena (arena.ai)results as of Nov 9, 2026 |
| Epoch Capabilities Index | Epoch AI's single capability score, combining results from many benchmarks. | Epoch AIhub updated Nov 11, 2026 |
| Vals Index | Agentic tasks in finance, law, tax and software, weighted by each sector's share of US GDP, run by Vals AI. Cost per task is Vals's published cost per test. | Vals AIresults as of Nov 9, 2026 |
| Artificial Analysis Coding Agent Index | Models run inside coding-agent tools on three evaluations: FrontierCode, SWE-Atlas QnA and Terminal-Bench v4. The tool used is shown with each score; several are the developer's own tool. Cost per task is Artificial Analysis's average API cost per task. | Artificial Analysisboard shows no update date, retrieved Nov 11, 2026 |
| Terminal-Bench 4.1 | Tasks carried out in a computer terminal, run inside named agent tools. Share of trials solved. Cost here is derived: the board's total cost divided by its number of trials (cost per trial). The board does not say which results it verified. | Terminal-Bench (tbench.ai)latest result Sep 29, 2026, board retrieved Nov 11, 2026 |
| Arena WebDev | Human preference for web apps the models build from a prompt, judged side by side. | Arena (arena.ai)results as of Nov 10, 2026 |
| LiveBench Coding | LiveBench's coding suite, scored automatically from 0 to 100. A single run of 24 tasks per model, so a difference of a few points can be one task. Cost here is derived: LiveBench's total suite cost divided by its task count, as the LiveBench site itself does. | LiveBenchrelease Nov 6, 2026 |
| Arena Agent | Real agent sessions on arena.ai: how much better each model does than a baseline at finishing users' multi-step tasks, from task outcomes and user feedback. | Arena (arena.ai)results as of Nov 8, 2026 |
| LiveBench Browser | LiveBench's browser-agent suite: tasks completed by operating a web browser, scored from 0 to 100. A single run of 25 tasks per model, so a difference of a few points can be one task. Cost here is derived: LiveBench's total suite cost divided by its task count, as the LiveBench site itself does. | LiveBenchrelease Nov 6, 2026 |
| LiveBench Research | LiveBench's research suite, scored from 0 to 100. A single run of 22 tasks per model, so a difference of a few points can be one task. Cost here is derived: LiveBench's total suite cost divided by its task count, as the LiveBench site itself does. | LiveBenchrelease Nov 6, 2026 |
| LiveBench Long Agentic | LiveBench's long agentic suite: long, multi-step agent tasks, scored from 0 to 100. A single run of 18 tasks per model, so a difference of a few points can be one task. Cost here is derived: LiveBench's total suite cost divided by its task count, as the LiveBench site itself does. | LiveBenchrelease Nov 6, 2026 |
| Vending-Bench 2 | A simulated year running a vending-machine business: ordering stock, setting prices, dealing with suppliers. Score is the money left at the end (start: $500). Cost per task is the mean API cost of one full simulated year. | Andon Labsboard shows no update date, retrieved Nov 11, 2026 |
| LiveBench Data Analysis | LiveBench's data-analysis suite, scored from 0 to 100. A single run of 150 tasks per model, so a difference of a few points can be one task. Cost here is derived: LiveBench's total suite cost divided by its task count, as the LiveBench site itself does. | LiveBenchrelease Nov 6, 2026 |
| Arena Vision | Human preference on prompts that include images, such as charts, photos and screenshots. | Arena (arena.ai)results as of Nov 9, 2026 |
| Vectara Hallucination Rate | How often a model adds facts that are not in a document it is asked to summarise. Lower is better. | Vectararesults as of Oct 22, 2026 |
| FrontierMath (Tiers 1-3) | Unpublished, expert-written research-level maths problems, run by Epoch AI. | Epoch AIresults as of Nov 6, 2026 |
| GPQA Diamond | Graduate-level multiple-choice questions in biology, chemistry and physics, run by Epoch AI. Top models are close to the ceiling, so small gaps mean little. | Epoch AIresults as of Nov 6, 2026 |
| ARC-AGI-2 | Visual grid puzzles where each task needs a new rule to be worked out, run by the ARC Prize Foundation. Cost per task is ARC Prize's measured cost per task. | ARC Prizeresults as of Nov 10, 2026 |
Benchmarks measure narrow tasks under one set of conditions, and results move quickly as models and agent tools change. Gaps of a few points are often within the margin of error. A model that ranks first here may not be the best on your own documents, processes or data. List prices change and some are introductory; open-weight prices depend on the host. Several well-known boards are not used to rank: SWE-bench, SWE-Bench Pro, OSWorld, tau-bench, BFCL and Aider have results for few or none of the models listed here, and METR's time-horizon board is no longer updated. This page is general information, not purchasing advice. Chmltech has no commercial relationship with any model developer or benchmark listed.
Spotted an error or a model we missed? Let us know and we will review it.
Work with us
Benchmarks narrow the field. The decision comes from testing two or three candidates on your own documents, processes and data, with cost, privacy and reliability in the scorecard. Chmltech helps Canadian businesses run that evaluation and build the agents that follow.