Back to the trend of the day: Google's Gemini can click and type on Macs for paying US subscribers
Poly Rating · from the trend of the day

Gemini Computer Use Price vs Cost per Finished Job (2026) · 8 places

On the hardest public desktop test, one agent try costs anywhere from $2.40 to $76. Once failed tries are counted, a finished job costs about $52 at best and over $400 at worst.

  • No. 1: Gemini computer use on Mac (Google AI Plus, Pro, Ultra)
  • 8 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

    Gemini computer use on Mac (Google AI Plus, Pro, Ultra)

    Cheapest way in, but Google has published no completion rate, so its cost per finished job is unknown. AI Plus is quoted at both $4.99 and $7.99.

    Included it: Claude, ChatGPT, Meta, Grok, StepFun, Z.ai, DeepSeek, Qwen. Did not include: Nemotron, MiMo.

    In their lists: Claude: 2 · ChatGPT: 10 · Meta: 1 · Grok: 4 · StepFun: 1 · Z.ai: 1 · DeepSeek: see list · Qwen: 3

    What each AI noted 8
    • Claudeplace 2

      The cheapest door in: $7.99 a month on AI Plus ($95.88 a year), but no independent table yet says how often it finishes a job.

      Source ↗
    • ChatGPTplace 10

      The new Mac feature has an entry price, but no published completion rate—so its cost per finished job cannot yet be ranked.

      Source ↗
    • Metaplace 1

      $4.99 a month buys the cheapest Mac computer-use entry, a quarter of Claude Pro, but only in US English on a personal account.

      Source ↗
    • Grokplace 4

      $19.99/mo lifts limits and model access while keeping the same local Mac background-window pattern

    • StepFunplace 1

      $4.99 a month buys the only computer-use agent that interrupts you only before the risky clicks.

      Source ↗
    • Z.aiplace 1

      $7.99 buys the same Mac agent Google gives $249.99 Ultra users - computer use is the rare frontier feature that landed at the bottom of the price ladder first.

      Source ↗
    • DeepSeek

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

      see list
    • Qwenplace 3

      The lowest flat entry price at $4.99 a month, but Google publishes zero completion rates, turning every task into a supervision gamble.

      Source ↗
    8 AIs put it here
  2. 2

    Claude Pro or Max with Cowork computer use

    $240 a year buys access to the model family with the best measured finish rates. There is no completion rate for the product itself, and the flat fee holds only until the usage cap.

    Included it: Claude, Meta, Grok, MiMo, StepFun, Z.ai, DeepSeek. Did not include: ChatGPT, Nemotron, Qwen.

    In their lists: Claude: 1 · Meta: 2 · Grok: 2 · MiMo: 1 · StepFun: 2 · Z.ai: 7 · DeepSeek: 9

    What each AI noted 7
    • Claudeplace 1

      $240 a year buys access to the model family with the highest independently measured OSWorld 2.0 completion: Opus 5 finished 44% of tasks where Opus 4.7 finished 14%.

      Source ↗
    • Metaplace 2

      The $20 plan sits on the model family that tops OSWorld, and deletion still always needs approval.

      Source ↗
    • Grokplace 2

      $20/mo Pro includes research-preview computer use on your machine; token burn still adds up on long OSWorld-style jobs

    • MiMoplace 1

      The cheapest honest entry: a subscription you may already pay, and it stops before every key action.

      Source ↗
    • StepFunplace 2

      The most-measured finisher in the benchmark, riding a $20 plan many people already pay for.

      Source ↗
    • Z.aiplace 7

      $3/$15 per million tokens and the best-measured OSWorld score you could buy at launch (61.4%) - the metered route, where every screenshot is a real dollar and every approval is your time.

      Source ↗
    • DeepSeekplace 9

      8.3% binary at $22.30 a task: about $269 per finished job — worse than the same model at medium effort.

      Source ↗
    7 AIs put it here
  3. 3

    ChatGPT Pro agent / ChatGPT Work

    Most expensive way in: Pro starts at $100 a month with agent runs capped at 400. Agent mode was retired in 2026, and sources disagree on which tier keeps computer use.

    Included it: Claude, Meta, Grok, MiMo, StepFun, Z.ai, DeepSeek. Did not include: ChatGPT, Nemotron, Qwen.

    In their lists: Claude: 5 · Meta: 10 · Grok: 6 · MiMo: 5 · StepFun: 3 · Z.ai: 9 · DeepSeek: 3

    What each AI noted 7
    • Claudeplace 5

      $1,200 to $2,400 a year for a model that finishes about 27% of OSWorld 2.0 tasks, with agent runs capped at 400 a month on Pro.

      Source ↗
    • Metaplace 10

      First dot included sounds free, until the $100 Pro gate makes it the dearest entry for always-on work.

