A METR horizon is not running time. It is the length of task a human expert would need that the model finishes at 50% success. At 80% success the same models manage only about 1–3 hours, and error bars run roughly a factor of two each way.
Back to the trend of the day: Some ChatGPT Pro users can now make always-on helpers on their phones
Poly Rating · from the trend of the day
ChatGPT Dots Availability: Top 7 AI Agents of 2026
For $20 a month, Claude Pro runs cloud tasks on a model METR rates at about 12 hours of expert work. ChatGPT dots costs $100, its GPT-6 Astra model has no METR score yet, and the Pro plan that includes dots isn't sold to the roughly 530 million people of the EEA, the UK and Switzerland.
- No. 1: Manus (Cloud Computer, Automations, Cue)
- 7 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
Manus (Cloud Computer, Automations, Cue)
$20 Pro with 20 scheduled cloud tasks, sold worldwide. No METR score because it routes across models. Users report $20 of credits used up in a weekend.
Included it: Claude, ChatGPT, Grok, Gemini, StepFun, Kimi, DeepSeek. Did not include: Meta, Wenxin.In their lists: Claude: 6 · ChatGPT: 5 · Grok: 9 · Gemini: see list · StepFun: 3 · Kimi: 2 · DeepSeek: 5
What each AI noted 7
- Claudeplace 6
Free right now with an invitation code, but Manus has not said what it will cost afterwards.
Source ↗ - ChatGPTplace 5
A $20 entry-level paid plan combines a recurring-task allowance with a finite credit budget: an always-on workflow is not the same as unlimited autonomous work.
- Grokplace 9
Always-on web-task agent with free credits then $20
Source ↗ - Gemini
Named differently in its own list: open the full list.
see list - StepFunplace 3
$20 a month for a general agent that works in its own cloud session — a fifth of the price of dots, but with no METR horizon published for the harness.
Source ↗ - Kimiplace 2
Server-side background tasks on a model METR measured near 12 hours — for one-fifth of Dots' price, at $20.
- DeepSeekplace 5
A full cloud PC for $20 a month: your laptop can die and the job keeps running.
7 AIs put it here - 2
Grok Bot (xAI + Cursor)
From $20 (Cursor Pro) or $30 (X Premium), though quotes run up to $300. No published region block, so it works in Europe. Grok 4 measured only 1h49m on METR.
Included it: Claude, ChatGPT, Meta, Grok, Kimi, DeepSeek. Did not include: Gemini, StepFun, Wenxin.In their lists: Claude: 5 · ChatGPT: 4 · Meta: 4 · Grok: 5 · Kimi: 3 · DeepSeek: 4
What each AI noted 6
- Claudeplace 5
Opened to a limited free trial on 21 August 2026, but the best-measured Grok model (Grok 4) sits at only 1 h 49 min on METR.
Source ↗ - ChatGPTplace 4
An always-on bot priced below Dots in independent comparisons, but its shared cloud-computer setup and undisclosed underlying model make performance-per-dollar difficult to establish.
Source ↗ - Metaplace 4
Price dropped 10× on Aug 26 2026 — $20 Pro now includes bot that launched at $200-300
Source ↗ - Grokplace 5
Included from $30 X Subscription Premium or legacy Cursor $20; global but usage-pooled
Source ↗ - Kimiplace 3
The cheapest way into an always-on cloud-computer agent — $20 via Cursor Pro — and the one always-on agent with no published region limit, so it works in Europe today.
Source ↗ - DeepSeekplace 4
The only agent that lets one subscriber run about 50 bots on one cloud computer — and xAI won't say which model powers them.
Source ↗
6 AIs put it here - 3
Meta Muse
The only one you can start for free ($20 Power tier), running in its own cloud VM. But it's sold only in the US and Canada, about 4.7% of the world (estimate). No METR score.
Included it: ChatGPT, Meta, Grok, Kimi, DeepSeek. Did not include: Claude, Gemini, StepFun, Wenxin.In their lists: ChatGPT: 3 · Meta: 3 · Grok: 4 · Kimi: 4 · DeepSeek: 6
What each AI noted 5
- ChatGPTplace 3
The most striking price contrast: a cloud personal agent with a free entry point, versus Dots at $100 a month.
