AI agents, compared honestly.

Personal agent that works on its own secure computer: email and messaging, calendar, travel and dining, shopping, bills, health and fitness.

Runs on

Muse Spark 1.3

Where it runs

Its own secure computer, run by Meta

What it can connect to

Email and messaging, calendar, travel and dining, shopping, bills, health and fitness

Price

• FreeUsage limits apply$0
• Power$20/month
• Max$100/month

Meta has not published exact paid-tier quotas.

Source: TechCrunch, September 8, 2026 · Last checked October 5, 2026

Availability

US only, ages 18+. Web, iOS, Android and WhatsApp; a Mac version followed in September.

Benchmarks

No independent benchmark published.

Safeguards the maker describes

• Meta says Muse data will not feed its advertising systems (Meta's claim).

Reported concerns

• TechCrunch noted Meta's history, including its FTC settlement, and said Meta's privacy claims for Muse need independent testing. TechCrunch, September 8, 2026

  Maker's response: Meta says Muse data will not feed its advertising systems.

Announced, not yet shipped

• Smart-glasses support
• A realtime avatar
• Fuller Mac computer control
• An email address for Muse

Announced September 23, 2026, all marked "coming soon" (TechCrunch).

Sources: TechCrunch: Meta debuts Muse (Sep 8, 2026) · TechCrunch: Everything new coming to Muse (Sep 23, 2026)

Always-on agents, each with its own cloud computer, that keep working toward goals without being prompted.

Runs on

GPT-6 Astra

Where it runs

Each dot has its own cloud computer

What it can connect to

4,000+ apps. Works through ChatGPT (desktop, web, mobile), Slack, Teams and voice calls; texting is coming.

Price

• First dotPro and Business PremiumIncluded in the plan
• Additional dotsOpenAI says more dots will be offered later at extra costNot announced

Source: OpenAI: Introducing Dots · Last checked October 5, 2026

Availability

Pro and Business Premium in eligible markets. Enterprise, Edu and Healthcare get an admin-enabled beta.

Benchmarks

No independent benchmark published.

Safeguards the maker describes

• A separate cloud computer for each dot
• Read-only permissions for background work
• Auto-review of sensitive actions
• Custom Rules for approvals
• Passwords kept away from the model
• OpenAI says "dots can still make mistakes, so always review consequential work."

Sources: OpenAI: Introducing Dots · OpenAI: DevDay 2026 recap

Persistent agent inside Microsoft 365 with its own identity, memory and workspace, built on the open-source OpenClaw framework.

Runs on

Not disclosedReported as including GPT-6 Astra and Claude models; Microsoft does not say which does what.

Where it runs

Inside Microsoft 365, with its own identity, memory and workspace

What it can connect to

Microsoft 365

Price

• Copilot subscriptionEnterprise rate; required$30/month
• Usage creditsBilled separatelyNot announced

Source: VentureBeat · Last checked October 5, 2026

Availability

Private preview, expanding. Managed runtime and SDK in public preview.

Benchmarks

No independent benchmark published.

Safeguards the maker describes

• Microsoft describes policy checks with an audit trail for each action (Microsoft's claim).

Sources: VentureBeat: Copilot revamp with Autopilot (Sep 2026) · TechCrunch: Microsoft launches Scout (Jun 2, 2026)

Agent teammate with its own cloud computer that works across apps, including ones without clean APIs, and can run several bots under a coordinator.

Runs on

Not disclosedSpaceXAI says tasks are routed to models automatically.

Where it runs

Its own cloud computer

What it can connect to

Works across apps, including ones without clean APIs; Team Bots work in Slack and Teams

Price

• x.aix.ai says usage is separate from existing plansNo price stated
• Individuals via Cursor UltraReported by VentureBeat$200/month
• Cursor Teams PremiumReported by VentureBeat$120 per seat/month

Sources disagree, so both figures are shown with who said them.

Source: x.ai: Introducing Grok Bot · Last checked October 5, 2026

Availability

Beta for SuperGrok and Cursor Pro, Pro+, Ultra and Teams plans; enterprise waitlist.

Benchmarks

No independent benchmark published.

Sources: x.ai: Introducing Grok Bot · VentureBeat: Grok Bot at $120 per month · TechCrunch: SpaceX to acquire Cursor (Jun 16, 2026)

Always-on agent across Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube and Maps, with scheduled tasks and custom skills.

Runs on

Not disclosed

Where it runs

Not disclosed

What it can connect to

Gmail, Calendar, Drive, Docs, Sheets, Slides, YouTube and Maps

Price

• Google AI Pro and UltraPlan requiredIncluded
• Spark itselfNot disclosed

Source: Google: Gemini Spark · Last checked October 5, 2026

Availability

Google AI Pro and Ultra subscribers, 18+, in select countries; still rolling out.