      Source ↗
    • Grokplace 6

      $200/mo flat buys hosted browser automation; higher certainty on web jobs, far higher monthly floor than Gemini Plus

    • MiMoplace 5

      OpenAI retired its computer-using product and now sells the paperwork instead — planning over clicking.

      Source ↗
    • StepFunplace 3

      The $20 plan that bought a computer agent — until OpenAI retired the agent that did the clicking.

      Source ↗
    • Z.aiplace 9

      $200 for 400 messages is Plus's math ten times over - $0.50 a message either way, so you bought volume, not cheaper finished jobs.

      Source ↗
    • DeepSeekplace 3

      23.14% binary at $26.62 a task: about $115 per finished job, but the plan that includes agentic computer use starts at $1,200 a year.

      Source ↗
    7 AIs put it here
  4. 4

    Open-weight models in your own agent (MiniMax M3, Qwen 3.7-Plus, Kimi K2.6)

    Cheapest measured finished job: MiniMax M3 costs $2.40 a try but finishes only 4.6%, so about $52 per finish. Qwen and Kimi come to about $136–143.

    Included it: Claude, ChatGPT, Meta, Grok, MiMo, DeepSeek, Qwen. Did not include: Nemotron, StepFun, Z.ai.

    In their lists: Claude: 9 · ChatGPT: 2 · Meta: 6 · Grok: see list · MiMo: 10 · DeepSeek: 5 · Qwen: 7

    What each AI noted 7
    • Claudeplace 9

      The cheapest attempt ($2.40) finishes 1 task in 22, so about $52 per finished task. Nobody puts a price on the 21 failed tries a human has to clean up.

      Source ↗
    • ChatGPTplace 2

      About $136 per finished benchmark job, despite a $3.80 average attempt.

      Source ↗
    • Metaplace 6

      Free tier with a cloud computer is the cheapest way to try delegation, but it is not control of your own Mac.

      Source ↗
    • Grok

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

      see list
    • MiMoplace 10

      Free software, your own machine, and the model bill is the only bill.

      Source ↗
    • DeepSeekplace 5

      $6.60 a task and 4.6% binary: about $143 per finished job, and 21 failures for every completion.

      Source ↗
    • Qwenplace 7

      Costs just $2.40 per attempt on OSWorld, but a 4.6% success rate means the human spends 20 times longer fixing failures than the AI saves.

      Source ↗
    7 AIs put it here
  5. 5

    Simular Sai desktop agent

    About $56 per finish ($15.70 a try, 28.25% finished). That would rank 2nd, but it is self-reported by a benchmark co-author, so it stays below the confirmed rows.

    Included it: Claude, Meta, MiMo, StepFun, DeepSeek. Did not include: ChatGPT, Grok, Nemotron, Z.ai, Qwen.

    In their lists: Claude: 3 · Meta: see list · MiMo: 4 · StepFun: 5 · DeepSeek: 2

    What each AI noted 5
    • Claudeplace 3

      $15.70 an attempt at 28% completion works out to about $56 per finished task, the lowest per-finish figure among agents that finish at least one task in five, though the vendor reported it.

      Source ↗
    • Meta

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

      see list
    • MiMoplace 4

      The only seller that prices the job, not the month: $15.70 per attempted task.

      Source ↗
    • StepFunplace 5

      Free to start with no card — but you review what it ran, and your minutes are the bill.

      Source ↗
    • DeepSeekplace 2

      28.25% binary completion at $15.70 a task — about $56 per finished job, less than a sixth of Claude Opus 4.8's $350.

      Source ↗
    5 AIs put it here
  6. 6

    Perplexity Comet and Perplexity Computer

    The browser is free, but maximum Computer use sits on Max at $200 a month, and one Max user reported a single task using up the month's budget.

    Included it: Claude, Meta, StepFun, Z.ai, Qwen. Did not include: ChatGPT, Grok, Nemotron, MiMo, DeepSeek.

    In their lists: Claude: 8 · Meta: 3 · StepFun: 9 · Z.ai: 5 · Qwen: 10

    What each AI noted 5
    • Claudeplace 8

      The browser is free, but the long-running agent sits behind tiers reported at $200 to $500 a month, and it has no OSWorld 2.0 score.

      Source ↗
    • Metaplace 3

      Free browser entry means $0 per trial, but the agent lives in the browser, not across your whole Mac.

    • StepFunplace 9

      $200 a month for 'Maximum Computer usage' — maximum until one task eats the month.

    • Z.aiplace 5

      $17 a month (annual) is where Perplexity's computer use stops being a sampler - 'Expanded Computer access' plus 4,000 bonus credits, one tier below the $167 Max wall.

    • Qwenplace 10

      Runs fully local on Mac for $5 a month, keeping data private but relying entirely on the user's hardware for long-horizon completion.