Source ↗ - Metaplace 3
$0 entry for always-on agent — cheapest per hour is infinite, but only 4.6% of world (US + Canada) can turn it on
Source ↗ - Grokplace 4
Free tier plus $20 Power; only US+Canada (~4–5 % of world)
Source ↗ - Kimiplace 4
The only always-on agent you can start for free — but it is locked inside two countries holding under 5% of humanity.
Source ↗ - DeepSeekplace 6
Free is the headline; the US-and-Canada-only launch and a 'forget' button that doesn't delete are the catch.
Source ↗
5 AIs put it here - 4
Microsoft 365 Copilot Cowork / Autopilot
Cowork became generally available worldwide, Europe included, at about $30 a seat, but only for businesses. The always-on Autopilot agent is still in private preview.
Included it: Claude, Meta, Grok, StepFun, DeepSeek. Did not include: ChatGPT, Gemini, Kimi, Wenxin.In their lists: Claude: 10 · Meta: 6 · Grok: 6 · StepFun: 6 · DeepSeek: 10
What each AI noted 5
- Claudeplace 10
The cheapest always-on seat on paper, $5 a month plus $0.40 an hour, but only for US organisations in preview.
Source ↗ - Metaplace 6
$30 seat + meter — $0.01 per credit, 3,000 credits = another $30 seat
- Grokplace 6
Enterprise-grade always-on inside M365; $30/user + credits
Source ↗ - StepFunplace 6
The only always-on agent here that went generally available worldwide on day one — $30 a seat, including the European countries where dots is not on personal sale.
Source ↗ - DeepSeekplace 10
A $30 seat plus metered Copilot Credits, and you still can't buy it: private preview only.
5 AIs put it here - 5
Cursor Cloud Agents
$20 Pro runs coding agents in cloud sandboxes, sold worldwide. No per-task fee, but they draw on the editor's quota. Models vary, so there's no single METR figure.
Included it: ChatGPT, Gemini, Kimi, DeepSeek. Did not include: Claude, Meta, Grok, StepFun, Wenxin.In their lists: ChatGPT: 9 · Gemini: 4 · Kimi: 8 · DeepSeek: 8
What each AI noted 4
- ChatGPTplace 9
Useful for developers who want background coding work, but agent consumption and task limits matter more than a headline subscription price.
Source ↗ - Geminiplace 4
Provides 1.6 hours of offline codebase refactoring for $40/month, delivering top economic efficiency for working developers.
Source ↗ - Kimiplace 8
Background coding agents on Cursor's cloud from $20 — the plan that also doubles as the cheapest door into Grok Bot.
Source ↗ - DeepSeekplace 8
Coding agents bundled into a $20 IDE plan: no per-task fee, but they eat the same request quota as your editor.
Source ↗
4 AIs put it here - 6
Devin (Cognition)
$20 to start, but work is billed at $2.25 per ACU (about 15 minutes), and older coverage still quotes a $500 minimum. Coding only, with no product METR score.
Included it: Gemini, StepFun, DeepSeek. Did not include: Claude, ChatGPT, Meta, Grok, Kimi, Wenxin.In their lists: Gemini: 3 · StepFun: 4 · DeepSeek: 7
What each AI noted 3
- Geminiplace 3
Delivers the highest autonomy benchmark at 3.5 hours of human-expert software engineering, but demands $500 monthly.
Source ↗ - StepFunplace 4
$20 a month to rent an async software engineer that keeps its own machine running after you close the tab — the only agent here priced per unit of work (ACUs).
Source ↗ - DeepSeekplace 7
The cloud engineer that bills by the 15-minute ACU — $20 gets you in, but a day of work costs more than the subscription.
Source ↗
3 AIs put it here - 7
Perplexity Computer
The priciest general agent here: $200 Max with 10,000 credits (a cheaper Pro route is reported but unconfirmed). One reviewer used up $100 of credits in an hour.
