Computer and browser use
We track 77 benchmarks in computer and browser use. They run from everyday level to research frontier level.
OpenAI and Anthropic held the top computer and browser use score about equally, 35% and 34% of the contested time
That's about 16 of the 46 months since January 2023, averaged over these benchmarks. We only count time while other labs are still testing, and stop once a model passes 85%. Faded stretches no longer count.
OpenAI ByteDance Seed Google DeepMind Microsoft Qwen Alibaba Group Amazon Anthropic Moonshot AI
Everyday Most adults, without training
University Undergraduates and entry-level professionals
Expert Specialists with years of training, often over hours of work
Sources: Official leaderboards, boards copied by Epoch AI (CC BY 4.0) and labs’ model cards
Share of contested time at number one
OpenAI 35% Anthropic 34% 
Qwen 15% 
ByteDance Seed 6% 
Alibaba Group 6% 
Google DeepMind 4% Moonshot AI 1%
Qwen 15% ByteDance Seed 6% Alibaba Group 6% Google DeepMind 4% Moonshot AI 1% Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
Closed out: 8 benchmarks passed 85%
OpenAI 3 Google DeepMind 2 Anthropic 1 ByteDance Seed 1 Qwen 1
Google DeepMind 2 Anthropic 1 ByteDance Seed 1 Qwen 1 Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
- Jun 2026 OSWorld-Verified Claude Fable 5
- May 2026 AndroidWorld Gemini 3.5 Flash
- Mar 2026 Online-Mind2Web GPT-5.4
- Feb 2026 BrowseComp Gemini 3.1 Pro
- Dec 2025 ScreenSpot-Pro GPT-5.2
- Apr 2025 ScreenSpot-v2 UI-TARS-1.5
- Jan 2025 ScreenSpot Qwen2.5-VL-72B
- Jan 2025 WebVoyager OpenAI Computer-Using Agent (CUA)
Each benchmark counts from the one source its own page leads with: its official leaderboard when that is current, else Epoch AI’s own runs, else labs’ own numbers. We date independent runs by the model’s release and lab reports by when they came out.
OpenAI DeepSeek Qwen Alibaba Group Google DeepMind Anthropic Moonshot AI Z.ai ByteDance Seed
Everyday Most adults, without training
DeepSeek-V3.2 University Undergraduates and entry-level professionals
Expert Specialists with years of training, often over hours of work
Sources: Official leaderboards and boards copied by Epoch AI (CC BY 4.0)
Share of contested time at number one, independent runs only
Anthropic 39% OpenAI 34% Qwen 11% Alibaba Group 5% ByteDance Seed 5% Google DeepMind 3% Z.ai 2% Moonshot AI 1%
Qwen 11% Alibaba Group 5% ByteDance Seed 5% Google DeepMind 3% Z.ai 2% Moonshot AI 1% Sources: Official leaderboards and Epoch AI (CC BY 4.0)
Closed out: 3 benchmarks passed 85%
Anthropic 1 Google DeepMind 1 OpenAI 1
Google DeepMind 1 OpenAI 1 Sources: Official leaderboards and Epoch AI (CC BY 4.0)
- Jun 2026 OSWorld-Verified Claude Fable 5
- May 2026 AndroidWorld Gemini 3.5 Flash
- Mar 2026 Online-Mind2Web GPT-5.4
Every independent run: each benchmark’s own leaderboard, Epoch AI’s runs and boards Epoch AI copied, dated by the model’s release. We leave out benchmarks that only labs report.