      Source ↗
    5 AIs put it here
  7. 7

    Manus cloud computer agent

    From $19–20 a month in credits. Failed runs burn credits too, and there is no OSWorld completion rate to price a finished job.

    Included it: Meta, Grok, StepFun, Z.ai, Qwen. Did not include: Claude, ChatGPT, Nemotron, MiMo, DeepSeek.

    In their lists: Meta: 7 · Grok: 8 · StepFun: 7 · Z.ai: 6 · Qwen: 8

    What each AI noted 5
    • Metaplace 7

      $19 buys a cloud sandbox agent, not Mac control, and one email hijack reached its connected-account tokens in a researcher test.

      Source ↗
    • Grokplace 8

      Cloud VM or local app; credit or subscription models sit between open-source free and $200 Pro tiers

    • StepFunplace 7

      A rented computer with a credit meter: 4,000 credits a month, and the meter runs while you sleep.

      Source ↗
    • Z.aiplace 6

      From $20 for a credit meter that watches every retry - the only plan here that shows you, in credits, exactly what each failure costs.

      Source ↗
    • Qwenplace 8

      Burns credits starting at $20 a month for delegated tasks, with completion rates gated behind a proprietary cloud environment.

      Source ↗
    5 AIs put it here
  8. 8

    Microsoft Copilot computer use (Copilot Studio, Copilot Cowork)

    The only option that bills per step: about $0.04 a step, or $0.16 for a 4-step workflow, on top of a Microsoft 365 Copilot seat. Windows-first.

    Included it: Claude, Meta, MiMo. Did not include: ChatGPT, Grok, Nemotron, StepFun, Z.ai, DeepSeek, Qwen.

    In their lists: Claude: 6 · Meta: 9 · MiMo: 8

    What each AI noted 3
    • Claudeplace 6

      Every task gets its own bill: $0.01 a credit, with light tasks reported at 100–300 credits, charged on top of a Microsoft 365 Copilot seat.

      Source ↗
    • Metaplace 9

      Metered at about $0.04 a step, a 4-step workflow is about $0.16, the clearest per-job price here, but it is Windows-centric and enterprise-shaped.

    • MiMoplace 8

      Priced per seat at $21 and aimed at the work apps — the office route to the same hands.

      Source ↗
    3 AIs put it here

02Worth knowing

What the AIs noticed while building the list.

The cheapest try is not the cheapest finish: MiniMax M3 costs $2.40 a try but needs about 22 tries per success; Opus 4.8 costs $72.40 a try and ends near $351 per finished job.

Noticed by:
Included it: Claude, ChatGPT, DeepSeek, Qwen. Did not include: Meta, Grok, Nemotron, MiMo, StepFun, Z.ai.
Source ↗

More thinking can mean fewer finished jobs: Sonnet 4.6 at max effort finished 8.3% at $22.30 a task; at medium effort it finished 9.3% at $14.40.

Noticed by:
Included it: Claude, DeepSeek. Did not include: ChatGPT, Meta, Grok, Nemotron, MiMo, StepFun, Z.ai, Qwen.
Source ↗

'Mostly done' is not done: the paper's best agent earns 54.8% partial credit but finishes only 20.6% of jobs that take a skilled human about 1.6 hours.

Noticed by:
Included it: ChatGPT, StepFun, Z.ai, DeepSeek. Did not include: Claude, Meta, Grok, Nemotron, MiMo, Qwen.
Source ↗

Approval minutes turn into reflex: people accept 93% of permission prompts, so most supervision time is spent clicking yes.

Noticed by:
Included it: Claude, MiMo. Did not include: ChatGPT, Meta, Grok, Nemotron, StepFun, Z.ai, DeepSeek, Qwen.
Source ↗

An independent audit of 50 OSWorld 2.0 tasks found scoring or task errors in 36, a caveat on reading benchmark completion as real-world job success.

Noticed by:
Included it: ChatGPT. Did not include: Claude, Meta, Grok, Nemotron, MiMo, StepFun, Z.ai, DeepSeek, Qwen.
Source ↗

03Questions people ask

Is Gemini computer use free?

No. Computer use on Mac needs a paid Google AI Plus, Pro or Ultra plan in the US. AI Plus is quoted at $4.99 a month in launch coverage and $7.99 in Google's I/O post.

Source ↗
What is Gemini computer use?

A Gemini for Mac feature, launched November 19, 2026, that clicks, types and fills in forms across apps, files and websites in a background window. You choose the apps and approve key actions.

Source ↗
What is the Gemini computer use model?

On Mac, Google recommends Gemini Hydrogen 4 or newer. For developers, Gemini 2.5 Computer Use (October 2025) was a separate API model; one source says it was shut down on July 28, 2026.

Source ↗
Is Gemini 2.5 computer use free?

No free tier was found: it was billed at $1.25 per million input tokens and $10 per million output tokens (prompts up to 200K).

