Back to the trend of the day: Apple is building an AI helper that can use your Mac for you
Poly Rating · from the trend of the day

Best AI Agent to Control Your Computer 2026

Claude's computer use completes 85.96% of real desktop tasks with only Screen Recording and Accessibility, while OpenClaw asks for Full Disk Access and has published no score at all.

  • No. 1: Qwen3.8 Max (Alibaba)
  • 10 places

01Final list

The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.

Order: from most to least AI agreement.

A bright badge means this AI put the entry on its list. A dim badge means it did not.

  1. 1

    Qwen3.8 Max (Alibaba)

    Highest listed OSWorld-Verified score, 86.1%, but only Alibaba reports it. Steel says the visual model tested isn't public, so you can't yet run it in your own sandbox.

    Included it: Claude, ChatGPT, Meta, Grok, MiMo, Kimi, DeepSeek, Qwen. Did not include: Gemini, Wenxin.

    In their lists: Claude: 5 · ChatGPT: 1 · Meta: see list · Grok: 3 · MiMo: 5 · Kimi: 2 · DeepSeek: 5 · Qwen: 1

    What each AI noted 8
    • Claudeplace 5

      It holds the highest listed OSWorld-Verified number (86.1%), but the weights are unconfirmed and the score is vendor-reported only.

      Source ↗
    • ChatGPTplace 1

      86.1% leads the current OSWorld-Verified model table—but the published row does not disclose its agent scaffold.

      Source ↗
    • Meta

      Named differently in its own list: open the full list.

      see list
    • Grokplace 3

      Leads many aggregators at 86.1% Verified with open weights.

      Source ↗
    • MiMoplace 5

      Tops the desktop scoreboard at 86.1% — listed as open while the tested visual model isn't.

      Source ↗
    • Kimiplace 2

      86.1% on OSWorld-Verified — and the weights are downloadable (a 27B checkpoint hits 84.3%), so the top desktop result needs zero access to your machine.

      Source ↗
    • DeepSeekplace 5

      86.1% on OSWorld-Verified—the highest listed—but the agent's local permission footprint is not publicly detailed.

      Source ↗
    • Qwenplace 1

      86.1% on OSWorld-Verified, the highest verified desktop score yet, while allowing granular tool-permission controls.

      Source ↗
    8 AIs put it here
  2. 2

    Claude computer use (Anthropic: Claude Code, Cowork)

    Highest officially run score from a mainstream agent, 85.96%, with Screen Recording and Accessibility only, no Full Disk Access. The asterisk is July's SharedRoot sandbox escape.

    Included it: Claude, ChatGPT, Meta, Grok, MiMo, Kimi, DeepSeek, Qwen. Did not include: Gemini, Wenxin.

    In their lists: Claude: 1 · ChatGPT: 2 · Meta: 1 · Grok: 10 · MiMo: 1 · Kimi: 6 · DeepSeek: 3 · Qwen: 3

    What each AI noted 8
    • Claudeplace 1

      Scores 85.0% on OSWorld-Verified and 66.1% partial on OSWorld 2.0, with a setup of Screen Recording plus Accessibility and no Full Disk Access.

      Source ↗
    • ChatGPTplace 2

      85% puts it within 1.1 points of the top score; the benchmark table attributes the result to Anthropic.

      Source ↗
    • Metaplace 1

      85.96% OSWorld-Verified for Fable 5 and 81.8% OSWorld 2.1 partial for Opus 5.5, with cloud-sandbox default, not full-disk by default.

      Source ↗
    • Grokplace 10

      Session-scoped Accessibility + Screen Recording with per-app approval prompts.

      Source ↗
    • MiMoplace 1

      The only agent here that runs its commands in a cage — and the cage cut permission prompts 84%.

      Source ↗
    • Kimiplace 6

      The consumer packaging of the same stack: it asks per app, advises against banking or health apps, and reaches your computer only while the desktop app is open.

      Source ↗
    • DeepSeekplace 3

      85% on OSWorld-Verified with just Accessibility and Screen Recording—no Full Disk Access—and a per-app override tier in macOS 27.3.

      Source ↗
    • Qwenplace 3

      Hits 85.0% on real desktop tasks but relies on full screen-recording and accessibility overrides.

      Source ↗
    8 AIs put it here
  3. 3

    OpenAI ChatGPT agent / Codex computer use

    Two modes with two access costs: the cloud ChatGPT agent never sees your desktop, while Codex on your Mac needs Screen Recording and Accessibility. GPT-5.5 scores 78.7% on OSWorld-Verified.

    Included it: Claude, Meta, Grok, MiMo, Kimi, DeepSeek, Qwen. Did not include: ChatGPT, Gemini, Wenxin.

    In their lists: Claude: see list · Meta: 2 · Grok: 8 · MiMo: 3 · Kimi: 4 · DeepSeek: 1 · Qwen: see list

    What each AI noted 7
    • Claude

      Named differently in its own list: open the full list.

      see list
    • Metaplace 2

      78.7% OSWorld-Verified for GPT-5.5 and 62.6% OSWorld 2.0 partial for GPT-5.6 Sol, while the agent runs in an OpenAI sandbox and does not see your desktop by default.

      Source ↗
    • Grokplace 8

      Early 38.1% OSWorld; later sandboxed virtual computer reduces local full-control needs.

      Source ↗
    • MiMoplace 3

      The strongest new computer-use brain at one-third the cost — with its own safety card listing workplace mishaps.

      Source ↗
    • Kimiplace 4

      The most capable native agent on your own Mac (75% OSWorld-Verified) is also the neediest: Screen Recording, Accessibility, per-app approvals — plus an opt-in Locked Use mode that can temporarily unlock the Mac.

      Source ↗
    • DeepSeekplace 1

      Delivers 38.1% on OSWorld from a cloud browser sandbox—almost no local disk or Accessibility permission needed.

      Source ↗
    • Qwen

      Named differently in its own list: open the full list.

      see list
    7 AIs put it here
  4. 4

    Simular Sai (Agent S)

    Can run on a cloud desktop instead of your Mac. It claims 73% partial on OSWorld 2.0 but fully finishes only 28.25% of jobs, and Simular co-wrote the benchmark and reported the run itself.