Z.ai Qwen Amazon ByteDance Seed OpenAI Google DeepMind Microsoft Alibaba Group Anthropic
Everyday Most adults, without training
University Undergraduates and entry-level professionals
Expert Specialists with years of training, often over hours of work
Sources: Official leaderboards, boards copied by Epoch AI (CC BY 4.0), labs’ model cards and claims collected by Epoch AI (CC BY 4.0)
Share of contested time at number one, claimed scores included
Anthropic 38% OpenAI 25% ByteDance Seed 16% Alibaba Group 9% Qwen 7% Z.ai 5% Google DeepMind 1%
ByteDance Seed 16% Alibaba Group 9% Qwen 7% Z.ai 5% Google DeepMind 1% Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
Closed out: 8 benchmarks passed 85%
Google DeepMind 2 OpenAI 2 ByteDance Seed 2 Anthropic 1 Qwen 1 Sources: Official leaderboards, Epoch AI (CC BY 4.0) and labs’ model cards
- Jun 2026 OSWorld-Verified Claude Fable 5
- May 2026 AndroidWorld Gemini 3.5 Flash
- Feb 2026 BrowseComp Gemini 3.1 Pro
- Dec 2025 ScreenSpot-Pro GPT-5.2
- Sep 2025 Online-Mind2Web UI-TARS-2
- Apr 2025 ScreenSpot-v2 UI-TARS-1.5
- Jan 2025 ScreenSpot Qwen2.5-VL-72B
- Jan 2025 WebVoyager OpenAI Computer-Using Agent (CUA)
Every score, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model.
From everyday to research frontier level
Grouped by who could solve the tasks, from things most adults can do to open research problems. The bar under each name is its best score.
Everyday
Most adults, without training
- OSWorld-Verified
- AndroidWorld
- OSWorld
- MobileWorld
- WebVoyager
- Online-Mind2Web
- ScreenSpot-v2
- ScreenSpot
- BU Bench V2.1
- WebArena
- CLASHarness
- GroundUI
- OSWorld-G
- WebArena-Verified
- VisualWebBench
- WebTailBench
- MyPCBench
- OSWorld-MCP
- Realtime-Bench
- WARC-Bench
- Windows Agent Arena
- ClawBench V2
- OpenComputer
- ClawBench
- iOSWorld
- AndroidDaily
- CreativeWork
- UI-Vision Element Grounding
- VisualAgentBench
- AndroidControl
- DeepShop
- MobileGym
- WebHarbor
- CRAB
- WebCanvas
- WebLINX
- Android in the Zoo (AitZ)
- Mind2Web
- MobileMiniWoB++
- UI-Vision Layout Grounding
- MMBench-GUI
- WebBench (Halluminate)
- WorldGUI
- AssistantBench
- BU Bench V2
- GUI Odyssey
- LifeVal
- MiniWoB++
- OSGym
- UI-Vision
- UI-Vision Action Prediction
- Verifiable Browsing Task (Meta)
- WebShop
University
Undergraduates and entry-level professionals
Expert
Specialists with years of training, often over hours of work
Research frontier
Researchers, on open problems or tasks that take days
By kind
Desktop apps
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| OSWorld-Verified 369 open-ended tasks in real desktop apps. Agent acts through screenshots, mouse and keyboard. Execution-based checkers score success. | Everyday | Checked automatically | 86% | In use |
| OSWorld 2.0 Agents complete 108 long-horizon workflows across 31 self-hosted websites and desktop apps, averaging 318 tool calls. Scored by binary completion and checkpoint partial credit. | University | Checked automatically | 20.6% | In use |
| OSWorld Agents complete 369 real computer tasks in Ubuntu desktop apps through screenshots and mouse and keyboard actions. Execution-based scripts check the final state. | Everyday | Checked automatically | 72.1% | Fading |
| OSWorld 2.1 Agents complete 108 long-horizon workflows across 31 self-hosted websites and desktop apps, averaging 318 tool calls. Scored by binary completion and checkpoint partial credit. | University | Checked automatically | 81.8% | In use |
| RecreationBench Agents explore a running reference app and rebuild it, across 250 tasks on five platforms. Programmatic tests and a VLM judge score the result. | Expert | Checked automatically | 58.1% | In use |
| ScienceBoard Computer-use agents do scientific workflows in desktop software such as ChimeraX, Celestia, GRASS GIS and Lean 4. The apps are modified to expose state for checking. | Expert | Checked automatically | 68% | Not in cards |
| CLASHarness Computer-use agents work in multi-agent setups where peers report stale states, give contradictory instructions or compete for files. Agents must check claims against the environment. | Everyday | Checked automatically | 28.7% | Not in cards |