Source ↗

04How we counted

In plain words, no formulas.

1The question

Top 10 ways to buy an AI that works your computer, ranked by cost per job it actually finishes

2Why this topic

Buyers see that the sticker price is the smallest line in the bill: completion rates, retries and approvals vary far more between products than subscriptions do.

3Where it applies

Global; list prices in US dollars

4What we looked at

  • What a year of subscription buys once failures and retries are counted
  • Human supervision still required before key actions (every command, each app, or sensitive steps only)
  • Where the compute runs: your machine or the vendor's cloud
  • OSWorld 2.0 paper (XLANG Lab, University of Hong Kong) — 108 long desktop jobs: compute dollars per attempt, binary completion (whole job done) and partial credit
  • Snorkel AI OSWorld 2.0 leaderboard — Independent runs with effort level, binary %, partial % and estimated cost per task
  • Steel.dev OSWorld 2.0 leaderboard — Vendor-reported and author-run scores side by side, with self-reported results flagged; cost per task or per run
  • Google: Gemini computer use on Mac announcement — Which paid tiers include it, where it runs, the approval pattern, country and language limits
  • AgentStackWatch price compilation — US list prices and plan terms for personal AI agents, linked to vendor pages

5Who ranked

The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT, Meta, Grok, Nemotron, MiMo, StepFun, Z.ai, DeepSeek, 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 computer use on Mac and Windows (Claude Pro, $20/month, Cowork and Claude Code)

    $240 a year buys access to the model family with the highest independently measured OSWorld 2.0 completion: Opus 5 finished 44% of tasks where Opus 4.7 finished 14%.

    Source ↗
  2. 2
    Gemini computer use on Mac (Google AI Plus or Pro, 2026)

    The cheapest door in: $7.99 a month on AI Plus ($95.88 a year), but no independent table yet says how often it finishes a job.

    Source ↗
  3. 3
    Simular Sai (Simular AI desktop agent)

    $15.70 an attempt at 28% completion works out to about $56 per finished task, the lowest per-finish figure among agents that finish at least one task in five, though the vendor reported it.

    Source ↗
  4. 4
    Claude Opus 5 computer use via API (pay per task)

    No subscription and no cap: about $28 per attempt, and roughly one success in every two to three tries.

    Source ↗
  5. 5
    ChatGPT agent / ChatGPT Work (ChatGPT Pro, $100–$200/month)

    $1,200 to $2,400 a year for a model that finishes about 27% of OSWorld 2.0 tasks, with agent runs capped at 400 a month on Pro.

    Source ↗
  6. 6
    Microsoft Copilot Cowork (pay-as-you-go Copilot Credits)

    Every task gets its own bill: $0.01 a credit, with light tasks reported at 100–300 credits, charged on top of a Microsoft 365 Copilot seat.

    Source ↗
  7. 7
    Gemini computer use on Mac via Google AI Ultra ($99.99 or $199.99/month)

    Same agent as the $7.99 tier at 12 to 25 times the price. It pays off only if you hit the lower tiers' limits every month.

    Source ↗
  8. 8
    Perplexity Comet browser and Perplexity Computer

    The browser is free, but the long-running agent sits behind tiers reported at $200 to $500 a month, and it has no OSWorld 2.0 score.

    Source ↗
  9. 9
    Open-weight models in a self-run agent (MiniMax M3, Kimi K2.6, Qwen 3.7-Plus; UI-TARS Desktop)

    The cheapest attempt ($2.40) finishes 1 task in 22, so about $52 per finished task. Nobody puts a price on the 21 failed tries a human has to clean up.

    Source ↗
  10. 10
    Gemini 2.5 Computer Use model (Gemini API, developers)

    Billed by the token ($1.25 in, $10 out per million), and you run the sandbox and the approvals yourself. Injection screening is an opt-in flag.

    Source ↗

What Claude noticed

The cheapest attempt is not the cheapest finish. MiniMax M3 costs $2.4 a task but finishes 4.6%, about 22 tries per success; Claude Opus 5 costs about 12 times more per attempt but finishes up to 44%.

More thinking can mean fewer finished jobs. Claude Sonnet 4.6 at max thinking finished 8.3% at $22.3 a task; at medium thinking it finished 9.3% at $14.4.

'Mostly done' is not done. Claude Opus 5 scores 77.67% partial credit but 44.33% binary, so more than half its tasks end unfinished despite visible progress.

Human approval minutes turn into a reflex. Anthropic found users accept 93% of permission prompts anyway, so most supervision time is spent clicking yes.

🇺🇸 ChatGPT10 places · 3 notes

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

  1. 1
    Kimi 2.6 computer-use API stack

    The cheapest measured finish is still about $52—not $2.40—because 95.4% of attempts did not finish.