    Included it: Claude, Meta, Grok, MiMo, Kimi, DeepSeek, Qwen. Did not include: ChatGPT, Gemini, Wenxin.

    In their lists: Claude: 6 · Meta: 3 · Grok: 4 · MiMo: 7 · Kimi: 7 · DeepSeek: 6 · Qwen: 8

    What each AI noted 7
    • Claudeplace 6

      Costs $15.70 per task on OSWorld 2.0, with a 28.25% binary score.

      Source ↗
    • Metaplace 3

      73% partial on OSWorld 2.0 but only 28.25% binary - the widest partial-versus-finished gap in this ranking.

      Source ↗
    • Grokplace 4

      Highest reported partial on OSWorld 2.0/2.1 (~73–79%) at lower cost than pure frontier models.

      Source ↗
    • MiMoplace 7

      Sold by the job: $15.70 per finished task — the only agent with a price tag on the result.

      Source ↗
    • Kimiplace 7

      69.9% on OSWorld — within 2.5 points of the human baseline — from an open framework whose Behavior Best-of-N buys the points with extra rollouts, not extra permissions.

      Source ↗
    • DeepSeekplace 6

      72.6% on OSWorld—above the human baseline—in an open-source harness that runs with Accessibility and Screen Recording.

      Source ↗
    • Qwenplace 8

      Achieves 73% on OSWorld 2.0, balancing long-horizon professional tasks with strict local sandboxing.

      Source ↗
    7 AIs put it here
  5. 5

    Gemini computer use for Mac (Google)

    Scores 83% on OSWorld-Verified (Google-reported) while working in a separate window, only in apps you pick, and asking before key actions. Reviewers warn that injected text can still steer it.

    Included it: Claude, ChatGPT, Meta, MiMo, Kimi, DeepSeek. Did not include: Grok, Gemini, Wenxin, Qwen.

    In their lists: Claude: 7 · ChatGPT: see list · Meta: 5 · MiMo: 2 · Kimi: 3 · DeepSeek: 8

    What each AI noted 6
    • Claudeplace 7

      It works in a separate window, but Google's own documentation warns about prompt injection across apps.

      Source ↗
    • ChatGPT

      Named differently in its own list: open the full list.

      see list
    • Metaplace 5

      83.0% OSWorld-Verified, launched November 19, 2026 - the same day Apple’s own Mac agent was reported.

      Source ↗
    • MiMoplace 2

      Late to the table — and the only guest that shows its work before it touches an app.

      Source ↗
    • Kimiplace 3

      78.4% OSWorld-Verified with 'full: you run the environment' — the agent ships with no Mac permission sheet because the sandbox is yours.

      Source ↗
    • DeepSeekplace 8

      83% on OSWorld-Verified, but the access profile of the agent harness is not publicly documented.

      Source ↗
    6 AIs put it here
  6. 6

    Holo3 / Holo4 (H Company)

    Best score the benchmark team ran itself, 82.6%, from open weights with 3B active parameters, so you pick the sandbox. Holo4-27B claims 85.2% but is research-only.

    Included it: Claude, ChatGPT, Meta. Did not include: Grok, Gemini, MiMo, Kimi, DeepSeek, Wenxin, Qwen.

    In their lists: Claude: 3 · ChatGPT: 7 · Meta: 4

    What each AI noted 3
    • Claudeplace 3

      The top independently run OSWorld-Verified entry (82.6%) comes from a model with only 3B active parameters.

      Source ↗
    • ChatGPTplace 7

      82.6% is the strongest independently benchmark-run result among these top ten table entries; it is a specialized computer-use model.

      Source ↗
    • Metaplace 4

      85.2% on OSWorld for 27B at $0.08 per task - the best score-per-dollar here, but the 27B variant is research-only.

      Source ↗
    3 AIs put it here
  7. 7

    Meta Muse Spark / metacua

    Scores 80.67% on the official OSWorld-Verified board, but metacua drives your real Mac with no sandbox by default. A Muse incident helped prompt Apple's new warning.

    Included it: Claude, Meta, MiMo. Did not include: ChatGPT, Grok, Gemini, Kimi, DeepSeek, Wenxin, Qwen.

    In their lists: Claude: 9 · Meta: 6 · MiMo: see list

    What each AI noted 3
    • Claudeplace 9

      Posts the best binary OSWorld 2.0 score on the board, 32.0%, despite a lower partial score of 66.9%.

      Source ↗
    • Metaplace 6

      80.67% verified on OSWorld-Verified, but it drives your actual machine - the access cost is the whole machine, not a sandbox.

      Source ↗
    • MiMo

      Named differently in its own list: open the full list.

      see list
    3 AIs put it here
  8. 8

    Manus

    Needs no access to your Mac because each task gets its own cloud VM. Its 72.5% OSWorld figure comes from one market report only and isn't on any official board, so treat it as an estimate.

    Included it: Meta, DeepSeek. Did not include: Claude, ChatGPT, Grok, Gemini, MiMo, Kimi, Wenxin, Qwen.

    In their lists: Meta: 8 · DeepSeek: 2

    What each AI noted 2
    • Metaplace 8

      A fully isolated cloud VM per task with zero-trust claims - strong sandbox, no published OSWorld score to prove the work.

      Source ↗
    • DeepSeekplace 2

      Runs in a cloud VM with root access, so its 72.5% OSWorld score comes without handing over your Mac's disk.

    2 AIs put it here
  9. 9

    Perplexity Personal Computer (Comet)

    Asks for all three big Mac permissions, Full Disk Access included. It has approval checkpoints but no published OSWorld score to show what you get in return.

    Included it: Meta, Wenxin. Did not include: Claude, ChatGPT, Grok, Gemini, MiMo, Kimi, DeepSeek, Qwen.

    In their lists: Meta: 9 · Wenxin: 9

    What each AI noted 2
    • Metaplace 9

      Asks for Accessibility, Screen Recording and Full Disk Access on your real Mac - the highest local permission ask - with no OSWorld score and two public security incidents.