| OSWorld-Science Computer-use agents do 146 scientific software tasks, such as molecular drawing, across several domains. Produced artifacts are checked to score each task. | Expert | Checked automatically | 73.7% | Not in cards |
| MyPCBench Computer-use agents work in a personalized Linux desktop with 17 simulated logged-in web apps and a user's history, across 184 tasks. | Everyday | Checked automatically | 62% | Not in cards |
| OSWorld-MCP Computer-use agents complete OSWorld desktop tasks while they can also call 158 MCP tools across 7 common applications. The benchmark checks whether agents pick a tool call or GUI actions at the right time. It reports task accuracy, tool invocation rate and average completion steps. | Everyday | Checked automatically | 56.6% | Not in cards |
| WeaveBench Computer-use agents do 114 long tasks in 8 work domains that mix GUI actions with command-line and code operations in one trajectory. Results are checked against verifiable artifacts. | Expert | Checked automatically | 41.2% | Not in cards |
| Windows Agent Arena Agents complete more than 150 tasks in a real Windows OS with its apps, tools and browsers. Scripts adapted from OSWorld check the final state. | Everyday | Checked automatically | 50.6% | Dropped |
| OpenComputer Computer-use agents complete 1,000 desktop tasks across 33 real applications, including browsers, office tools, creative software, development environments and messaging apps. App-specific verifiers inspect the final application state, and the harness gives partial credit from the recorded trajectory. | Everyday | Checked automatically | 68.3% | Not in cards |
| OSWorld 2.1 (offline subset) Agents complete the 82 OSWorld 2.1 tasks on the official offline list, with no internet access. Scored by checkpoint partial credit and strict pass rate. | University | Checked automatically | 89.3% | In use |
| CreativeWork Agents complete productivity tasks in Notion, Canva and Figma by mixing GUI actions with MCP tool calls. | Everyday | Checked automatically | 42.5% | In use |
| Spider2-V Multimodal agents do data science and engineering workflows in real enterprise apps such as BigQuery, dbt and Airbyte, on a desktop computer. Execution-based checks score each task. | Expert | Checked automatically | 14% | Not in cards |
| CUA-World-Long Computer-use agents do one long task in each of 200 real applications, such as 3D Slicer, GIMP and OpenEMR, often needing 500 or more steps. Checklist verifiers with privileged state score each task. | Expert | Checked automatically | 27.5% | Not in cards |
| OSWorld 2.0 (Aug 2026) Agents complete 108 long-horizon workflows across 31 self-hosted websites and desktop apps, averaging 318 tool calls. Scored by binary completion and checkpoint partial credit. | University | Checked automatically | 77.9% | System cards only |
| CRAB Multimodal agents do tasks that span an Ubuntu desktop and an Android phone at once. Graph-based evaluators give credit for intermediate steps. | Everyday | Checked automatically | 38% | Not in cards |
| Internal computer use safety benchmark (OpenAI) Models complete computer-use tasks while avoiding unsafe actions. | Research frontier | Graded by a model | 2.40% | System cards only |
| WorldGUI Desktop GUI agents do 611 tasks that start from varied initial states, such as the target app not being open. Tasks come with instructional videos and project files. | Everyday | Checked automatically | 45.8% | Not in cards |
| OSGym Infrastructure running 1,000+ parallel OS sandboxes for computer-use agent training and evaluation. Tasks span browsing, documents, SWE and multi-app work. | Everyday | Checked automatically | Not in cards |
Phones
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| AndroidWorld Agents complete 116 tasks across 20 real Android apps in an emulator. Each task checks the device's system state to judge success. | Everyday | Checked automatically | 100% | In use |
| MobileWorld Agents complete 201 tasks across 20 mobile apps, including tasks that need user clarification or MCP tools. Backend checks verify success. | Everyday | Checked automatically | 55.3% | In use |