Included it: Claude, StepFun. Did not include: ChatGPT, Meta, Grok, Gemini, Kimi, DeepSeek, Wenxin.In their lists: Claude: 4 · StepFun: 8
What each AI noted 2
- Claudeplace 4
Model-agnostic and priced like a car lease at about $200 a month on Max, so no single METR number applies.
- StepFunplace 8
$200 a month and 10,000 credits for a cloud agent that researches, codes and deploys — the priciest general assistant here, with users reporting $100 gone in an hour.
2 AIs put it here
02Worth knowing
What the AIs noticed while building the list.
METR revised Claude Opus 4.6 from ~14.5h to ~12.0h in March 2026 after a regularization fix, so pages still quoting 14.5h are out of date.
Grok Bot's entry price fell about tenfold, from $200–300 at the 11 August beta to $20 via Cursor Pro on 26 August.
The cheapest agent and the widest-reaching one are different products: Meta Muse is free but sold only in the US and Canada, while ChatGPT dots reaches most of the world from $100.
03How we counted
In plain words, no formulas.
1The question
Top 10 Always-On AI Agents: Hours of Work Against Monthly Price and Who Can Buy One
2Why this topic
It is the only list that puts an agent's lab ability, its price and the number of people allowed to buy it side by side, on the week dots launched.
3Where it applies
Global consumer and prosumer AI assistants that keep working in the cloud while your device is off
4What we looked at
- Length of task the underlying model finishes alone at 50% success (METR time horizon)
- Monthly price of the cheapest plan that includes the always-on agent
- Share of the world's population living where that plan and feature are sold
- Whether the agent keeps acting when the phone or computer is off
- METR — Task-Completion Time Horizons — Length of a human-expert task, in hours, that a model completes alone at 50% and 80% success; METR says values above 16 hours are unreliable
- METR — note on modelling assumptions (March 2026) — A fix that lowered recent models' 50% horizons by up to 20% (Claude Opus 4.6 from ~14.5h to ~12h)
- OpenAI Help Center — Getting started with your dot — Plans that include dots (Pro from $100, Business Premium, Enterprise) and the regions left out (EEA, UK, Switzerland)
- Anthropic — Claude pricing and Cowork cloud docs — Cheapest Claude plan with cloud Cowork sessions and scheduled tasks that run with the device off
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, StepFun, Kimi, DeepSeek, Wenxin.
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.
04Each AI's own list
🇺🇸 Claude10 places · 4 notes
The same question, answered by Claude alone. The final list above combines all the AIs.
- 1Claude Code routines (Anthropic, Claude Pro/Max)
The cheapest ticket to the longest-horizon model: Pro is about $20 a month, against a METR 50% horizon of about 12 hours for Opus 4.6 and 16+ hours for Mythos Preview.
Source ↗ - 2Gemini Spark (Google AI Pro / Ultra)
The widest reach of any always-on agent: about 160 countries and $19.99 a month on the Pro tier, with the EEA, UK, Switzerland and Nigeria left out.
Source ↗ - 3ChatGPT dots (OpenAI, Pro from $100, announced Sept/Oct 2026)
The agent the question is named for: each dot gets its own cloud computer, but the cheapest ticket is $100 a month and Pro buyers in the EEA, UK and Switzerland are shut out.
Source ↗ - 4Perplexity Computer (Max, reportedly also Pro)
Model-agnostic and priced like a car lease at about $200 a month on Max, so no single METR number applies.
- 5Grok Bot (SpaceXAI + Cursor, beta)
Opened to a limited free trial on 21 August 2026, but the best-measured Grok model (Grok 4) sits at only 1 h 49 min on METR.
Source ↗ - 6Manus Cue (early access)
Free right now with an invitation code, but Manus has not said what it will cost afterwards.
Source ↗ - 7Sai by Simular
A small player with a $50 entry plan, but an always-on computer only on the $500 Unlimited plan.
- 8Claude Managed Agents (Anthropic API)
No monthly fee, but a 24/7 session runs about $58 a month in runtime before tokens.