      Source ↗
    • Wenxinplace 9

      Extends agent orchestration directly to local Mac hardware, 24/7 operation optimized for Mac mini deployments

    2 AIs put it here
  10. 10

    OpenClaw (formerly Moltbot / Clawdbot)

    Most access, least proof: Full Disk Access for Terminal plus Accessibility and Screen Recording, control of your real Mac with no sandbox, and no OSWorld score.

    Included it: Claude, Meta. Did not include: ChatGPT, Grok, Gemini, MiMo, Kimi, DeepSeek, Wenxin, Qwen.

    In their lists: Claude: 10 · Meta: 10

    What each AI noted 2
    • Claudeplace 10

      Its own docs list its macOS permission needs, which makes it the most transparent open option, but it has no benchmark score of its own.

      Source ↗
    • Metaplace 10

      Maximum access, no benchmark: Full Disk Access to Terminal plus Accessibility plus Screen Recording, with no published OSWorld score to justify it.

      Source ↗
    2 AIs put it here

More about the top three ↓

02More about the top three

The ranking is above. Here is more about the three places at the top: what they are, why they are here, and what is worth knowing.

The scores are not quite like for like: Qwen's 86.1% is Alibaba's own report, Claude's 85.96% is from the official board (other tables show 85.0%), and OpenAI's 78.7% is for GPT-5.5.

1Qwen3.8 Max (Alibaba)

No. 1 — Qwen3.8 Max (Alibaba). On the lists of 8 of 10 AIs; median place there: 3. The AIs put it first because it has the highest listed desktop score, though only Alibaba reports it and the tested model isn't public.

Alibaba Group provisional office at Xiong'an (20180503164635)
N509FZ · CC BY-SA 4.0 · Wikimedia Commons
Alibaba Binjiang Park
Danielinblue , designed by HASSELL (architects) [ 1 ] · CC BY-SA 4.0 · Wikimedia Commons
Alibaba Building at Wangjing
钉钉 · CC BY-SA 4.0 · Wikimedia Commons

Did you know

  • Qwen3.8-Max is a huge sparse model: 2.4 trillion parameters in total, but only about 95 billion are active for each token. That is roughly 4% of the model working at any moment. It was released August 3, 2026.

What it is
Alibaba's huge AI model that can operate a desktop.
Its edge
Highest listed OSWorld-Verified score of the three, at 86.1%.
Who it suits
Suits people who follow benchmarks; not those who need to run it themselves today.
  • OSWorld-Verified score86.1% (Alibaba-reported)

Wikipedia · Wikidata

Why it is here
  • The 86.1 is Alibaba's own number. Artificial Analysis measured the model taking 64 turns per task, against 14 for Qwen3.7-Max, so the top score comes with many more steps.

History
  • Alibaba unveiled its Qwen models in 2023 and released its first open-weight versions in August 2023. The early license allowed academic use, but commercial use needed separate permission.

Keep in mind
  • The 86.1 on OSWorld-Verified comes from Alibaba's own announcement. DataCamp calls the results vendor-run showcases, so nobody outside Alibaba has checked them. Treat the number as a claim, not a proven score.

    Source ↗ ✓ confirmed by several AIs
What people say
  • Those demonstrations remain company-produced and have not yet been broadly replicated by independent evaluators. — Carl Franzen, VentureBeat

Where to go further
  • Read the DataCamp review first. It lists OSWorld-Verified 86.1 for Qwen3.8-Max next to Fable 5 at 85.0, and says vendor runs are 'vendor showcases, not independent reproductions.' Look at that line before the table.

Video: How Alibaba Makes Billions · CNBC International

Compare the top three ↓

2Claude computer use (Anthropic: Claude Code, Cowork)

No. 2 — Claude computer use (Anthropic: Claude Code, Cowork). On the lists of 8 of 10 AIs; median place there: 2.5. The AIs rank it high for an officially run 85.96% score with only two Mac permissions, held back by the SharedRoot sandbox escape.

Did you know

  • The Cloud Security Alliance note says Anthropic closed the SharedRoot report as informative without a fix. Local execution, the mode the bug targets, stays unpatched. The newer cloud-by-default setting avoids it.

What it is
Anthropic's agent that sees and controls your Mac screen.
Its edge
Top officially run score with only Screen Recording and Accessibility, no Full Disk Access.
Who it suits
Suits Mac users who want strong results with few permissions; careful with local Cowork.
  • OSWorld-Verified score85.96%
Why it is here
  • The official OSWorld-Verified board lists claude-fable-5 at 86.0%, first in that list. That is a score run on the benchmark's own board, not a number from the maker.

  • Claude's local Cowork setup had a flaw called SharedRoot. Reports say it could reach files of the logged-in Mac user, and that about 500,000 macOS users with local sessions might be exposed. Cloud mode appeared unaffected.

    Source ↗ ✓ confirmed by several AIs
History
  • Anthropic launched computer use as a public beta on October 22, 2024, with an upgraded Claude 3.5 Sonnet. Anthropic called it experimental and said it could be cumbersome and error-prone.

What people say
  • We connected a folder to a fresh Claude Cowork session, sent one short message, and watched the agent escape the sandbox, — Oren Yomtov, Accomplish AI (quoted by The Hacker News)

Where to go further
  • Read Anthropic's computer-use safety page first. It says computer use has no sandbox between Claude and your apps, and that Claude asks before accessing each app. Check the app blocklist and the per-app permission settings.

Claude, Meta, MiMo put it first. What they saw in it
  • Claude: Scores 85.0% on OSWorld-Verified and 66.1% partial on OSWorld 2.0, with a setup of Screen Recording plus Accessibility and no Full Disk Access.

  • Meta: 85.96% OSWorld-Verified for Fable 5 and 81.8% OSWorld 2.1 partial for Opus 5.5, with cloud-sandbox default, not full-disk by default.

  • MiMo: The only agent here that runs its commands in a cage — and the cage cut permission prompts 84%.