Benchmarks

No independent benchmark published.

Safeguards the maker describes

• App connections are off by default (Google's description)
• Google says it checks with you before major actions

Sources: Google: Gemini Spark

Agent for documents, files and multi-step work on desktop and mobile, merged with Claude chat into one product.

Claude is also the AI assistant used to research this page. This entry uses only Anthropic's announcements and press coverage.

Runs on

Claude Opus 5.5 and Sonnet 5.5See the model cards

Where it runs

Desktop and mobile apps

What it can connect to

Documents and files

Price

• Claude plansSee the Claude entries on the Models page

Source: TechCrunch, September 16, 2026 · Last checked October 5, 2026

Availability

Rolling out to Pro and Max first, Free and Team later.

Benchmarks

No independent benchmark published.

Sources: TechCrunch: Anthropic merges Claude chat and Cowork (Sep 16, 2026)

OpenAI's coding agent: a command-line tool, a cloud agent, desktop apps and IDE plugins, merged into the ChatGPT desktop app on July 9, 2026.

Runs on

GPT-6.1 Sol and other modelsSee the model cards

Where it runs

Your machine (CLI, desktop, IDE) or OpenAI's cloud, which keeps working with your laptop closed

What it can connect to

GitHub repositories (Code Review, Codex Security Cloud), your local files and editor

Price

• ChatGPT Plus, Pro, Business, Healthcare, Education, EnterpriseIncluded in the plan

Source: OpenAI: DevDay 2026 recap · Last checked October 5, 2026

Availability

ChatGPT paid plans listed above.

Benchmarks

No independent benchmark published.

Sources: OpenAI: DevDay 2026 recap · Wikipedia: OpenAI Codex (AI agent)

Anthropic's coding agent. This entry will be filled in from Anthropic's own pages.

Claude is also the AI assistant used to research this page. This entry uses only Anthropic's announcements and press coverage.

Runs on

Not disclosed

Where it runs

Not disclosed

What it can connect to

Not disclosed

Price

Not disclosed here yet.

Source: Anthropic announcements · Last checked October 5, 2026

Availability

Not disclosed here yet.

Benchmarks

No independent benchmark published.

Sources: Anthropic announcements

Agent-first coding platform that runs several agents in parallel, with a desktop app, CLI and SDKs.

Runs on

Gemini 3.5 Flash

Where it runs

Desktop app and CLI on your machine

What it can connect to

Not disclosed

Price

Not disclosed.

No pricing found.

Source: Cybernews · Last checked October 5, 2026

Availability

Not disclosed

Benchmarks

No independent benchmark published.

Sources: Cybernews: Google I/O 2026

Coding agent that aims to take a whole ticket end to end rather than answer one question.

Runs on

Not disclosed

Where it runs

Cognition's cloud

What it can connect to

Not disclosed

Price

• FreeLight quota, limited models$0/month
• Pro$20/month
• Max$200/month
• Teams$80/month + $40 per developer seat
• EnterpriseCustom, billed in compute units

Listed for the free documentation and free tier; the useful quotas are paid.

Source: Devin's pricing page · Last checked September 17, 2026

Availability

Free, Pro, Max, Teams and Enterprise plans.

Benchmarks

• SWE-bench (original, unassisted)13.86%

  share of real GitHub issues resolved end to end

  Cognition's own technical report · 15 March 2024 · maker's own figures

  The maker's own figure, and now two years old.

• SWE-bench Verified48.2% or 61.7%, depending on the source

  share of human-checked GitHub issues resolved end to end

  Dataku (48.2%) and Tensorfeed (61.7%) · 9 June 2025 / undated · independent

  The two aggregators contradict each other and neither is the official leaderboard. Treat as unverified.

Devin's numbers are a mess to cite honestly: Cognition's only first-party report predates the SWE-bench Verified set, and the two aggregators that list a current figure disagree by 13 points. Both are below.

Figures copied from the sources above, checked 29 September 2026.

Sources: Devin's pricing page

Open-source coding agent for VS Code, JetBrains, the terminal and the cloud.

Runs on

Whichever model you plug in

Where it runs

Your editor or terminal, or the cloud

What it can connect to

Your editor and files

Price

• IndividualOpen source, free forever$0
• Model useOr $0 with the free model routingProvider cost, zero markup
• Teams14-day trial$15/user/month
• EnterpriseCustom

Source: Kilo's pricing page · Last checked September 17, 2026

Availability

Individuals, teams and enterprise.