Source ↗ - 9Microsoft Scout (Frontier preview)
The only big-lab always-on agent that a person cannot simply buy: it needs an admin, Intune and about $99 per user.
- 10Windows 365 for Agents (Microsoft)
The cheapest always-on seat on paper, $5 a month plus $0.40 an hour, but only for US organisations in preview.
Source ↗
What Claude noticed
The model with the best lab score is not the agent most people can switch on. Claude Mythos Preview is at 16+ hours on METR, yet the cheapest always-on route to Anthropic models is a roughly $20 Pro plan with a 5-runs-a-day cap.
Google undercuts OpenAI on entry price by about five times. Spark is reported at $19.99 on AI Pro, against $100 for ChatGPT Pro to get dots.
Europe is the common blind spot. Dots on Pro and Spark both exclude the EEA, UK and Switzerland, about 6.5% of the world's people by my UN-based estimate, while Business Premium dots do reach those regions.
METR's frontier doubles roughly every 89 to 131 days depending on the fit, but the suite is now saturated. Above 16 hours the numbers are unreliable.
🇺🇸 ChatGPT10 places · 4 notes
The same question, answered by ChatGPT alone. The final list above combines all the AIs.
- 1ChatGPT Dots (OpenAI, 2026)
The clearest match for the brief: a cloud agent that can keep working after your phone or computer is switched off, but costs at least $100 monthly for an eligible individual.
Source ↗ - 2Claude Cowork Cloud (Anthropic, 2026)
A $20 monthly subscription can now support cloud-running work without leaving your laptop open, making it substantially cheaper than individual Dots access.
Source ↗ - 3Meta Muse (Meta, 2026)
The most striking price contrast: a cloud personal agent with a free entry point, versus Dots at $100 a month.
Source ↗ - 4Grok Bot (xAI, 2026)
An always-on bot priced below Dots in independent comparisons, but its shared cloud-computer setup and undisclosed underlying model make performance-per-dollar difficult to establish.
Source ↗ - 5Manus AI Agent (2026)
A $20 entry-level paid plan combines a recurring-task allowance with a finite credit budget: an always-on workflow is not the same as unlimited autonomous work.
- 6Gemini Spark (Google, 2026)
Google's 24/7 personal-agent ambition is substantial, but its Ultra subscription and supported-country restrictions make reach and affordability central to the comparison.
Source ↗ - 7Hermes Agent (Nous Research, 2026)
The software can be free, but the server, model inference and maintenance are not necessarily free; total cost depends on how it is hosted.
Source ↗ - 8OpenClaw (2026)
Zero license cost is not zero operating cost: the appeal is control over deployment, with cloud uptime and reliability left to the operator.
Source ↗ - 9Cursor Cloud Agents (2026)
Useful for developers who want background coding work, but agent consumption and task limits matter more than a headline subscription price.
Source ↗ - 10Google AI Ultra with Gemini Agent Workflows (2026)
A powerful high-tier route into Google's agent ecosystem, but the distinction between a cloud agent and a desktop agent remains decisive.
Source ↗
What ChatGPT noticed
METR's time horizon is often misunderstood: a two-hour horizon does not mean an agent runs autonomously for two hours. It estimates the human-expert duration of tasks at a specified predicted success probability.
The price gap is unusually large: independent comparisons put Meta Muse at free entry-level use, Claude Cowork at $20 monthly, and ChatGPT Dots at $100 monthly for individual Pro access.
A device-off claim should be tested at the workflow level. A cloud agent can keep running after a laptop closes, while an agent that depends on local files or an active desktop session may not.
The first independent Dots-versus-Muse hands-on comparison found impressive multistep preparation but continued human-editing needs and problems executing real-world purchases. Autonomy is not equivalent to dependable end-to-end completion.
🇺🇸 Meta10 places · 4 notes
The same question, answered by Meta alone. The final list above combines all the AIs.