Compare the top three ↓

3OpenAI ChatGPT agent / Codex computer use

No. 3 — OpenAI ChatGPT agent / Codex computer use. On the lists of 7 of 10 AIs; median place there: 3. The AIs place it third because it offers a cloud mode and a Mac mode with two permissions, but scores lower, at 78.7%.

Did you know

  • Codex computer use on a Mac can keep working even when the Mac is locked, a feature added in May 2026. OpenAI's docs say this mode is narrow: other apps and local processes can't use it to unlock the computer.

What it is
OpenAI's cloud ChatGPT agent and Codex, which can use your Mac.
Its edge
Cloud mode never sees your desktop; Mac mode needs just two permissions.
Who it suits
Suits Mac users outside the EU and UK who want a choice of modes.
  • OSWorld-Verified score78.7% OSWorld-Verified
Why it is here
  • On a Mac, Codex computer use asks for two permissions: Screen Recording to see the app, and Accessibility to click and type. It does not list Full Disk Access.

History
  • On July 17, 2025, OpenAI launched ChatGPT agent, which can use its own computer to finish tasks. It folded in the earlier Operator tool, which was then phased out.

Keep in mind
  • Codex computer use is not yet available in the EU or the UK, and it is Mac-focused. One reviewer found it failed twice until Mac privacy settings were updated to grant access.

    Source ↗ ✓ confirmed by several AIs
What people say
  • Codex now features the best computer use feature I have ever tested in any LLM or desktop agent — Federico Viticci, MacStories

Where to go further
  • Open OpenAI's Computer Use page and find 'Settings > Computer use.' It asks before using an app, and on macOS it needs Screen Recording and Accessibility. Then read the sandbox note on file reads, edits and shell commands.

DeepSeek put it first. What they saw in it
  • DeepSeek: Delivers 38.1% on OSWorld from a cloud browser sandbox—almost no local disk or Accessibility permission needed.

Compare the top three ↓

The top three side by side

Qwen3.8 Max (Alibaba)Claude computer use (Anthropic: Claude Code, Cowork)OpenAI ChatGPT agent / Codex computer use
In the AIs' lists8 / 108 / 107 / 10
OSWorld-Verified score86.1% (Alibaba-reported)85.96%78.7% OSWorld-Verified
Put first byChatGPT, QwenClaude, Meta, MiMoDeepSeek

Places with the same number of AIs are ordered by their average place in the AIs' own lists.

03Worth knowing

What the AIs noticed while building the list.

macOS 27.3 splits the old Device Control setting into System & App Protection Override and Accessibility. It adds an ask-each-time option and a menu-bar indicator, but permissions granted before 27.3 stay active.

Noticed by:
Included it: Claude, ChatGPT, Meta, Kimi, DeepSeek. Did not include: Grok, Gemini, MiMo, Wenxin, Qwen.
Source ↗

The official OSWorld-Verified #1 is Intelligence-Indeed Agent at 90.19%. It was self-submitted, and nothing is published about the access it needs.

Noticed by:
Included it: Meta, Grok, MiMo, Kimi. Did not include: Claude, ChatGPT, Gemini, DeepSeek, Wenxin, Qwen.
Source ↗

The effort setting alone moves Claude Opus 5 from 18.81% to 44.33% binary completion on OSWorld 2.0.

Noticed by:
Included it: MiMo. Did not include: Claude, ChatGPT, Meta, Grok, Gemini, Kimi, DeepSeek, Wenxin, Qwen.
Source ↗

Gemini's Mac computer use launched on November 19, 2026, the same day Apple was reported to be building its own Mac agent for as early as late 2027.

Noticed by:
Included it: Claude, Meta. Did not include: ChatGPT, Grok, Gemini, MiMo, Kimi, DeepSeek, Wenxin, Qwen.
Source ↗

04How we counted

In plain words, no formulas.

1The question

Top 10 AI Agents That Control Your Computer: How Well They Work Against How Much Access They Ask For

2Why this topic

The reader sees the benchmark result and the price paid in access side by side, which shows that the strongest agent is not always the one that needs the least of your Mac.

3Where it applies

Global; AI agents that operate a desktop (Mac, Windows, Linux), including desktop apps and browsers

4What we looked at

  • Task-completion score on OSWorld-Verified (real desktop tasks), with the tested model, agent setup and benchmark version named
  • Score on the harder OSWorld 2.0 long-horizon tasks, shown separately and never mixed with OSWorld-Verified
  • Breadth of system access the agent's setup asks for: Full Disk Access, screen recording, Accessibility control, and the override tier in macOS 27.3
  • Whether the agent can run in a sandbox, or in a mode that asks permission each time
  • OSWorld-Verified official board (OSWorld team) — Share of real desktop tasks completed, checked against the final app or file state; the OSWorld team runs these rows itself at 100 steps
  • BenchLM OSWorld-Verified tracker — 36 models, each marked as vendor-reported or run by OSWorld-Verified
  • OSWorld 2.0 / 2.1 (Snorkel AI, Steel) — Partial and strict binary completion on 108 long-horizon workflows; its own scale, never mixed with OSWorld-Verified
  • macOS 27.3 permission model (MacRumors) — Which tier an app requests: ask-each-time, Accessibility, Full Disk Access, System & App Protection Override

5Who ranked

The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT, Meta, Grok, Gemini, MiMo, Kimi, DeepSeek, Wenxin, Qwen.

6How the places were set

The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.

05Each AI's own list

🇺🇸 Claude10 places · 4 notes

The same question, answered by Claude alone. The final list above combines all the AIs.

  1. 1
    Claude Fable 5 (Anthropic computer use, 2026)

    Scores 85.0% on OSWorld-Verified and 66.1% partial on OSWorld 2.0, with a setup of Screen Recording plus Accessibility and no Full Disk Access.

    Source ↗
  2. 2
    Claude Opus 5 (Anthropic, 2026)

    Its OSWorld 2.0 result was reproduced independently by Snorkel (31.43% binary), and it posts the lowest published injection rate in GUI tests (0.54%).

    Source ↗
  3. 3
    Holo3 (H Company, 35B-A3B)

    The top independently run OSWorld-Verified entry (82.6%) comes from a model with only 3B active parameters.

