TL;DR: Claude Duet, launched as a US-only research preview on October 21, 2026, is a proactive agent on Claude Opus 5.5 that monitors your connected apps, sends a morning brief, and runs tasks on its own cloud VM. It is included in Pro ($20/mo) and Max (from $100/mo) — no new tier — but every background check burns your existing usage allowance. It is not a coding tool, and there is no Claude Code or GitHub integration yet.

	Claude Duet	OpenAI Dots	Claude Code + cron
Best for	Cross-app monitoring and delegated office tasks on a plan you already pay for	The same job, if you’re already on ChatGPT Pro	Recurring repo-scoped jobs you control end to end
Price / cost	$0 extra on Pro ($20/mo) or Max (from $100/mo), but shares plan usage limits	First dot needs ChatGPT Pro ($100/mo) or Business Premium	The Claude plan or API key you already pay for; cron is $0
The catch	US waitlist only; no API, no GitHub/MCP/Claude Code hooks documented; background checks eat your limits	No API; dot-started Codex tasks bill against a shrinking allowance	You write and babysit the automation yourself; nothing watches your inbox

Honest take: If you’re a US developer already on Claude Pro or Max, join the waitlist — it costs nothing on top and the morning-brief-plus-monitoring loop is genuinely new at the $20 price point. Do not expect it to touch your repos: for code automation, a cron line driving claude -p still does more, today, in every country, on the same subscription.

Anthropic shipped Claude Duet on October 21, 2026 as a research preview, and the Hacker News thread that followed filed it under AI coding news — which is how it landed in our topic scan with a garbled premise. Before this article existed, the claim circulating was “two Claude instances working together, one proposing and one reviewing.” That is not what Duet is. We verified against Anthropic’s own announcement and product page on October 28, 2026, and the real product is more interesting for some developers and less interesting for others than the rumor suggested. Here’s what it actually is, what it costs in practice, and the honest boundary between Duet and the coding tools this site covers.

What is Claude Duet and what does it actually do?

Claude Duet is a standalone app — iOS, Android, and claude.ai on the web — where Claude acts as a proactive agent for one person rather than a chat assistant waiting for prompts. Anthropic announced it October 21, 2026 as a research preview built around Claude Opus 5.5, with a safety and security update following on October 23.

Four mechanics define it, per the official announcement:

• Morning briefs. Duet sends a short daily summary of the most important items across your connected apps.
• Monitoring. It checks connected accounts every few hours and messages you only when something needs attention. You configure what you hear about and how often.
• Its own virtual machine. Tasks run on Anthropic’s cloud VM with its own browser, so your laptop or phone doesn’t need to be on — the same cloud-computer architecture OpenAI shipped with Dots in September.
• One thread, with approvals. You describe tasks in a single ongoing conversation (typed or voice). Before Duet sends a message, makes a payment, signs a document, or deletes anything, it shows you the action and waits. Trading securities and moving money between accounts are off-limits even with your approval.

It carries over memory, skills, and connected apps from your existing Claude account, and reaches over 6,000 connectors and plugins from the Claude Marketplace — Anthropic names Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn, DocuSign, HubSpot, Canva, and Shopify as launch examples.

Is Claude Duet a coding tool?

No — and the “two Claudes pair-programming” framing that spread through aggregators is wrong. Nothing in Anthropic’s announcement or product page describes two model instances collaborating, a proposer/reviewer loop, or any Claude Code integration. Claude Duet is Anthropic’s entry in the proactive-personal-agent category, squarely against OpenAI Dots (covered in our October 3 review) and Meta’s Muse assistant push.

The developer-relevant facts, as of October 28, 2026:

• No API. Duet is a consumer surface. You cannot trigger it programmatically or wire it into CI.
• No documented GitHub, MCP, or Claude Code integration. The 6,000-connector marketplace is office-and-commerce shaped (email, CRM, payroll, storefronts), and the announcement does not mention repository access.
• It can ship code — loosely. Anthropic’s page cites building dashboards, prototyping web apps, and says the Duet team itself uses Duet to ship code. But that work happens inside Duet’s VM and browser; there is no documented path from Duet into your repos, your terminal, or your review flow.
• Same safety framework as Claude Code. Duet is built on Anthropic’s trustworthy-agents framework — the one behind Claude Code and Claude in Chrome — with the Managed Agents credential design, meaning Duet never directly sees your passwords or sign-in tokens.

So the right mental model for a developer: Duet is the thing that watches your inbox, Slack, and invoices while Claude Code watches your repo. Those two jobs don’t overlap yet, and Anthropic has not announced plans to connect them.

What does Claude Duet cost on Pro and Max?

Nothing extra — and that’s the headline against OpenAI Dots. Duet is included in Claude Pro ($20/month, or $17/month billed annually at $200/year) and Max (from $100/month), verified against claude.com/pricing on October 28, 2026. There is no separate Duet tier and no per-task fee. Access is a waitlist for US subscribers at claude.com/product/claude-duet; Team and Enterprise plans are not included yet.