- 1Anthropic Claude Cowork (Claude Opus 4.6)
718 minutes at 50% — about 12 hours of expert work for $20, the first full workday horizon
Source ↗ - 2OpenAI ChatGPT Dots Pro 100 (GPT-6 Astra)
First dot included at $100/mo — projected 45 hours horizon, reaches ∼93.6% of world after EEA/UK/CH exclusion
Source ↗ - 3Meta Muse Free Tier
$0 entry for always-on agent — cheapest per hour is infinite, but only 4.6% of world (US + Canada) can turn it on
Source ↗ - 4xAI Grok Bot via Cursor Pro
Price dropped 10× on Aug 26 2026 — $20 Pro now includes bot that launched at $200-300
Source ↗ - 5Google Gemini Spark (Google AI Pro)
$20 for Spark — runs while you sleep, but still US-only beta in July 2026
Source ↗ - 6Microsoft Copilot Autopilot
$30 seat + meter — $0.01 per credit, 3,000 credits = another $30 seat
- 7OpenAI Dots Business Premium
Same Astra brain but $100 per user billed annually ($125 monthly) — the only way to get dots in EEA/UK today
Source ↗ - 8Meta Muse Power ($20)
20× the token budget of free — still US/Canada only, so 0.6 hours per dollar if 1h horizon
Source ↗ - 9Google Gemini Enterprise Agent
Private preview Oct 8 2026 — no published price, no GA date, gets own Workspace email
Source ↗ - 10OpenAI Dots Pro 500 ($500 Ultrafast)
8× faster Codex at $500/mo — same Astra, but credit pool 62,500 expiring Dec 31 for early Pro
Source ↗
What Meta noticed
Free beats $100 on price, but not on reach — Muse free reaches 4.7% of world vs Dots Pro 100 reaching 93.6% because EEA+UK+CH are only ∼6.4% of 8.16B
12-hour horizon is 180× the 4-minute GPT-4 horizon in March 2023, doubling roughly every 129 days since 2023 per METR correction March 3 2026
Grok Bot entry price fell 10× from $200-300 at Aug 11 beta to $20 via Cursor Pro on Aug 26, the biggest price move among always-on agents
Measurements above 16 hours flagged unreliable by METR — Claude Mythos preview 16h+ estimate had 95% CI 8.5h to 55h
🇺🇸 Grok10 places · 4 notes
The same question, answered by Grok alone. The final list above combines all the AIs.
- 1ChatGPT Dots (OpenAI, GPT-6 Astra)
First true always-on cloud agent with own computer; included in $100 Pro outside EEA/UK/CH or Business Premium worldwide
- 2Gemini Spark / Task mode (Google)
Cheapest broad always-on agent at $19.99; excluded only from EEA/UK/CH/Nigeria
Source ↗ - 3Claude Cowork (Anthropic)
Highest measured METR horizons (12–17 h) and cloud continuation from $20 Pro
Source ↗ - 4Meta Muse
Free tier plus $20 Power; only US+Canada (~4–5 % of world)
Source ↗ - 5Grok Bot (xAI)
Included from $30 X Subscription Premium or legacy Cursor $20; global but usage-pooled
Source ↗ - 6Microsoft 365 Copilot Cowork / Copilot Studio agents
Enterprise-grade always-on inside M365; $30/user + credits
Source ↗ - 7Zapier Agents
No-code always-on across 9 000+ apps; free tier then $20+
Source ↗ - 8n8n (self-hosted or cloud)
Open-source always-on workflows from €20; full control
Source ↗ - 9Manus AI
Always-on web-task agent with free credits then $20
Source ↗ - 10Hark Pro / Instinct (early always-on text agents)
Text-to-agent from free/invite or $20; still limited geography
Source ↗
What Grok noticed
ChatGPT Dots and Gemini Spark both exclude the EEA/UK/Switzerland at launch for consumer tiers, leaving ~450–500 million people in high-AI-use Europe temporarily locked out of the newest always-on features while Business/Ultra plans fill the gap.
METR’s 50 % time horizon has roughly doubled every 3–4 months since 2024; the current ceiling (~17 h for early Claude Mythos) already exceeds a typical workday, yet measurements above 16 h are flagged unreliable on the present task suite.
Meta Muse is free to start but available only in the US and Canada—roughly 4.7 % of world population—while Claude Cowork at $20 reaches essentially every market where Claude itself is sold.