| iOSWorld Phone agents work in a native iOS simulator across 26 purpose-built apps that share one persistent user's data. Tasks span single apps, multi-app flows and memory or personalization from the user's history. | Everyday | Checked automatically | 51.9% | Not in cards |
| AndroidDaily Agents complete 350 daily tasks in 94 real closed-source Android apps. A process-aware reviewer model checks each trajectory against written guidelines. | Everyday | Graded by a model |  61.9% | Fading |
| AndroidControl Agents predict the next action on Android app screens from either a high-level goal or a low-level step instruction. Scored by step accuracy against human demonstrations. | Everyday | Checked automatically |  91.2% | Dropped |
| MobileGym Mobile GUI agents use 28 simulated apps in a browser-hosted phone, across 416 parameterized task templates. Deterministic judges check the app state. | Everyday | Checked automatically | 58.8% | Not in cards |
| Android in the Zoo (AitZ) Models predict the next action on Android app screens, given the goal and past steps. Scored by exact match of the predicted action on the test split. | Everyday | Checked automatically | 83.2% | Dropped |
| MobileMiniWoB++ Agents complete MiniWoB++ web tasks inside the Android emulator of AndroidWorld. Scored by task success rate. | Everyday | Checked automatically | 68% | Dropped |
| AndroidLab Mobile agents complete 138 tasks across nine apps on predefined Android virtual devices, in text-only and multimodal modes. Checks on the final device state score each task. | University | Checked automatically | 31.2% | Not in cards |
| GUI Odyssey Mobile GUI agents carry out tasks that cross several apps, such as sharing a post from one social app to another. The dataset has 8,334 episodes averaging 15.3 steps, over 212 apps, 1,357 app combinations and six devices. Agents are scored by how often each predicted action matches the recorded step. | Everyday | Checked automatically | Not in cards |
Web browsing
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| BrowseComp Hard-to-find facts that need persistent web browsing. Short verifiable answers graded for exact correctness. | Expert | Checked automatically | 93.3% | In use |
| BrowseComp Extended Agents browse the web to find hard-to-locate facts with short answers. Mercor adds 100 new expert-written tasks to the 130 public ones and reports both scores. | Expert | Checked automatically | 73% | Not in cards |
| WebVoyager Agents complete user tasks on 15 popular real websites by browsing from screenshots and page text. A GPT-4V judge grades whether each task succeeded. | Everyday | Graded by a model |  89.3% | Fading |
| Online-Mind2Web Web agents do 300 tasks on 136 live websites. An LLM judge reads the agent's trajectory and decides whether the task succeeded. | Everyday | Graded by a model | 92.7% | Fading |
| ApprenticeBench Agent joins a simulated company and processes 100 accounts-payable bills in Odoo via GUI or API, learning from records and feedback. Cumulative success rate. | University | Checked automatically | 72% | Not in cards |
| BU Bench V2.1 Browser agents do 200 real web tasks. A findings judge checks tool evidence, answers and files, and deterministic integrity checks run as well. | Everyday | Graded by a model | 81.3 | Not in cards |
| WebArena Agents complete 812 realistic tasks on self-hosted shopping, forum, GitLab, CMS and map websites. Functional graders check the outcome. | Everyday | Checked automatically | 68% | Fading |
| WebArena-Verified Agents complete 812 audited WebArena tasks on self-hosted websites. Deterministic evaluators check the agent's answer and captured network traces. | Everyday | Checked automatically |  69% | In use |
| WebTailBench Computer-use agents complete 609 hand-verified tasks on live websites across 11 categories, such as bookings, real estate and job search. A companion set of 111 harmful tasks tests refusals. | Everyday | Graded by a model | 57.4% | Fading |
| Realtime-Bench Agents act in 50 live game and web tasks where a late action fails: 30 web tasks (travel, shopping, services and security) and 20 games. The environment keeps moving while the model thinks. Three clocks separate speed from skill: Wall Clock runs on real time, Paused Clock stops while the model thinks, and Token Clock advances with each generated token. Scores are task success rates over three runs. | Everyday | Checked automatically | 42.7% | Not in cards |