Benchmarks

• KiloBench (Terminal-Bench 2.0, run in Kilo's harness)76.2% at $87.41 per attempt — 1st

  share of real command-line tasks completed unaided, with the API cost of trying · tested: GPT-5.6 Sol driven by Kilo

  Kilo's own KiloBench board · checked September 2026 · maker's own figures

• KiloBench (Terminal-Bench 2.0, run in Kilo's harness)75.3% at $112.27 per attempt — 3rd

  share of real command-line tasks completed unaided, with the API cost of trying · tested: Gemini 3.8 Flash driven by Kilo

  Kilo's own KiloBench board · checked September 2026 · maker's own figures

Kilo now publishes its own board — KiloBench — which runs Terminal-Bench 2.0 through Kilo's actual agent harness and reports the cost of each attempt alongside the pass rate. These are Kilo's own figures, not an independent leaderboard, and each one belongs to the model plugged in rather than to Kilo itself. The 88% SWE-bench figure floating around one comparison blog still has no primary source, so it stays out.

Figures copied from the sources above, checked 29 September 2026.

1 more published figure on its own page.

Sources: Kilo's pricing page

Terminal-first, open-source coding agent.

Runs on

Your own model key, or models through the Zen gateway

Where it runs

Your machine (terminal)

What it can connect to

Your files and terminal

Price

• OpenCodeOpen source, bring your own model key$0
• Zen gatewayStarts with a $20 balancePay as you go, no subscription
• Zen — Claude Opus 4.5$5.00 in / $25.00 out per million tokens
• Zen — Claude Haiku 4.5$1.00 in / $5.00 out per million tokens

Source: OpenCode Zen pricing · Last checked September 17, 2026

Availability

Free to download.

Benchmarks

• Terminal-Bench 2.051.7% — 66th place

  share of real command-line tasks completed unaided · tested: OpenCode driving Claude Opus 4.5

  Terminal Trove comparison · checked September 2026 · independent

  Third-party comparison site, not the official leaderboard.

OpenCode publishes no score of its own; the one public figure tests it as a harness driving somebody else's model.

Figures copied from the sources above, checked 29 September 2026.

Sources: OpenCode Zen pricing

Coordinates several agents on one job instead of running a single assistant.

Runs on

Your own model key

Where it runs

Your machine

What it can connect to

Not disclosed

Price

• HerdrApache 2.0, installed and run by you$0
• Model useWhatever your own API key costs

Herdr has no pricing page; there is nothing to buy from them.

Source: the Herdr site · Last checked September 17, 2026

Availability

Free to download.

Benchmarks

No published score. Herdr isn't a model — it runs several agents side by side, so there's no capability score to publish. The only figures its site quotes are speed and processor use against tmux, which measure the terminal, not the AI. A benchmark for this kind of orchestration now exists (ClawArena-Team, July 2026), but nobody has run Herdr on it, so there is still nothing to quote.

Sources: Herdr site

Open-source, self-improving agent with memory and skills.

Runs on

The model of your choice, through your own API key

Where it runs

Your own machine or server, or a third-party host

What it can connect to

Not disclosed

Price

• SoftwareMIT license, run it yourself$0
• Model useNo fee to HermesWhatever your own API key costs
• Managed hosting — 7-day trialIncludes built-in model access, 500 conversations every 5 hours$12 one-off
• Managed hosting — monthlyThird-party hosts run this; Hostinger's managed plan starts at $5.99/month and renews at $11.99/monthFrom about $6/month
• Your own server instead$4–$25/month VPS plus your model spend

The project itself sells nothing; the hosted figures come from the third-party hosts that package it.

Source: the Hermes Agent site · Last checked September 17, 2026

Availability

Free to download.

Latest version

See releases

Benchmarks

Published scores belong to the Hermes 4 405B model, not the agent.

• MMLU-Pro82.9%

  hard multiple-choice questions across professional subjects · tested: Hermes 4 405B

  Benchmark List, citing the Hermes 4 report · 13 August 2025 report · maker's own figures

• GPQA Diamond72.7%

  graduate-level science questions · tested: Hermes 4 405B

  Nous Research technical report · 13 August 2025 · maker's own figures

The agent itself has never been benchmarked. These are scores for Hermes 4 405B, the model it runs on, and the headline numbers come from Nous Research's own technical report.

Figures copied from the sources above, checked 29 September 2026.

1 more published figure on its own page.

Sources: Hermes Agent site · GitHub releases

Open-source personal agent that became the base for other companies' agents, including Microsoft's.

Runs on

Not disclosed

Where it runs

Your own machine

What it can connect to

Not disclosed

Price

• OpenClawOpen source$0

Source: TechCrunch, February 15, 2026 · Last checked October 5, 2026

Availability

Free to download.

Benchmarks

No independent benchmark published.

Sources: TechCrunch: OpenClaw creator joins OpenAI (Feb 15, 2026)