The cheapest always-on entry points ($0–$20) are now Meta Muse free, Gemini Spark $19.99, Claude Cowork $20 and Zapier free tier; the $100 ChatGPT Pro tier remains the highest-capability consumer option with its own cloud computer.
🇺🇸 Gemini10 places · 1 notes
The same question, answered by Gemini alone. The final list above combines all the AIs.
- 1OpenAI ChatGPT Dots (2026)
Carries a 2.8-hour METR capability horizon on dedicated cloud sandboxes, but locks out 520 million European consumer subscribers at rollout.
Source ↗ - 2Anthropic Claude Computer Use Agent (2026)
Balances 2.2 hours of autonomous desktop task navigation at $20/month with broad compliance across 159 countries.
Source ↗ - 3Cognition Devin (2026)
Delivers the highest autonomy benchmark at 3.5 hours of human-expert software engineering, but demands $500 monthly.
Source ↗ - 4Cursor Background Agent (2026)
Provides 1.6 hours of offline codebase refactoring for $40/month, delivering top economic efficiency for working developers.
Source ↗ - 5GitHub Copilot Workspace (2026)
Taps Azure cloud containers across 190 countries for $39/month while sustaining 1.4-hour autonomous task completions.
Source ↗ - 6Google Gemini Astra Workspace Agent (2026)
Bundles Google Workspace and cloud agent automation into Google One for $20/month with 80% global consumer reach.
- 7Factory Code Droid (2026)
Focuses on 2.5-hour automated migration loops, but charges $250/month and sells almost exclusively to Western enterprise teams.
- 8Manus AI Autonomous Agent (2026)
Delivers general-purpose browser execution for $40/month, completing 1.1-hour human research tasks unattended.
- 9Adept Work Navigator Cloud (2026)
Operates enterprise CRM and back-office pipelines for $30/month, but caps out at 0.9 hours of unattended reliability.
- 10MultiOn Cloud Agent (2026)
Offers the lowest barrier to entry at $29/month, though it stumbles on complex multi-step human workflows beyond 0.6 hours.
What Gemini noticed
OpenAI's mobile rollout of ChatGPT 'dots' introduces dedicated cloud computers per user, but regulatory hurdles in the EU and UK mean personal prosumers in the European Economic Area cannot purchase individual dots despite holding paid Pro accounts.
🇨🇳 StepFun10 places · 3 notes
The same question, answered by StepFun alone. The final list above combines all the AIs.
- 1Claude Cowork (Anthropic, 2026)
$20 a month buys the longest measured work streak in this table — 14.5 hours at 50% success, about 0.7 hours of agent work per dollar.
- 2ChatGPT dots (OpenAI, October 2026)
The always-on helper that named the category: $100 a month, and personal Pro buyers in the EU, UK and Switzerland — about 537 million people — are left out at launch.
Source ↗ - 3Manus (Butterfly Effect, 2026)
$20 a month for a general agent that works in its own cloud session — a fifth of the price of dots, but with no METR horizon published for the harness.
Source ↗ - 4Devin (Cognition, 2026)
$20 a month to rent an async software engineer that keeps its own machine running after you close the tab — the only agent here priced per unit of work (ACUs).
Source ↗ - 5Genspark Super Agent (2026)
$24.99 a month for an agent that drives other labs' frontier models — cheap access, but you are renting someone else's METR score.
Source ↗ - 6Microsoft 365 Copilot Cowork (2026)
The only always-on agent here that went generally available worldwide on day one — $30 a seat, including the European countries where dots is not on personal sale.
Source ↗ - 7Zapier Agents (2026)
Always-on in everything but name for years: $33 a month and it never sleeps — but its output is workflow runs, not a lab time horizon.
- 8Perplexity Computer (2026)
$200 a month and 10,000 credits for a cloud agent that researches, codes and deploys — the priciest general assistant here, with users reporting $100 gone in an hour.
- 9Gemini Agent (Google AI Ultra, 2026)
The newest model in the ranking sits in the smallest market: US-only, English-only, $99.99 a month, while AI Ultra itself sells in 140+ countries.