    Source ↗
  4. 4
    GPT-6 Astra (OpenAI)

    It leads the offline-subset OSWorld 2.0 claims among OpenAI models at 72.6% partial, but OpenAI has not published a matching OSWorld-Verified figure that I could find.

    Source ↗
  5. 5
    Qwen3.8-Max (Alibaba)

    It holds the highest listed OSWorld-Verified number (86.1%), but the weights are unconfirmed and the score is vendor-reported only.

    Source ↗
  6. 6
    Simular Sai

    Costs $15.70 per task on OSWorld 2.0, with a 28.25% binary score.

    Source ↗
  7. 7
    Gemini computer use on Mac (Google, 2026)

    It works in a separate window, but Google's own documentation warns about prompt injection across apps.

    Source ↗
  8. 8
    Claude Sonnet 5 (Anthropic)

    81.2% on OSWorld-Verified from a mid-tier model with the same two-permission setup as the flagships.

    Source ↗
  9. 9
    Muse Spark 1.3 (Meta)

    Posts the best binary OSWorld 2.0 score on the board, 32.0%, despite a lower partial score of 66.9%.

    Source ↗
  10. 10
    OpenClaw (open-source Mac agent)

    Its own docs list its macOS permission needs, which makes it the most transparent open option, but it has no benchmark score of its own.

    Source ↗

What Claude noticed

On OSWorld 2.0 the best independent binary result (about 31%) is far below the 66–78% partial scores vendors quote, so the choice of metric changes the story.

Top OSWorld-Verified models sit within about a point of each other, which suggests the benchmark is nearing saturation.

macOS 27.3 splits the old broad permission into System & App Protection Override and Accessibility, and lets users require ask-each-time approval.

Apple is reportedly hiring to build its own Mac-controlling agent, possibly launching as soon as the end of 2027.

🇺🇸 ChatGPT10 places · 2 notes

The same question, answered by ChatGPT alone. The final list above combines all the AIs.

  1. 1
    Qwen3.8 Max (2026)

    86.1% leads the current OSWorld-Verified model table—but the published row does not disclose its agent scaffold.

    Source ↗
  2. 2
    Claude Fable 5 (2026)

    85% puts it within 1.1 points of the top score; the benchmark table attributes the result to Anthropic.

    Source ↗
  3. 3
    Claude Mythos 5 (2026)

    85% ties the runner-up; the published figures do not establish which agent setup produced the result.

    Source ↗
  4. 4
    Qwen3.8-27B (2026)

    84.3% is less than two points behind the leading Qwen3.8 Max result.

    Source ↗
  5. 5
    Claude Opus 4.8 (2026)

    83.4% on the shorter benchmark contrasts with only 20.6% full completion on OSWorld 2.0.

    Source ↗
  6. 6
    Gemini 3.6 Flash (2026)

    83% is a provider-reported result; the table does not supply enough setup detail for a like-for-like product comparison.

    Source ↗
  7. 7
    Holo3-35B-A3B (2026)

    82.6% is the strongest independently benchmark-run result among these top ten table entries; it is a specialized computer-use model.

    Source ↗
  8. 8
    MiMo-V2.6-Pro (2026)

    82% keeps it in the top eight, but the result is provider-reported rather than a disclosed end-user agent trial.

    Source ↗
  9. 9
    Claude Sonnet 5 (2026)

    81.2% places it ahead of the next model by 0.4 points, but the reported row omits its agent scaffold.

    Source ↗
  10. 10
    MiMo-V2.6-Flash (2026)

    80.8% rounds out the top ten, 5.3 points below the top result.

    Source ↗

What ChatGPT noticed

The 82.6% Holo3-35B-A3B row is marked as run by OSWorld-Verified; most higher-scoring rows are provider-reported, so the top-line ordering is not a uniform independent bake-off.

macOS 27.3 adds a per-request permission option and splits Accessibility from System & App Protection Override; Apple says previously granted app permissions remain in effect.

🇺🇸 Meta10 places · 3 notes

The same question, answered by Meta alone. The final list above combines all the AIs.

  1. 1
    Claude Cowork (Anthropic, 2026)

    85.96% OSWorld-Verified for Fable 5 and 81.8% OSWorld 2.1 partial for Opus 5.5, with cloud-sandbox default, not full-disk by default.

    Source ↗
  2. 2
    OpenAI ChatGPT agent (GPT-5.6 Sol, 2026)

    78.7% OSWorld-Verified for GPT-5.5 and 62.6% OSWorld 2.0 partial for GPT-5.6 Sol, while the agent runs in an OpenAI sandbox and does not see your desktop by default.

    Source ↗
  3. 3
    Simular Sai (Agent S, 2026)

    73% partial on OSWorld 2.0 but only 28.25% binary - the widest partial-versus-finished gap in this ranking.

    Source ↗
  4. 4
    H Company Holo4 computer-use agent (2026)

    85.2% on OSWorld for 27B at $0.08 per task - the best score-per-dollar here, but the 27B variant is research-only.

    Source ↗
  5. 5
    Google Gemini computer use for Mac (Gemini 3.6 Flash, 2026)

    83.0% OSWorld-Verified, launched November 19, 2026 - the same day Apple’s own Mac agent was reported.

    Source ↗
  6. 6
    Meta Muse / metacua (Muse Spark 1.1, 2026)

    80.67% verified on OSWorld-Verified, but it drives your actual machine - the access cost is the whole machine, not a sandbox.

    Source ↗
  7. 7
    Microsoft 365 Copilot Researcher with Computer Use (2026)

    The cleanest access model here - a fresh sandbox per conversation - but no reproducible OSWorld score to weigh against it.

    Source ↗
  8. 8
    Manus 2.0 (2026)

    A fully isolated cloud VM per task with zero-trust claims - strong sandbox, no published OSWorld score to prove the work.

    Source ↗
  9. 9
    Perplexity Comet Personal Computer (2026)

    Asks for Accessibility, Screen Recording and Full Disk Access on your real Mac - the highest local permission ask - with no OSWorld score and two public security incidents.