The catch is in the metering. Duet shares your plan’s usage limits with the rest of Claude, and it consumes them differently than chat:

1. Background checks count. Every few-hours scan of your connected accounts draws from the same allowance as your Claude conversations.
2. Big tasks draw big. Anthropic says larger jobs — building a financial model, producing a deck — require more model turns and more background work, all metered against your plan.
3. It runs Opus 5.5. Duet is built on Anthropic’s second-most-expensive model (see our Opus 5.5 backend review), not a cheap routing tier, so the turns it burns are premium turns.

The problem this creates is concrete: a Pro subscriber who also uses Claude heavily for daily work — or runs Claude Code on the same subscription — now has an always-on agent silently drawing down the same pool every few hours. Anthropic publishes no hard token numbers for plan limits, so you can’t budget this in advance. The practical fix until real-world usage reports accumulate: configure Duet’s monitoring scope and frequency down (both are user-configurable), connect only the two or three apps whose alerts you’d act on, and treat Max — with its 5x or 20x Pro allowance — as the tier where “always-on” is actually sustainable. If you find your Pro limits exhausted by mid-afternoon after enabling Duet, the monitoring cadence is the first dial to turn, not the subscription.

How does Claude Duet compare to OpenAI Dots for a developer?

Duet undercuts Dots on entry price by 5x and loses on nothing a developer has confirmed access to yet. Both are cloud-VM proactive agents with connector ecosystems and no API; the structural differences, comparing Anthropic’s October 21 launch against what we verified for our Dots review on October 3:

	Claude Duet	OpenAI Dots
Cheapest plan that includes it	Pro, $20/mo	ChatGPT Pro, $100/mo
Model	Claude Opus 5.5	GPT-6 Astra
Connector ecosystem	6,000+ (Claude Marketplace)	4,000+ apps
Coding delegation path	None documented	Dot can start Codex cloud tasks
Usage metering	Shares plan limits, including background checks	Dot chats free; dot-started Codex/Work tasks bill against allowances (first month exempt)
Hard-blocked actions	Trading securities, moving money between accounts	Password changes, money transfers always hand back
Availability	US waitlist, Pro/Max only, iOS/Android/web	Eligible markets, Pro/Business Premium+

Two of those rows deserve emphasis. First, Dots has a confirmed coding path — a dot can delegate work to Codex and hand back pull requests — while Duet has none; if your “agent works while I sleep” wishlist is code-shaped, Dots is currently ahead on capability and behind by $80/month on price. Second, the metering models differ in direction: OpenAI exempted dot usage for the first month then bills Codex delegation against allowances, while Anthropic meters everything from day one but adds no tier above what you already pay. Neither company lets you compute the real monthly cost in advance, which remains this category’s shared defect.

For who should pick which: a developer already paying for ChatGPT Pro should work their included dot and skip Duet’s waitlist until it gains repo access. A developer on Claude Pro or Max — a much lower bar — gets Duet for free and gives up only the Codex delegation they weren’t using anyway. Nobody should buy a $100 plan from either company just to get a proactive agent in October 2026.

Can a scheduled Claude Code job replace Claude Duet?

For repo work — entirely, and Duet doesn’t compete. For cross-app monitoring — no, and that’s the only reason Duet earns a slot. The boundary is worth drawing precisely, because the same Claude subscription powers both.

Claude Code’s non-interactive mode runs any prompt on a schedule, with tool permissions you control, per the official headless docs:

# crontab: nightly triage at 02:00, auto-approved read-only tools
0 2 * * * cd ~/src/myapp && claude -p \
  "Review open issues labeled 'bug', try to reproduce each against main, \
   and write findings to triage-$(date +\%F).md" \
  --allowedTools "Read,Bash(git *),Edit" >> ~/logs/triage.log 2>&1

Exit code 0 means the run succeeded; non-zero means it failed, so a wrapper script can branch, retry, or alert. Add --output-format json and the payload includes total_cost_usd per run — which is precisely the per-task cost visibility Duet doesn’t offer. Our Dynamic Workflows guide covers the multi-agent version of the same pattern.

What the cron approach cannot do is the other half of Duet’s job: it has no standing connection to your Gmail, Slack, or Stripe, no browser of its own in the cloud, no notion of “message me only when something needs attention,” and it stops existing between invocations. You can approximate pieces with MCP servers and more scripting, but “watch twelve accounts and interrupt me correctly” is a product, not a crontab line — and at $0 on top of a Pro plan, Duet is the cheapest version of that product anyone is shipping.

The two-tool answer for a developer on Max: schedule Claude Code against your repos, hand Duet the inbox and the invoices, and watch your usage meter, because both now drink from it.

What are the security guardrails — and what’s still unproven?

On paper, Duet launches with a stricter posture than most coding agents shipped with in 2025. Verified from the announcement and its October 23 security update: the trustworthy-agents framework shared with Claude Code and Claude in Chrome; the Managed Agents credential design so Duet never sees passwords or tokens directly; an auto-mode classifier approving actions; a dedicated classifier plus system prompt guarding against leaking sensitive data (credit card numbers and government IDs are refused outright; health and personal-finance data shared only on explicit instruction); external penetration testing; and coverage under Anthropic’s public bug bounty.