Source ↗ - 10Lindy (2026)
Sold in 100+ countries from $29.99 a seat — but ask three pricing pages and you get three different numbers for the same plan.
What StepFun noticed
The price gap is the story: Claude's $20 Pro plan carries a model METR clocks at 14.5 hours of work at 50% success, while dots costs $100 for a GPT-6 whose only published agentic horizon is UK AISI's 30.9 minutes — five times the price with no METR number yet.
Google's AI Ultra plan is sold in 140+ countries, but the Gemini Agent inside it is US-only and English-only: the newest always-on agent has the smallest addressable market in the table.
METR itself warns the ranking's headline number is soft: Opus 4.6's 14.5-hour 50% horizon carries a 95% confidence interval of 6 to 98 hours, and the benchmark only covers software tasks.
🇨🇳 Kimi10 places · 4 notes
The same question, answered by Kimi alone. The final list above combines all the AIs.
- 1ChatGPT Dots (OpenAI, 2026)
The only always-on agent running on a frontier model measured at a ~16–20h solo task horizon — but a $100 ticket that 532 million Europeans can't buy on Pro.
Source ↗ - 2Claude Cowork (Anthropic, 2026)
Server-side background tasks on a model METR measured near 12 hours — for one-fifth of Dots' price, at $20.
- 3Grok Bot (xAI, 2026)
The cheapest way into an always-on cloud-computer agent — $20 via Cursor Pro — and the one always-on agent with no published region limit, so it works in Europe today.
Source ↗ - 4Meta Muse (Meta, 2026)
The only always-on agent you can start for free — but it is locked inside two countries holding under 5% of humanity.
Source ↗ - 5Manus Cue (Manus, 2026)
Named personal agents that take phone calls for you — free in early access, if you can find an invite code.
Source ↗ - 6Manus 2.0 Automations (Manus, 2026)
Twenty concurrent scheduled cloud tasks for $20 — the most scheduled automations per dollar of any paid agent plan.
- 7OpenAI Codex cloud agent (2026)
The cloud coding teammate inside ChatGPT Plus — always-on cloud runs for $20, if you accept tighter task limits than Pro.
Source ↗ - 8Cursor cloud agents (Cursor, 2026)
Background coding agents on Cursor's cloud from $20 — the plan that also doubles as the cheapest door into Grok Bot.
Source ↗ - 9Zentor cloud agent (2026)
The fixed-price antidote to Manus's credit roulette: $20 with 1,000 included credits and scheduled browser tasks.
- 10Vellum personal assistant (2026)
A permanent free tier with free cloud hosting — the only $0 way to keep an always-on agent running, if you accept Claude-only models in the managed cloud.
What Kimi noticed
The price ladder for 'an agent that works while you sleep' now spans $0 to $500: Muse is free (US/Canada), Grok Bot and Cowork start at $20–30 with no meaningful region wall, while Dots starts at $100 and is the only one gated by regulation rather than money.
METR's own numbers say the underlying models can already work ~12–20 hours alone at 50% reliability — yet no subscription sells you that much autonomous compute; every plan meters the agent with weekly or usage caps that no vendor publishes in task-hours.
Europe inverts the usual rollout: EU/UK/Swiss consumers are the only Pro subscribers on Earth who cannot get Dots, so the practical always-on choice there is Grok Bot — a SpaceXSI product — according to hands-on comparisons from early October 2026.
ChatGPT's supported-country list (~208 territories) still excludes about 1.7 billion people — China, Russia, Iran, North Korea, Cuba, Belarus, Hong Kong, Macau, Syria — so even the widest-reach agent on this list is unreachable to roughly one in five humans.
🇨🇳 DeepSeek10 places · 3 notes
The same question, answered by DeepSeek alone. The final list above combines all the AIs.
- 1ChatGPT Dots (OpenAI, 2026)
The only always-on agent sold across 208 countries — and the $100 Pro plan is the price of entry, not a premium add-on.
Source ↗ - 2Claude Duet (Anthropic, 2026)
The best model per dollar in the category — if you live in the one country where it is sold.