    Source ↗
  10. 10
    OpenClaw (formerly Moltbot / Clawdbot, 2026)

    Maximum access, no benchmark: Full Disk Access to Terminal plus Accessibility plus Screen Recording, with no published OSWorld score to justify it.

    Source ↗

What Meta noticed

Gemini computer use for Mac launched on November 19, 2026, the same day TechCrunch reported Apple is building its own Mac-using AI agent that could launch as soon as end of 2027.

macOS 27.3 splits Device Control and Data Access into System and App Protection Override and Accessibility, adds per-folder Full Disk Access, an ask-every-time option, and a menu-bar indicator when one app controls another.

The verified OSWorld board’s top entry is not a consumer model: Intelligence-Indeed Agent scored 90.19% verified, ahead of Claude Fable 5 at 85.96%, showing how much scaffold and framework add on top of a model.

🇺🇸 Grok10 places · 1 notes

The same question, answered by Grok alone. The final list above combines all the AIs.

  1. 1
    Intelligence-Indeed Z-Agent / InAgent (2026)

    First verified agent to clear 90% on OSWorld-Verified (90.19%).

    Source ↗
  2. 2
    Claude Fable 5 / Mythos 5 computer-use (Anthropic, 2026)

    85% Verified, near the top while Anthropic also leads long-horizon partial scores.

    Source ↗
  3. 3
    Qwen3.8 Max computer-use setup (Alibaba, 2026)

    Leads many aggregators at 86.1% Verified with open weights.

    Source ↗
  4. 4
    Simular Sai (2026)

    Highest reported partial on OSWorld 2.0/2.1 (~73–79%) at lower cost than pure frontier models.

    Source ↗
  5. 5
    Claude Opus 5 / 5.5 computer-use (Anthropic, 2026)

    Top binary/partial on long-horizon OSWorld 2.0/2.1 (up to 44% binary, 81.8% partial).

    Source ↗
  6. 6
    Coasty CUA (2026)

    82.81% official Verified entry; emphasizes real desktop control.

    Source ↗
  7. 7
    Pointer Agent + Opus 4.7 (2026)

    83.64% Verified with hybrid coding + GUI approach.

    Source ↗
  8. 8
    OpenAI ChatGPT Agent / Operator lineage (2025–2026)

    Early 38.1% OSWorld; later sandboxed virtual computer reduces local full-control needs.

    Source ↗
  9. 9
    Microsoft Mage-VL-Cowork 4B (2026)

    66.0% Verified at 4B scale—best reported open compact hybrid GUI+shell agent.

    Source ↗
  10. 10
    Claude Code computer-use (Anthropic, 2026)

    Session-scoped Accessibility + Screen Recording with per-app approval prompts.

    Source ↗

What Grok noticed

Agents now routinely exceed the ~72% human baseline on OSWorld-Verified, yet binary completion on OSWorld 2.0 long-horizon tasks stays under 50% even for leaders.

🇺🇸 Gemini10 places · 1 notes

The same question, answered by Gemini alone. The final list above combines all the AIs.

  1. 1
    Anthropic Claude 3.7 Sonnet Computer Use (2025)

    Delivers 44.2% on OSWorld-Verified while offering strict Dockerized API sandboxing and human-in-the-loop per-action approval hooks.

  2. 2
    ByteDance UI-TARS-2 (2025)

    Outperforms closed models with a 53.1% OSWorld-Verified score and runs locally on consumer hardware without granting root remote-cloud telemetry.

    Source ↗
  3. 3
    Agent S2.5 w/ o3 (2025)

    Leverages reasoning tokens to reach 56.4% on verified desktop tasks, but requires invasive Accessibility and raw file-system read/write permissions.

  4. 4
    OpenAI Operator (2025)

    Scores 46.8% on desktop browser-assisted OS tasks while restricting its blast radius primarily to browser automation and isolated VM tabs.

  5. 5
    Coasty Autonomous Agent (2025)

    Hits 42.1% OSWorld task completion with built-in remote browser and desktop sandboxing that isolates your local OS from arbitrary clicks.

  6. 6
    OS-Copilot / FRIDAY (2024)

    Pioneered self-learning desktop agents with 35.6% on OSWorld, but demands unconditional Bash root shell and Full Disk Access.

  7. 7
    Simular AI Agent (2025)

    Achieves 38.3% on verified benchmarks using native multi-modal GUI coordinates with a strict prompt-before-execute confirmation threshold.

  8. 8
    OpenAdapt Desktop Automation (2024)

    Records screen states and user demonstrations to reach 29.8% on desktop workflows while running transparently via open-source local logs.

    Source ↗
  9. 9
    Adept ACT-1 / Workday Agent (2024)

    Pioneered enterprise software UI actions with 31.2% verified task rates, confined inside browser DOM wrappers rather than raw OS level control.

  10. 10
    Apple Prototype Mac Agent (2026)

    Targets system-level Siri and App Intents integration to run cross-app Mac tasks under Apple's native macOS 27.3 security guardrails.

    Source ↗

What Gemini noticed

Apple's upcoming macOS 27.3 update introduces strict warning dialogues whenever third-party agents request simultaneous Screen Capture and Accessibility override privileges, directly challenging autonomous desktop wrappers.

🇨🇳 MiMo10 places · 4 notes

The same question, answered by MiMo alone. The final list above combines all the AIs.

  1. 1
    Claude Code (Anthropic, 2026)

    The only agent here that runs its commands in a cage — and the cage cut permission prompts 84%.

    Source ↗
  2. 2
    Gemini for Mac computer use (Google, 2026)

    Late to the table — and the only guest that shows its work before it touches an app.

    Source ↗
  3. 3
    GPT-6.2 Astra computer use (OpenAI, 2026)

    The strongest new computer-use brain at one-third the cost — with its own safety card listing workplace mishaps.

    Source ↗
  4. 4
    Claude Cowork computer use (Anthropic)

    Top-tier scores — and an honest footnote in its own docs: no sandbox between the agent and your screen.