| WARC-Bench Screenshot-only agents do short web subtasks, such as date pickers and scroll containers, on real sites replayed from Web Archive snapshots. A program checks the end state. | Everyday | Checked automatically | 64.8% | Not in cards |
| ClawBench V2 Browser agents do 130 new everyday tasks on live websites under six supported harnesses. Request interception and a more lenient LLM judge score each run. | Everyday | Graded by a model | 44.6% | Not in cards |
| VisualWebArena Multimodal web agents complete 910 visually grounded tasks on self-hosted classifieds, shopping and Reddit-style sites. Execution-based checks score each task. | University | Checked automatically | 16.4% | Not in cards |
| ClawBench Browser agents do 153 everyday tasks on live websites, such as booking travel, ordering food and applying for jobs. Request interception stops real side effects, and an LLM judge scores the five-layer traces. | Everyday | Graded by a model | 61.4% | Not in cards |
| CocoaBench Unified digital agents do long human-designed tasks that mix vision, search and coding. An automatic function checks each final output. | Expert | Checked automatically | 45.1% | Not in cards |
| SaaS-Bench 106 long-horizon GUI workflows across 23 locally deployed real SaaS apps. Scored from app databases with weighted checkpoints and full-resolution rate. | University | Checked automatically | 31.1% | In use |
| VisualAgentBench Large multimodal models act as agents in embodied, GUI and visual design environments, including mobile, WebArena-based web and Minecraft tasks. | Everyday | Checked automatically | 36.2% | Not in cards |
| DeepShop Web agents answer complex shopping queries on live sites with product attributes, search filters and sorting preferences. An LLM judge checks whether results meet each requirement. | Everyday | Graded by a model |  42.3% | Fading |
| WebHarbor Agents navigate local copies of 153 websites to search, sign in, book services, and change stored state. The suite contains 3,518 tasks, including 306 verified test tasks and 2,288 training tasks. The test score measures the percentage of tasks passed by scripted verifiers. | Everyday | Checked automatically | 86.7% | Not in cards |
| WebCanvas Web agents do tasks on live websites, scored on key intermediate states rather than one fixed action path, so the test survives site changes. | Everyday | Checked automatically | 28.8% | Not in cards |
| WebLINX Agents follow a user's multi-turn chat instructions to navigate real websites. 100,000 interactions from 2,300 expert demonstrations give turn-level metrics. | Everyday | Checked automatically | 10.4 | Not in cards |
| Mind2Web Models predict the next web action for over 2,000 crowdsourced tasks on 137 real websites in 31 domains. Scored by element accuracy and step success rate. | Everyday | Checked automatically | 55.8% | Dropped |
| WebBench (Halluminate) Browser agents complete 2,454 tasks on 452 live websites from the global top 1,000 by traffic. Tasks either read information or write data: creating, updating and deleting records, logging in and downloading files. The board compares browser agent products. | Everyday | 59.8% | Not in cards | |
| AssistantBench Web agents do 214 realistic, time-consuming tasks on the open web, such as monitoring real-estate markets. Answers are checked automatically against gold answers. | Everyday | Checked automatically | Not in cards | |
| BU Bench V2 Browser agents do real web tasks with the Browser Use library. A judge scores each run against a rubric. | Everyday | Graded by a model | Not in cards | |
| LifeVal Assistants carry out everyday personal tasks rather than advise on them: finding childcare, booking a car inspection, ordering groceries on a budget, handling government paperwork. Tasks come from time-use surveys and from about 100 people who recorded tasks they wanted an assistant to handle. Score is average reward across tasks. | Everyday | Not in cards | ||