    Source ↗
  2. 2
    Qwen 3.7-Plus computer-use API stack

    About $136 per finished benchmark job, despite a $3.80 average attempt.

    Source ↗
  3. 3
    MiniMax M3 computer-use API stack

    A $6.60 attempt becomes roughly $143 per finished job at a 4.6% completion rate.

    Source ↗
  4. 4
    Claude Opus 4.7 computer use, batched actions

    Batching lowers the measured cost to about $185 per completed benchmark job, versus about $258 without batching.

    Source ↗
  5. 5
    GPT-5.5 computer-use API stack

    It uses far fewer output tokens than Opus, but its lower completion rate leaves the estimated finished-job cost near $196.

    Source ↗
  6. 6
    Claude Opus 4.7 computer use, single actions

    Turning batching off raises the estimated price per finished job by about $73 versus its batched result.

    Source ↗
  7. 7
    Claude Sonnet 4.6 computer-use API stack

    Its $22.30 attempt looks modest until the 8.3% completion rate is counted.

    Source ↗
  8. 8
    Claude Opus 4.8 computer use, batched actions

    The benchmark leader finishes one in five jobs, but its $72.40 attempt averages about $351 per finish.

    Source ↗
  9. 9
    Claude Opus 4.8 computer use, single actions

    The non-batched run costs more per attempt and completes fewer jobs than the batched configuration.

    Source ↗
  10. 10
    Gemini computer use for Mac (US, 2026)

    The new Mac feature has an entry price, but no published completion rate—so its cost per finished job cannot yet be ranked.

    Source ↗

What ChatGPT noticed

Dividing measured attempt cost by completion changes the apparent bargain: Kimi’s $2.40 attempt estimates to about $52 per finish, while the top-scoring Opus 4.8 batched run estimates to about $351.

OSWorld 2.0’s median task takes a skilled human about 1.6 hours; the benchmark’s strongest agent completes 20.6% of tasks, despite earning 54.8% partial credit.

OSWorld-Human reports 1.4–2.7 times more steps than necessary for leading agents. Approval time is not included in the cost estimates here, and no open study found measures the minutes spent on production approvals.

🇺🇸 Meta10 places · 3 notes

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

  1. 1
    Gemini computer use on Mac (2026, US)

    $4.99 a month buys the cheapest Mac computer-use entry, a quarter of Claude Pro, but only in US English on a personal account.

    Source ↗
  2. 2
    Claude Cowork (2026)

    The $20 plan sits on the model family that tops OSWorld, and deletion still always needs approval.

    Source ↗
  3. 3
    Perplexity Comet (2026)

    Free browser entry means $0 per trial, but the agent lives in the browser, not across your whole Mac.

  4. 4
    OpenClaw (2026)

    $0 software shifts the bill to tokens, so a cheap model can beat a $20 subscription and a frontier model can blow past $100.

    Source ↗
  5. 5
    Fellou (2026)

    Credits make the math visible: Plus is about $2.38 per attempted task at 2,000 Sparks for $19.

    Source ↗
  6. 6
    Meta Muse (2026)

    Free tier with a cloud computer is the cheapest way to try delegation, but it is not control of your own Mac.

    Source ↗
  7. 7
    Manus (2026)

    $19 buys a cloud sandbox agent, not Mac control, and one email hijack reached its connected-account tokens in a researcher test.

    Source ↗
  8. 8
    Dia (2026)

    $10 is the cheapest paid AI browser, but chat-first scope means many desktop jobs are not finished at all.

  9. 9
    Microsoft Copilot Studio computer-use agents (2026)

    Metered at about $0.04 a step, a 4-step workflow is about $0.16, the clearest per-job price here, but it is Windows-centric and enterprise-shaped.

  10. 10
    OpenAI ChatGPT Dots (2026)

    First dot included sounds free, until the $100 Pro gate makes it the dearest entry for always-on work.

    Source ↗

What Meta noticed

A $72.40 OSWorld 2.0 attempt at 20.6% completion is about $351 per finished job, which is why a cheap subscription can still be expensive per job.

The later OSWorld 2.0 leaderboard lists Claude Opus 5 at 44.3% binary completion, more than double the paper's 20.6% top, showing how fast this number moves.

Gemini on Mac starts at $4.99, yet it is US-only, English-only and personal-account-only, so global buyers cannot use the cheapest headline price.