    Source ↗
  5. 5
    Qwen3.8 Max (Alibaba, 2026)

    Tops the desktop scoreboard at 86.1% — listed as open while the tested visual model isn't.

    Source ↗
  6. 6
    Claude Opus 5 computer use (Anthropic)

    Same model, same day: 18.8% or 44.3% of a long job finished, depending on how hard you tell it to try.

    Source ↗
  7. 7
    Simular Sai (Simular AI, 2026)

    Sold by the job: $15.70 per finished task — the only agent with a price tag on the result.

    Source ↗
  8. 8
    MiMo-V2.6 (Xiaomi, 2026)

    Open weights at 82% — near the commercial leaders without the lock-in.

    Source ↗
  9. 9
    Apple Mac agent (in development, 2027)

    The gatekeeper's own agent: it will help write the permission rules it then plays by.

    Source ↗
  10. 10
    Intelligence-Indeed Agent (2026)

    The scoreboard's mystery champion: 90.19%, above the human baseline — and almost nobody can name it.

    Source ↗

What MiMo noticed

Agents have beaten humans on the 369-task desktop test (human baseline 72.36% vs top score 90.19%) — yet the best machine finishes fewer than half of the long real jobs in OSWorld 2.0 binary terms.

Effort settings move results more than model choice: Opus 5 goes from 18.81% to 44.33% binary on the same benchmark just by turning effort from low to max.

The agent with the best access story runs in a sandbox and cut permission prompts 84%; the one with the best scores states plainly it has no sandbox at all.

Apple's new macOS release treats AI agents as the risk to design against — splitting disk permissions — while Apple builds its own agent for 2027.

🇨🇳 Kimi10 places · 4 notes

The same question, answered by Kimi alone. The final list above combines all the AIs.

  1. 1
    Claude Computer Use (Claude Opus 5.5, Anthropic)

    The best long-horizon desktop result on record — 48.7% binary on OSWorld 2.0 — delivered from inside a VM you supply, so it never touches your Mac's permission tiers.

    Source ↗
  2. 2
    Qwen3.8 Max (Alibaba)

    86.1% on OSWorld-Verified — and the weights are downloadable (a 27B checkpoint hits 84.3%), so the top desktop result needs zero access to your machine.

    Source ↗
  3. 3
    Gemini Computer Use (Google, Gemini 3.x)

    78.4% OSWorld-Verified with 'full: you run the environment' — the agent ships with no Mac permission sheet because the sandbox is yours.

    Source ↗
  4. 4
    ChatGPT Computer Use (OpenAI, GPT-5.4)

    The most capable native agent on your own Mac (75% OSWorld-Verified) is also the neediest: Screen Recording, Accessibility, per-app approvals — plus an opt-in Locked Use mode that can temporarily unlock the Mac.

    Source ↗
  5. 5
    Kimi K2.6 (Moonshot AI)

    73.1% OSWorld-Verified in an open-weight model — strong enough to matter, and free to run inside a disposable VM.

    Source ↗
  6. 6
    Claude Cowork / Claude desktop computer use (Anthropic)

    The consumer packaging of the same stack: it asks per app, advises against banking or health apps, and reaches your computer only while the desktop app is open.

    Source ↗
  7. 7
    Agent S3 (Simular AI)

    69.9% on OSWorld — within 2.5 points of the human baseline — from an open framework whose Behavior Best-of-N buys the points with extra rollouts, not extra permissions.

    Source ↗
  8. 8
    UI-TARS-2 (ByteDance)

    47.5% OSWorld with a privacy twist: the 7B variant runs on a single consumer GPU, so your screen never has to leave the machine.

    Source ↗
  9. 9
    OpenCUA-72B (HKU & Moonshot AI)

    45.0% OSWorld-Verified, the open-source SOTA — and because code, data and models are all public, the auditability is the access story.

    Source ↗
  10. 10
    Intelligence-Indeed Agent

    The 90.19% at the top of the official OSWorld-Verified board — a number almost nobody outside its own submission can reproduce.

    Source ↗

What Kimi noticed

The two OSWorld numbers describe different products: the same Claude Opus 4.8 scores 83.5% on OSWorld-Verified but only 20.6% binary completion on OSWorld 2.0's 108 long-horizon workflows — short-task demos are nearly solved, workday-length tasks are not.

macOS 27.3 splits the old Device Control & Data Access toggle into 'System & App Protection Override' and Accessibility, adds an 'Ask First' tier, per-app scoping, explicit 'Read, Change, Delete, and Export All Your Data' warnings, and a menu-bar indicator — but every permission granted before 27.3 stays in effect.

The state of the art in desktop control is downloadable: Alibaba open-sourced a Max-class checkpoint (a 27B variant scores 84.3% on OSWorld-Verified), making a fully sandboxed, self-hosted top-tier agent possible for the first time.

OSWorld-Verified is saturating — the top models cluster within ~1.1 points, above the 72.36% human baseline, which is exactly why OSWorld 2.0 exists and why agents are judged on access breadth rather than the last benchmark point.

🇨🇳 DeepSeek10 places · 4 notes

The same question, answered by DeepSeek alone. The final list above combines all the AIs.

  1. 1
    OpenAI Operator (2026)

    Delivers 38.1% on OSWorld from a cloud browser sandbox—almost no local disk or Accessibility permission needed.

    Source ↗
  2. 2
    Manus (Meta) desktop agent (2026)

    Runs in a cloud VM with root access, so its 72.5% OSWorld score comes without handing over your Mac's disk.

  3. 3
    Claude Computer Use (Claude Fable 5 / Opus 5 harness, 2026)

    85% on OSWorld-Verified with just Accessibility and Screen Recording—no Full Disk Access—and a per-app override tier in macOS 27.3.

    Source ↗
  4. 4
    Coasty CUA v1 (2026)

    82.81% verified on OSWorld-Verified while asking for screen recording and accessibility—not full disk access.

    Source ↗
  5. 5
    Qwen3.8 Max agent (Alibaba, 2026)

    86.1% on OSWorld-Verified—the highest listed—but the agent's local permission footprint is not publicly detailed.