The unproven part is the one this site keeps flagging in its security series: an always-on agent holding live grants to thousands of connectors is a standing prompt-injection target, because the blast radius of one poisoned email or calendar invite is the union of everything you’ve connected. That’s the same structural concern we walked through for persistent agent memory in the MemGhost analysis, and a research preview has, by definition, not absorbed a year of adversarial traffic. The approval gates on send/pay/sign/delete are the right mitigations; whether the classifier holds against crafted inputs is something only time and the bug bounty will show. Connect low-stakes accounts first.

Where Claude Duet falls short today

Anthropic’s own limitations list is unusually direct for a launch page, and three more gaps matter for this site’s readers:

• It may misread a thread or misjudge urgency (Anthropic’s words), and anything arriving between its few-hours check-ins waits for the next cycle — this is polling, not streaming.
• Personal errands are weaker than work tasks — booking tickets is called out as a weak spot.
• You can’t watch it work. It can’t yet show you its browser, so verifying how it did something means trusting the summary.
• US-only, waitlist-gated, consumer plans only. EU and India readers — a third of this site’s audience — get nothing yet, and neither do Team/Enterprise seats.
• No repo awareness. No GitHub connector is named, no MCP support documented, no Claude Code handoff exists.
• No cost telemetry. Unlike a claude -p run reporting total_cost_usd, Duet gives you no per-task accounting against your limits.

Verdict: who should join the waitlist?

Join it if you’re a US developer on Claude Pro or Max already: the price is $0 on top, the morning brief and monitoring are a real product category that cron can’t replicate, and the worst case is you turn the monitoring frequency down or ignore the app. On Max specifically, the usage-limit concern mostly evaporates.

Skip it if you wanted a coding agent: Duet does not touch repos, and the rumor that it was a proposer/reviewer pair of Claudes is false. Your subscription already includes the better developer automation — Claude Code on a schedule — and if you want always-on and code-shaped, OpenAI Dots’ Codex delegation is the only shipping version of that today, at 5x the entry price.

Wait if you’re on Pro and already brushing its usage limits with daily Claude or Claude Code work. An always-on Opus 5.5 agent metering against the same pool is a real cost even though the sticker says $0 — our Max plan cost math covers when the $100 tier starts paying for itself.

And if the part of Duet you actually want is “an agent that runs without a cloud VM I don’t control” — that exists, it’s just DIY: an open-weight model behind Ollama driving a local agent on your own hardware, per runaihome.com’s local coding LLM guide and the self-hosted agent landscape at aifoss.dev. You give up the connector marketplace; you keep the machine, the data, and the meter.

FAQ

Is Claude Duet two Claude instances reviewing each other’s work? No. That framing circulated on aggregators after launch but appears nowhere in Anthropic’s materials. Claude Duet is a single proactive agent (one ongoing thread, one user) built around Claude Opus 5.5 that monitors connected apps and runs delegated tasks on a cloud VM.

Does Claude Duet cost extra on top of Claude Pro or Max? No. It’s included in Pro ($20/month, $17/month billed annually) and Max (from $100/month) with no separate fee — but its background checks and tasks consume your plan’s existing usage limits. Verified October 28, 2026.

Can Claude Duet access my GitHub repos or run Claude Code? Not as of October 28, 2026. No GitHub connector, MCP support, or Claude Code integration is documented. Anthropic says its own team ships code with Duet, but inside Duet’s VM — there’s no path into your repositories.

How is Claude Duet different from OpenAI Dots? Same category, different entry price: Duet comes with a $20/month Pro plan, Dots requires ChatGPT Pro at $100/month. Dots can delegate coding tasks to Codex; Duet has no coding delegation. Duet meters everything against plan limits from day one; Dots exempted dot usage for the first month, then bills delegated tasks against allowances.

Is Claude Duet available outside the US? No. The research preview is waitlist-only for Pro and Max subscribers in the US. Team and Enterprise plans are excluded, and Anthropic has announced no international timeline.

Can I automate coding tasks without waiting for Duet? Yes, today, on the same subscription: claude -p "<task>" --allowedTools "..." in a crontab runs Claude Code non-interactively on any schedule, exits non-zero on failure, and reports per-run cost with --output-format json. See our Dynamic Workflows guide.

Sources

• Introducing Claude Duet — Anthropic (official announcement)
• Claude Duet product page and waitlist — claude.com
• Claude plans and pricing — claude.com
• Run Claude Code programmatically (headless mode) — Claude Docs
• Anthropic Just Introduced Claude Duet — Towards AI
• Claude Duet launch coverage and Meta Muse / OpenAI Dots positioning — TechNews

Last updated October 28, 2026. Claude Duet is a research preview — access, features, and metering behavior will change; verify current state at claude.com before relying on it.

Know which coding tool is worth paying for

Hands-on comparisons of AI coding assistants and what each one costs to run — including the local-model path. Sent only when something changes. Unsubscribe anytime.

Email address Subscribe