🇺🇸 Grok10 places

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

  1. 1
    Gemini computer use (Google AI Plus, Mac, 2026)

    $4.99/mo buys background clicks on your own Mac with key-action approval; lowest published entry for a finished-job path that stays local

  2. 2
    Claude computer use (Anthropic Pro / Max, desktop)

    $20/mo Pro includes research-preview computer use on your machine; token burn still adds up on long OSWorld-style jobs

  3. 3
    Open-source local agents (UI-TARS Desktop, Agent S3, Goose, Browser Use)

    Software free; you pay only local compute or cheap API tokens, but binary completion on long-horizon tasks stays well below frontier closed models

  4. 4
    Gemini computer use (Google AI Pro)

    $19.99/mo lifts limits and model access while keeping the same local Mac background-window pattern

  5. 5
    Claude API computer-use tool (pay-per-token)

    No flat fee, but OSWorld totals show high-effort runs easily $20–$76 per attempted long job before failures

  6. 6
    OpenAI Operator / ChatGPT Pro computer-use path

    $200/mo flat buys hosted browser automation; higher certainty on web jobs, far higher monthly floor than Gemini Plus

  7. 7
    Google AI Ultra + Gemini computer use

    $99.99–$199.99/mo for highest limits; still the same local Mac agent, just more headroom

  8. 8
    Manus / cloud-sandbox computer-use agents

    Cloud VM or local app; credit or subscription models sit between open-source free and $200 Pro tiers

  9. 9
    Self-hosted frameworks + frontier API (CrewAI, open agents + Claude/GPT)

    Software free; API + machine time still produces OSWorld-scale costs once retries are counted

  10. 10
    High-effort frontier API runs alone (no subscription wrapper)

    Pure token bills on OSWorld 2.0 can exceed $50–$70 per attempted long workflow before any human rescue minutes

🇨🇳 MiMo10 places · 4 notes

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

  1. 1
    Gemini computer use on Mac (Google AI Plus/Pro/Ultra)

    The cheapest honest entry: a subscription you may already pay, and it stops before every key action.

    Source ↗
  2. 2
    Claude Code with sandboxing (Anthropic)

    The only setup that measured its own supervision cost — and cut permission prompts 84%.

    Source ↗
  3. 3
    Meta Muse on Mac (free)

    Zero dollars and it runs on its own secure computer, not yours — the free tier of hands.

    Source ↗
  4. 4
    Simular Sai agent (per-task pricing)

    The only seller that prices the job, not the month: $15.70 per attempted task.

    Source ↗
  5. 5
    ChatGPT Work (OpenAI)

    OpenAI retired its computer-using product and now sells the paperwork instead — planning over clicking.

    Source ↗
  6. 6
    GPT-6.1 Sol via API (OpenAI)

    Keeps within 2.1 points of the strongest desktop brain at one-seventh the cost per task.

    Source ↗
  7. 7
    Grok Bot (xAI)

    It keeps its own cloud computer so the chore runs while your laptop is shut.

    Source ↗
  8. 8
    Microsoft 365 Copilot computer use (Windows)

    Priced per seat at $21 and aimed at the work apps — the office route to the same hands.

    Source ↗
  9. 9
    Holo4 desktop agent (H Company)

    Publishes its cost per task beside its scores — rare transparency in the cost column.

    Source ↗
  10. 10
    OpenClaw / Hermes Agent (open source)

    Free software, your own machine, and the model bill is the only bill.

    Source ↗

What MiMo noticed

The 'cost per finished job' flips several price tags: at $15.70 per attempt and 28.25% completion, Simular's effective cost per completed job is about $56 — while a 'free' subscription that needs you at every approval pays you in minutes instead.

The biggest hidden cost is supervision: real users approve 93% of the permission prompts they see, so an agent that asks constantly gets a rubber stamp, not a review.

Long jobs are still mostly failures: on OSWorld 2.0 the best system in the paper fully completes only 20.6% of tasks that take a skilled human over an hour — so any 'per job' math today is really 'per several attempts'.

OpenAI quietly removed the most convenient paid option — ChatGPT agent mode is discontinued — while Google shipped a new one; the price table of computer use is changing faster than the model rankings.

🇨🇳 StepFun10 places · 3 notes

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

  1. 1
    Gemini computer use on Mac — Google AI Plus (2026)

    $4.99 a month buys the only computer-use agent that interrupts you only before the risky clicks.

    Source ↗
  2. 2
    Claude Cowork — Claude Pro (2026)

    The most-measured finisher in the benchmark, riding a $20 plan many people already pay for.

    Source ↗
  3. 3
    ChatGPT Work — ChatGPT Plus (2026)

    The $20 plan that bought a computer agent — until OpenAI retired the agent that did the clicking.

    Source ↗
  4. 4
    Perplexity Comet assistant — Perplexity Pro (2026)

    Browser-scope help for $20, while the full cloud Computer agent is a $200-tier extra.

  5. 5
    Cyborb desktop agent — free plan (2026)

    Free to start with no card — but you review what it ran, and your minutes are the bill.

    Source ↗
  6. 6
    Sai by Simular — OSWorld 2.1 cost leader (2026)

    The only agent that publishes its own per-task bill: $14.34 an attempt, 41.7% of tasks finished.

