Claude Duet Pricing: What a Proactive Agent Costs to Run

Last updated: October 2026

Key takeaways

• Claude Duet is a proactive agent, so cost scales with wake-ups, not with conversations.
• Anthropic has not published Duet pricing, and it is a research preview for US Pro and Max subscribers on a waitlist.
• In an illustrative model, checking every 15 minutes for 16 hours a day costs about $72 per user per month before caching.
• The new Claude API version 2026-10-21 is a versioning change, and no price change was reported.
• Budget with a polling-frequency lever, a caching assumption and a hard cap per user.

What is Claude Duet?

According to TLDR AI's October 22 edition, Claude Duet is Anthropic's proactive agent. It is powered by Opus 5.5, monitors your connected apps, prioritizes tasks and works on its own, asking for approval before certain actions. It is a research preview, open first to US Pro and Max subscribers through a waitlist.

The pricing lesson is not about Duet itself. It is about a product shape that is spreading: agents that run without a user typing anything. If you build one, every background check is a billed API call that no customer asked for.

Does Claude Duet have a price?

Not in the coverage reviewed. No usage limits or per-token rates were reported for Duet, so any dollar figure you see is a model, not a quote. For reference, Anthropic's own rates for the API live on the Anthropic pricing page on Calcaas.

The useful question is not "what will Duet cost?" but "what would an agent like Duet cost me to run?" That one you can answer today.

What does a proactive agent cost per user per month?

Here is a simple model. Say your agent wakes up on a schedule, reads about 6,000 fresh input tokens from connected apps, and writes about 300 tokens of summary or actions. Assume it is active 16 hours a day for 30 days. We use round placeholder rates of $5 per million input tokens and $25 per million output tokens. These are illustrative, not Anthropic's price list.

| Check frequency | Checks per month | Input tokens | Output tokens | Cost per user (illustrative) | |---|---|---|---|---| | Every 60 minutes | 480 | 2.88M | 0.14M | $18.00 | | Every 15 minutes | 1,920 | 11.52M | 0.58M | $72.00 | | Every 5 minutes | 5,760 | 34.56M | 1.73M | $216.00 |

The jump from hourly to five-minute polling multiplies the bill by 12 without changing the model, the prompt or the user. That is the core pricing risk of proactive agents: the cost driver is a cron setting, not customer behavior.

How do you cut the bill?

Three levers matter most. First, cache the stable part of the context. If 80% of the input is a repeated prefix billed at a tenth of the normal rate, the 15-minute scenario drops from about $72 to about $31 per user (an assumption you should test against your provider's real cache pricing). Second, use cheap triage: a small model decides whether anything changed and only wakes the big model when it did. Third, cap spend per user per day so one runaway loop cannot eat a month of margin. Our post on cutting agent token spend by 80 percent walks through the same pattern.

What does the new Claude API version change for your costs?

The same newsletter reports that the Claude API shipped version 2026-10-21, its first official version release since 2023, with new major SDK and CLI versions that default to it. Anthropic says it will release versions more often so developers can handle breaking changes before a new model requires them. Version 2023-06-10 is not deprecated, and no pricing change was reported.

For a cost model, the practical point is operational: pin your API version, read the migration guide, and re-run your cost checks after upgrading SDKs, because a default change can alter request shapes. If you want the per-task view of a flagship model, see our breakdown of Claude Opus 5 cost per task.