    Source ↗
  6. 6
    Agent S3 (Simular, 2026)

    72.6% on OSWorld—above the human baseline—in an open-source harness that runs with Accessibility and Screen Recording.

    Source ↗
  7. 7
    UiPath Screen Agent (Claude Opus 4.5, 2026)

    67.1% on OSWorld-Verified in January 2026—respectable, but it asks for full UI automation control of the desktop.

  8. 8
    Gemini 3.6 Flash agent (Google, 2026)

    83% on OSWorld-Verified, but the access profile of the agent harness is not publicly documented.

    Source ↗
  9. 9
    Kimi K3 agent (Moonshot AI, 2026)

    84.8% on OSWorld-Verified as of November 8, 2026—high task success, but the local permission ask is unclear.

  10. 10
    UI-TARS-2 (ByteDance, 2026)

    47.5% on OSWorld and 50.6% on WindowsAgentArena—native GUI control, but it needs screen and input permissions.

What DeepSeek noticed

The access-for-results spread is extreme: Operator gives up 47 percentage points of OSWorld success compared with Claude Fable 5 (38.1% vs 85%) but never touches your local disk, while Claude needs only Accessibility and Screen Recording—not Full Disk Access—to reach 85%.

OSWorld 2.0 cuts everyone down: the best binary completion is 44.33% (Claude Opus 5, max effort, batch tool), even though the same family scores 85% on OSWorld-Verified. The longer horizon is a different game.

macOS 27.3 makes the access question explicit: Full Disk Access is being split and scoped, and System & App Protection Override gets its own tier with a menu-bar indicator—so an agent that asks for 'just' Accessibility is now visibly different from one that wants override control.

Coasty's 82.81% verified score is higher than most commercial agents and close to the top model scores, but it has no published OSWorld 2.0 row—showing how unevenly the long-horizon benchmark is being adopted.

🇨🇳 Wenxin10 places · 1 notes

The same question, answered by Wenxin alone. The final list above combines all the AIs.

  1. 1
    Apple In-House macOS AI Agent (Early 2027 Build)

    Zero third-party data exfiltration by design, native sandboxed operation, no external permissions required outside Apple's own permission model

    Source ↗
  2. 2
    OpenAI Agents API (2026 Beta)

    74% overall developer adoption rate, full sandbox support with OpenAI-hosted or self-hosted runtime options

  3. 3
    Perplexity Computer (2026)

    100+ parallel sub-agent orchestration, isolated compute environment with real filesystem and browser, no full desktop root access required

  4. 4
    Sai AI Agent (2026)

    GUI automation cost drops 40% on average after 3 workflow runs, free plan with daily limited cloud Windows/Linux sandbox credits

  5. 5
    Anthropic Claude Computer Use Agent (2026)

    75% overall developer usage tied with OpenAI, 49% of developers plan to work with it next year, the highest forward intent in the industry

  6. 6
    Meta Muse Agent (2026)

    Open-source native cross-platform desktop support, zero mandatory cloud data sharing for local task execution

    Source ↗
  7. 7
    OpenAI Codex Agent (2026)

    8% of developers name it their most loved tool, 17.6% plan to use it next, the third highest forward adoption intent globally

  8. 8
    OpenCode (2026)

    The top open-source coding desktop agent, 7% developer most-loved share, 11.7% next-use intent in the open-source developer community

  9. 9
    Perplexity Personal Computer Mac App (2026)

    Extends agent orchestration directly to local Mac hardware, 24/7 operation optimized for Mac mini deployments

  10. 10
    OpenAI Agents SDK (2026 Beta)

    Full custom loop control inside user own application, no mandatory OpenAI hosted runtime lock-in

What Wenxin noticed

Apple's upcoming first-party Mac AI agent will operate natively under macOS 27.3's new permission framework, requiring no permanent full disk, screen recording or accessibility access that all third-party agents currently need to request

🇨🇳 Qwen10 places · 3 notes

The same question, answered by Qwen alone. The final list above combines all the AIs.

  1. 1
    Qwen3.8 Max

    86.1% on OSWorld-Verified, the highest verified desktop score yet, while allowing granular tool-permission controls.

    Source ↗
  2. 2
    Claude Opus 5.5

    The only agent clearing 81% on the grueling OSWorld 2.0 108-task long-horizon suite.

    Source ↗
  3. 3
    Claude Fable 5

    Hits 85.0% on real desktop tasks but relies on full screen-recording and accessibility overrides.

    Source ↗
  4. 4
    Coasty

    A close second at 85.6% verified, specifically optimized to lower the compute and access footprint.

    Source ↗
  5. 5
    Kimi K3

    Secures 84.8% on OSWorld-Verified while offering a browser-sandboxed mode that bypasses full disk access.

    Source ↗
  6. 6
    Claude Fable 5.1

    Scores 77.9% on OSWorld 2.0, proving it can sustain multi-hour workflows when granted full Mac control.

    Source ↗
  7. 7
    Qwen3.8-27B

    The open-weight version hits 84.3% locally, but demands explicit user approval for Accessibility and Full Disk Access.

    Source ↗
  8. 8
    Sai by Simular

    Achieves 73% on OSWorld 2.0, balancing long-horizon professional tasks with strict local sandboxing.

    Source ↗
  9. 9
    GPT-6 Astra

    Manages 72.6% on OSWorld 2.0, but its cloud-first architecture means your local machine only grants screen-share permissions.

    Source ↗
  10. 10
    GPT-6.1 Sol

    Scores 71.4% on OSWorld 2.0, requiring the full macOS 27.3 override tier to navigate complex desktop states.

    Source ↗

What Qwen noticed

Apple's macOS 27.3 now forces AI agents to ask for permission every time they access sensitive folders, breaking the old 'set and forget' full disk access model.

Agents now pass the 72.36% human baseline on OSWorld-Verified, meaning AI is officially faster and more accurate than average humans at routine desktop clicks.

Long-horizon tasks in OSWorld 2.0 average over 250 agent steps, exposing how even top models like GPT-6 Astra struggle to maintain context over an hour of work.