  7. 7
    Manus Standard — cloud computer agent (2026)

    A rented computer with a credit meter: 4,000 credits a month, and the meter runs while you sleep.

    Source ↗
  8. 8
    The $100 tier — Claude Max / Google AI Ultra (2026)

    A hundred dollars a month doesn't cheapen a job; it just shortens the queue.

    Source ↗
  9. 9
    Perplexity Max with Computer (2026)

    $200 a month for 'Maximum Computer usage' — maximum until one task eats the month.

  10. 10
    Computer-use APIs — Claude and Gemini pay-per-token (2026)

    The meter runs per attempt and most attempts fail: up to $76 and 50 minutes for one run.

    Source ↗

What StepFun noticed

The entry price collapsed below $5: computer use now ships on Google's cheapest paid tier at $4.99/mo, while the best-measured desktop agent still costs $14.34 to $76 per attempt.

The approval pattern is a hidden meter: agents are already estimated at $6-8/hour of working time, and asking permission before every command adds minutes that dwarf the token bill.

Two flagship computer-use products were retired in 2026 — OpenAI's agent mode and the Atlas browser — so a $20 subscription can stop finishing jobs without any price change.

🇨🇳 Z.ai10 places · 4 notes

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

  1. 1
    Google AI Plus with Gemini computer use on Mac (2026)

    $7.99 buys the same Mac agent Google gives $249.99 Ultra users - computer use is the rare frontier feature that landed at the bottom of the price ladder first.

    Source ↗
  2. 2
    Perplexity Comet, free browser assistant (2026)

    $0 down and it drives the browser on your own machine - but the 'advanced models powering Comet Assistant' are a $167-$200 Max feature, so the free floor is also the gamble.

  3. 3
    Google AI Pro (2026)

    At $20 with 'four times more access' than Plus, you pay 2.5x Plus's price for the identical Mac agent - limits, not intelligence, are the upsell.

    Source ↗
  4. 4
    ChatGPT Plus with agent mode (2026)

    $20 buys 40 agent messages a month; at the ~13% long-horizon success OpenAI-class agents plateau at, that is roughly five jobs that actually finish - about $3.85 each (estimate).

    Source ↗
  5. 5
    Perplexity Pro with Perplexity Computer (2026)

    $17 a month (annual) is where Perplexity's computer use stops being a sampler - 'Expanded Computer access' plus 4,000 bonus credits, one tier below the $167 Max wall.

  6. 6
    Manus Pro, cloud computer agent (2026)

    From $20 for a credit meter that watches every retry - the only plan here that shows you, in credits, exactly what each failure costs.

    Source ↗
  7. 7
    Claude computer use API with Sonnet 4.5 (2025)

    $3/$15 per million tokens and the best-measured OSWorld score you could buy at launch (61.4%) - the metered route, where every screenshot is a real dollar and every approval is your time.

    Source ↗
  8. 8
    Claude Max with Claude for Chrome (2025-2026)

    The $100-$200 Max tier is Anthropic's retail wrapper around the same computer-use loop - Max users got Claude for Chrome first, with Opus 4.5 measuring just 1% on browser prompt injections.

    Source ↗
  9. 9
    ChatGPT Pro with agent mode (2026)

    $200 for 400 messages is Plus's math ten times over - $0.50 a message either way, so you bought volume, not cheaper finished jobs.

    Source ↗
  10. 10
    Google AI Ultra (2025-2026)

    At $249.99 with five times Pro's limits, Ultra is the best plan to fail with and the worst to underuse - light users pay the most per job that actually finishes.

    Source ↗

What Z.ai noticed

OSWorld 2.0's cost-aware table is the number that turns '$20 a month' into a real price: compute per attempted task spans about $2-$76, and even the best agent completes only 20.6% of long-horizon work at a 54.8% partial score - subscriptions sell attempts, you live on completions.

Quota scaling does not change unit economics: ChatGPT Plus's 40 messages and Pro's 400 both work out to $0.50 per agent message, roughly $3.85 per long-horizon job that actually finishes at 13% success (estimate) - you buy volume, not value.

Google shipped its Mac computer-use agent into every paid tier on day one, from $7.99 Plus to $249.99 Ultra (US, English) - the feature is not the upsell, the limits are.

The metered API route is quietly competitive per finished job: Sonnet 4.5 led OSWorld at 61.4% on $3/$15 tokens - if you can stomach DIY setup and per-action approvals no leaderboard scores.

🇨🇳 DeepSeek10 places · 4 notes

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

  1. 1
    MiniMax M3 API (OSWorld 2.0)

    $2.40 a task sounds cheapest, but 4.6% binary completion means about 22 attempts per finished job — roughly $52 in compute and far more human supervision.

    Source ↗
  2. 2
    Simular Sai (OSWorld 2.0 agent)

    28.25% binary completion at $15.70 a task — about $56 per finished job, less than a sixth of Claude Opus 4.8's $350.

