Anthropic has published a guide to building always-on agents, AI agents that run indefinitely and act without being asked, and says cheaper intelligence is part of why they are coming back. On one long-horizon coding benchmark, it says, “Opus 5.5 cost 78% less per solved task than Opus 5, compared with 40% less on typical workloads”.
The company’s cost documentation puts numbers on it: on its SWE-bench Pro subset, a solved task cost $1.01 on Opus 5 and $0.22 on Opus 5.5, 4.6 times cheaper by our arithmetic, at about the same score.
Business Pill · Switching Costs Explained
An evergreen Business Engineer explainer on switching costs: why it is hard to leave an AI supplier. It teaches the general idea only and does not discuss always-on agents.
The key insight: As we read the guide, Anthropic is tying a product pattern to its price curve: agents that wake themselves up only make sense when a solved task costs cents, and its documentation puts Opus 5.5 at $0.22 against $1.01 for Opus 5.
The guide, written by William Hu and dated 13 November, says “Over the past few months, we’ve seen a resurgence of always-on agents”. It defines them as “stateful and proactive, meaning they manage context across sessions and take actions without direct supervision”.
It gives two reasons for the return. The latest models “have better judgment”, it says, and “intelligence is becoming cheaper, with long-horizon tasks benefiting the most”. Anthropic writes that customers “have started to ask us how to integrate this type of experience into their own products”.
Cost per solved task on Anthropic’s SWE-bench Pro subset, from the Claude Platform documentation. Anthropic’s own tests; Opus 5.5 scored about the same as Opus 5.
The Price of a Solved Task
The cost page lists Opus 5.5 scoring 92.8% of tasks solved on the subset against Opus 5’s 91.9%, for 78% less per solved task, or “about 69% less at Claude Opus 5’s prices”. Sonnet 5.5 cost $0.11 per solved task there, half Opus 5.5’s figure by our arithmetic, after an 87% cut from Sonnet 5’s $0.84.
The long-horizon saving is about 1.9 times the typical one, by our arithmetic. The documentation’s advice is to “Compare on cost per solved task, not per token”, the same yardstick used in Garry Tan’s open-source harness test published on 14 November.
Five Parts of an Always-On Agent
The guide says the architecture around an always-on agent “has five core primitives: an event inbox, timers, memory, permissions, and an outbox”. The inbox queues events from webhooks, schedules and users so a worker can run them one at a time; timers cover self-set follow-ups and recurring schedules; memory lives in its own durable store.
On permissions, it says “Permissions are where always-on agents carry the most risk”, and recommends a credential proxy outside the agent’s environment, classifier-based auto mode and deferred tool calls for user approval. On messages, it advises that “it’s better for the agent to start quiet and then earn more of the user’s attention over time”.
The Structural Read
As we read the five primitives, most of the work is plumbing outside the model: an inbox, timers, a memory store, a permission layer and an outbox, with the agent loop itself unchanged.
As we read the permissions section, Anthropic names the main risk itself: a run holds the user’s access and operates unattended, so it recommends a credential proxy, auto mode and deferred approvals.
Anthropic, Building always-on agents, 13 November 2026
“Each run is brief, but the cycle keeps going indefinitely, and that’s what makes the agent feel like it’s always on.”
Three Implications
FOR TEAMS BUILDING AGENTS Anthropic recommends one inbox per user and only one active worker per inbox at a time.
FOR FINANCE TEAMS On Anthropic’s SWE-bench Pro subset, a solved task cost $0.22 on Opus 5.5 and $0.11 on Sonnet 5.5, per its documentation.
FOR SECURITY TEAMS The guide says permissions are where always-on agents carry the most risk and recommends a credential proxy outside the agent’s environment.
The framework argues: “The surface, not the model, governs autonomy.”
As we read it, Anthropic’s guide is a blueprint for that surface: the five primitives sit around an unchanged agent loop. The framework does not discuss always-on agents.
Cost-per-solved-task savings by model upgrade, from Anthropic’s documentation and its 13 November guide. Benchmarks differ by row.
What Is Not Established
The cost figures are from Anthropic’s own tests on its chosen benchmarks; the documentation notes the size of the saving “depends on the work”. The guide is a set of design recommendations, not a product launch, and it gives no pricing change or customer count for always-on agents.
We read the guide and the linked cost documentation; Anthropic was not contacted for this piece.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Anthropic’s new guide sets out five building blocks for agents that run without being asked, and ties their return to cost: on its coding benchmark, a solved task fell from $1.01 on Opus 5 to $0.22 on Opus 5.5, a 78% cut, against 40% on typical workloads.
A note on sourcing. The guide is Anthropic’s article of 13 November 2026; cost figures come from the Claude Platform documentation it links, read on 14 November 2026, and reflect Anthropic’s own tests. Multiples are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