Source ↗ - 3Gemini Spark (Google, 2026)
Google cut the price of a 24/7 agent from $100 to $20 and opened 160+ countries; the EEA, UK and Nigeria are still shut out.
Source ↗ - 4Grok Bot (xAI, 2026)
The only agent that lets one subscriber run about 50 bots on one cloud computer — and xAI won't say which model powers them.
Source ↗ - 5Manus Cloud Computer (2026)
A full cloud PC for $20 a month: your laptop can die and the job keeps running.
- 6Meta Muse (2026)
Free is the headline; the US-and-Canada-only launch and a 'forget' button that doesn't delete are the catch.
Source ↗ - 7Devin Cloud (Cognition, 2026)
The cloud engineer that bills by the 15-minute ACU — $20 gets you in, but a day of work costs more than the subscription.
Source ↗ - 8Cursor Cloud Agents (2026)
Coding agents bundled into a $20 IDE plan: no per-task fee, but they eat the same request quota as your editor.
Source ↗ - 9Hermes Cloud (Nous Research, 2026)
The reach champion: open-source software plus $15 hosting works in places no US lab sells to — if you can set it up.
Source ↗ - 10Microsoft Copilot Autopilot (2026)
A $30 seat plus metered Copilot Credits, and you still can't buy it: private preview only.
What DeepSeek noticed
Four always-on agents launched in 71 days — Grok Bot on 11 August, Muse on 8 September, Dots on 29 September and Claude Duet on 21 October — turning a category from zero to five major products in under three months.
The cheapest always-on agent is free: Meta Muse has a free tier, while ChatGPT Dots starts at $100/month — but the free one is US/Canada only, so the price floor and the reach ceiling do not coincide.
METR's own authors warn the 50% horizon is not unsupervised work time: at 80% reliability the same models manage about 1.5 hours, and error bars run roughly a factor of two each way.
🇨🇳 Wenxin10 places · 1 notes
The same question, answered by Wenxin alone. The final list above combines all the AIs.
- 1OpenAI Dots
Powered by GPT-6 Astra, it achieves 52.7 intelligence score for multi-step work, 149 t/s processing speed and supports 4000+ app integrations on dedicated cloud instances
Source ↗ - 2Claude Fable 5.1 Always-On Agent
Top-scoring agent on the 2026 Artificial Analysis Agentic Index at 57.9, beats all competitors on tool use and multi-step task completion performance
- 3Claude Opus 5 Always-On Agent
Delivers 94 t/s processing speed, half the input cost of Claude Fable 5.1 while retaining near-top-tier reasoning performance for agent tasks
- 4Qwen3.8 Max (0902) Always-On Agent
Delivers top-3 agent performance at just 2/1M input tokens, offering exceptional cost efficiency for continuous cloud agent workflows
- 5Muse Spark 1.3 Always-On Agent
Matches Claude Opus 5's 93 t/s processing speed at 75% lower input cost, ideal for high-volume fast agent loop workloads
- 6GLM 5.3 Always-On Agent
Delivers 143 t/s top-tier speed at just 0.079/1M input tokens, the lowest cost of any model in the top 7 agent ranking
- 7Grok 4.6 Always-On Agent
Optimized explicitly for fast agent loops, with 85 t/s processing speed to complete continuous tasks far quicker than general purpose models
- 8Kimi K2.6 Open-Weight Always-On Agent
Low cost open-weight model with strong tool use capabilities, priced at just0.68/0.68/3.41 per 1M input/output tokens for self-hosted or managed deployments
- 9GPT-5.4 Cloud Agent
Specialized for terminal and code execution agent tasks, with 2.50/15 per 1M input/output token pricing far lower than newer GPT-6 Astra
- 10Gemini 3.1 Pro Cloud Agent
Leading multimodal and long-context agent performance, with2/2/12 per 1M input/output token pricing optimized for long-running multimodal tasks
What Wenxin noticed
OpenAI Dots is the only always-on agent with formal support for over 4000 third-party app integrations via plugins, far exceeding the 500+ typical for competing products