    Source ↗
  3. 3
    GPT-5.6 Sol Max on ChatGPT Pro

    23.14% binary at $26.62 a task: about $115 per finished job, but the plan that includes agentic computer use starts at $1,200 a year.

    Source ↗
  4. 4
    Qwen3 7-Plus API (OSWorld 2.0)

    $3.80 a task and 2.8% binary completion: about $136 per finished job, but 36 attempts for every success.

    Source ↗
  5. 5
    Kimi K2.6 API (OSWorld 2.0)

    $6.60 a task and 4.6% binary: about $143 per finished job, and 21 failures for every completion.

    Source ↗
  6. 6
    Claude Sonnet 4.6 medium on Claude Pro

    9.3% binary at $14.40 a task: about $155 per finished job, with the $240-a-year Pro plan as the entry ticket.

    Source ↗
  7. 7
    Claude Opus 5 Max Thinking on Claude Max

    13.1% binary at $23.70 a task — about $181 per finished job, roughly three times Simular Sai.

    Source ↗
  8. 8
    GPT-5.5 xhigh on ChatGPT Pro

    13% binary at $25.50 a task: about $196 per finished job, nearly double GPT-5.6's cost for 10 points less completion.

    Source ↗
  9. 9
    Claude Sonnet 4.6 max on Claude Max

    8.3% binary at $22.30 a task: about $269 per finished job — worse than the same model at medium effort.

    Source ↗
  10. 10
    Claude Opus 4.8 max, batched tool calls

    20.6% binary at $72 a task — about $350 per finished job, the most expensive measured path to a completed task.

What DeepSeek noticed

The best binary completion on OSWorld 2.0 is 44.33%, while partial scores reach 77.67%: agents get most of the way through long tasks and then fail to verify the final state.

Simular Sai finishes 28.25% of tasks at $15.70 each; Claude Opus 4.8 finishes 20.6% at $72 each — a 4.6x cost gap for 7.6 points more completion.

Google's cheapest Mac computer-use plan is $4.99/month, but its support page names only Pro and Ultra for desktop apps — the entry price and the entry entitlement disagree.

Approval is not a safety brake if users approve 93–97% of prompts; 'asks before key actions' depends on attention that measured fatigue erodes after about 50 prompts.

🇨🇳 Qwen10 places · 3 notes

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

  1. 1
    Claude Opus 4.7 (Anthropic)

    Balances a $33.60 compute cost per attempt with an 18.2% success rate, making it the cheapest per finished job when human cleanup is counted.

    Source ↗
  2. 2
    OpenAI Codex (GPT-5.5)

    The most token-efficient agent at $25.50 per attempt, but a 13% success rate means humans must intervene for 87 out of 100 jobs.

    Source ↗
  3. 3
    Gemini Computer Use Mac (Google AI Plus)

    The lowest flat entry price at $4.99 a month, but Google publishes zero completion rates, turning every task into a supervision gamble.

    Source ↗
  4. 4
    GitHub Copilot Agent Mode

    Starts at $10 a month and delegates work to pull requests, capping runaway costs but forcing human review for every finished job.

    Source ↗
  5. 5
    Cursor Cloud Agents

    Costs $20 a month and keeps working when you close your laptop, mixing its own Composer model with Claude and GPT to hedge token costs.

    Source ↗
  6. 6
    Grok Bot (X Premium)

    Gives each bot its own virtual computer for $30 a month, bypassing local app permission prompts but drawing from a shared weekly usage pool.

    Source ↗
  7. 7
    Open-source Agents (OpenClaw + MiniMax M3)

    Costs just $2.40 per attempt on OSWorld, but a 4.6% success rate means the human spends 20 times longer fixing failures than the AI saves.

    Source ↗
  8. 8
    Manus AI

    Burns credits starting at $20 a month for delegated tasks, with completion rates gated behind a proprietary cloud environment.

    Source ↗
  9. 9
    Claude Opus 4.8

    Leads OSWorld 2.0 with a 20.6% success rate, but the $72.40 compute cost per attempt makes the finished job the most expensive on the list.

    Source ↗
  10. 10
    Perplexity Comet

    Runs fully local on Mac for $5 a month, keeping data private but relying entirely on the user's hardware for long-horizon completion.

    Source ↗

What Qwen noticed

OSWorld 2.0 shows that the highest-scoring agent (Claude Opus 4.8 at 20.6%) costs $72.40 per attempt, making the cheapest finished job belong to the slightly weaker Opus 4.7.

Google publishes no success rates for Gemini Computer Use on Mac, leaving its $4.99 entry price as a pure supervision gamble compared to metered API agents.

Open-source models like MiniMax M3 cost only $2.40 per attempt on OSWorld 2.0, but their 4.6% success rate means humans spend more time fixing failures than the AI saves.