Track the best-performing AI agents and models across browser automation, computer use, research, and coding benchmarks. Each leaderboard page includes methodology notes, scope labels, and source-linked results.
WebVoyager measures end-to-end browser task completion across live websites. It is widely used for comparing production-style browser agent systems.
Current top system: Jina (98.9%)
| Rank | System / Submission | Score | Organization | Notes | Source | Share |
|---|---|---|---|---|---|---|
| 1 | Jina New | 98.9% | Om Labs | Om Labs custom tracker; Jina multi-model system on self-hosted WebVoyager harness. | Source | |
| 2 | Alumnium New | 98.6% | Alumnium | Accessibility-based tree parsing with integrated visual reasoning. | Source | |
| 3 | Surfer 2 New | 97.1% | H Company | System-level orchestration with submitter-defined setup details. | Source | |
| 4 | Magnitude | 93.9% | Magnitude | Open-source architecture utilizing a modular agentic stack. | Source | |
| 5 | AIME Browser-Use New | 92.34% | Aime | Custom orchestration layer with specialized browser tooling. | Source | |
| 6 | Surfer-H + Holo1 New | 92.2% | H Company | Multi-modal action kernels integrated via H-Company research. | Source | |
| 7 | Browserable New | 90.4% | Browserable | Fine-tuned browser control models within a commercial framework. | Source | |
| 8 | Browser Use | 89.1% | Browser Use | Multi-step orchestration framework for open-source automation. | Source | |
| 9 | GLM-5V-Turbo New | 88.5% | Z.ai | Multimodal vision model optimized for GUI automation and coding tasks. | Source | |
| 10 | Agent Kura | 87.0% | Kura | 602/643 tasks (41 removed for invalid/auth issues); reported on trykura.com. | Source | |
| 10 | Operator | 87% | OpenAI | Native browser integration using proprietary vision-control models. | Source | |
| 12 | Skyvern 2.0 | 85.85% | Skyvern | DOM-level reasoning coupled with real-time error-correction. | Source | |
| 13 | Project Mariner | 83.5% | Gemini-powered reasoning with precise visual grounding. | Source | ||
| 14 | Agent-E | 73.1% | Emergence AI | Hierarchical planning modules within a multi-agent framework. | Source | |
| 14 | Notte New | 73.1% | Notte | Standardized operator stack for open-source performance evaluation. | Source | |
| 16 | WebSight | 68% | Academic Research | Navigation system prioritizing visual-only perceptual inputs. | Source | |
| 17 | Runner H 0.1 | 67% | H Company | Foundational agent architecture for general web interaction. | Source | |
| 18 | WebVoyager | 59.1% | Academic Research | Baseline implementation using standard multimodal LLM control. | Source | |
| 19 | Anthropic Computer Use 3.5 | 56.0% | Anthropic | Sampled 50/602 tasks for direct comparison; reported on trykura.com. | Source | |
| 20 | WILBUR | 53% | Academic Research | Research implementation using black-box optimization techniques. | Source | |
| 21 | GPT-4 (All Tools) | 30.8% | OpenAI | ChatGPT integrated tool baseline from original WebVoyager paper; reported on arxiv.org. | Source |
Browser agents
WebVoyager is a benchmark for browser agents operating on live websites. It focuses on practical tasks such as navigation, search, form completion, and multi-step workflows across a broad website mix.
Top: Jina (98.9%)
Caveat: Rows may use different evaluation settings and are not always strict apples-to-apples.
Research/search
BrowseComp targets difficult browse-and-synthesize research questions that are easy to verify but hard to answer without strong search and reasoning strategy.
Top: GPT-5.5 Pro (90.1%)
Caveat: Mixed-scope benchmark: model-only and tool-augmented rows are not inherently directly comparable.
Browser agents
WebArena evaluates browser agents in controlled, self-hosted web environments that represent realistic application patterns such as e-commerce, forums, and developer workflows.
Top: DeepSeek V3.2 (74.3%)
Caveat: Even with controlled environments, ranking rows can differ by setup and submission policy.
Coding
SWE-bench Verified evaluates software engineering performance on real GitHub issues with stricter quality controls than the broader SWE-bench set.
Top: Claude Mythos (93.9%)
Caveat: Model-focused benchmark, but harness and evaluation policy still affect outcomes.
Computer use
OSWorld evaluates computer-use agents across 369 real desktop tasks spanning Ubuntu, Windows, and macOS — covering web apps, desktop software, file I/O, and multi-application workflows.
Top: Mythos Preview (79.6%)
Caveat: Self-reported and independently verified rows coexist — check the source before comparing directly.
Model evals / reasoning
GAIA (General AI Assistants) evaluates agents on over 450 real-world questions with unambiguous, verifiable answers — requiring multi-step reasoning, tool use, web search, and file handling across three difficulty levels.
Top: OPS-Agentic-Search (92.36%)
Caveat: Top entries are multi-model ensembles — scores cannot be attributed to any single model.
Browser agents
ClawBench evaluates AI agents on 153 everyday tasks that real people need to complete regularly — booking appointments, completing purchases, submitting job applications, and filling in forms — across 144 live production websites in 15 categories.
Top: Claude Sonnet 4.6 (33.3%)
Caveat: Very new benchmark (April 2026) — published results cover only 7 frontier models. Expect the leaderboard to expand rapidly.
Browser agents
Online-Mind2Web is a live browser agent benchmark of 300 diverse, realistic tasks across 136 popular websites — spanning shopping, finance, travel, government, and more. Unlike static offline benchmarks, agents interact with real, dynamic pages as they exist at evaluation time.
Top: Browser Use Cloud (bu-max) (97.0%)
Caveat: Judge methodology varies significantly across submissions — human eval, WebJudge, and custom agentic judges produce different scores for the same agent. Always check the Notes column before comparing rows.
Model evals / reasoning
τ-bench (TAU-bench) evaluates AI agents in realistic enterprise tool-use scenarios across retail and airline domains — testing multi-turn conversation, policy adherence, database interactions, and rule-following consistency over many trials.
Top: Step-3.5-Flash (88.2%)
Caveat: Score comparisons across organizations require caution — prompt setup, tool schema, and trial count differ between submissions.
Model evals / reasoning
AgentBench evaluates LLMs as agents across 8 distinct interactive environments — including OS interaction, database querying, knowledge graph traversal, digital card games, lateral thinking puzzles, house-holding tasks, web browsing, and web shopping.
Top: AgentRL w/ Qwen2.5-32B-Instruct (70.4%)
Caveat: Community-submitted leaderboard — rows are self-reported and not independently verified. Check source links before drawing strong conclusions.
Methodology reminder: some benchmarks measure model capability, while others measure full systems (model + tools + policy). Mixed pages include both and should be read as directional unless setups are fully aligned.