| MiniWoB++ Agents operate over 100 small synthetic web pages, such as forms, menus and date pickers, from a natural-language instruction. Each task returns a success reward. | Everyday | Checked automatically | Not in cards | |
| Verifiable Browsing Task (Meta) Agents answer list-style information-seeking questions, written as realistic user requests, using browser search, open and find tools. | Everyday | Checked automatically | Not in cards | |
| WebShop Agents search and click through a simulated shopping site with 1.18 million real products to buy an item that matches a text instruction. Reward reflects how well the purchase matches. | Everyday | Checked automatically | Not in cards | |
| WorkArena Web agents do routine knowledge-work tasks in a live ServiceNow instance, such as filling forms and filtering lists. WorkArena++ adds 682 compositional tasks at levels 2 and 3. | University | Checked automatically | Not in cards | |
| WorkArena++ Web agents complete 682 compositional knowledge-work tasks in ServiceNow, which need planning, reasoning and memory across many steps. | Expert | Checked automatically | Not in cards |
Finding things on screen
| Benchmark | Level | Scored | Best score | Status |
|---|---|---|---|---|
| ScreenSpot-Pro GUI grounding on high-resolution professional-software screenshots. The model must locate the target UI element. | University | Checked automatically | 92.7% | In use |
| ScreenSpot-v2 Models locate a described element on mobile, desktop and web screenshots. A click inside the target box counts as correct. | Everyday | Checked automatically | 95.2% | Fading |
| ScreenSpot Model locates a described UI element in mobile, desktop and web screenshots. Scored by click-point accuracy. | Everyday | Checked automatically | 95.4% | Fading |
| GroundUI Agents get a screenshot and an instruction and must click the right UI element. About 18K test samples cover web, desktop and mobile screens. Scored by whether the point falls in the target box. | Everyday | Checked automatically | 35.2% | Dropped |
| OSWorld-G Models map 564 desktop instructions to the right screen element or point, covering text matching, element recognition, layout understanding and fine manipulation. | Everyday | Checked automatically | 71.3% | Fading |
| VisualWebBench Models answer questions about web page screenshots across seven tasks, such as captioning, OCR, element grounding and action prediction. | Everyday | Checked automatically | 48% | Dropped |
| ADeptS-Bench Desktop Agents choose safe interface actions on the Desktop test set. The published safety score combines task success and attack resistance. | Expert | Checked automatically | 76% | Not in cards |
| ADeptS-Bench Mobile Agents choose safe interface actions on the Mobile test set. The published safety score combines task success and attack resistance. | Expert | Checked automatically | 73.8% | Not in cards |
| UI-Vision Element Grounding Models locate interface elements from instructions. The score is the published overall grounding accuracy. | Everyday | Checked automatically | 8.27% | Not in cards |
| UI-Vision Layout Grounding Models identify interface layout regions. The score is the published intersection over union percentage. | Everyday | Checked automatically | 24.3% | Not in cards |
| MMBench-GUI GUI agents are tested at four levels on Windows, macOS, Linux, iOS, Android and the web: content understanding, element grounding, task automation and task collaboration. | Everyday | Checked automatically | 81.5% | System cards only |
| ADeptS-Bench Computer-use agents pick the next interface action from a screenshot and instruction. A deterministic grader scores whether the action is safe. | Expert | Checked automatically | System cards only | |
| UI-Vision Computer use agents are tested offline on human desktop demonstrations from 83 applications. Three tasks go from element grounding to layout grounding to next-action prediction for clicks, drags and key presses. | Everyday | Checked automatically | Not in cards | |
| UI-Vision Action Prediction Models predict the next desktop interface action. The score is the published step success rate. | Everyday | Checked automatically | Not in cards |