Anthropic shipped a feature on August 7 that quietly addresses one of the most common objections enterprises raise when deploying AI agents: what happens when a session runs away with your budget?

The answer is now a hard dollar cap. Developers building on Claude Managed Agents can set a budget parameter at session creation, specifying a maximum list cost in US dollars. When the session’s accumulated spend reaches that ceiling, it pauses rather than terminates, preserving all state and history. Raise or remove the cap, and work resumes automatically. No rebuild, no lost context.

The pricing covers three consumption types: model tokens at public list rates, web searches at $10 per thousand, and session running time at $0.08 per hour. The cap is enforced between model requests, not mid-request, so the final recorded cost can land a small amount above the limit. Anthropic describes this as expected and bounded by one model request per thread.

The same budget object can be attached to scheduled deployments, where it applies independently to each run. That means an automated agent that fires nightly can be capped per-run rather than against a cumulative total.

The Context Behind This

AI cost overruns have become a real operational concern. Gartner predicted in mid-2025 that over 40 percent of agentic AI projects would be cancelled by the end of 2027, citing unclear ROI and inadequate risk controls as the primary drivers. Cost predictability sits at the centre of that risk picture.

When an agent session is doing open-ended work, such as researching, writing, calling APIs, and iterating, there is no natural finish line the billing system can see coming. Without a cap, the session runs until the task is done or until someone notices a large invoice. For experiments, that is tolerable. For production systems processing hundreds of sessions daily, it is not.

The session budgets feature makes cost controls a first-class part of the agent runtime rather than an afterthought in the billing dashboard. Organisations that have been reluctant to run agents at scale due to unpredictable spend now have a mechanism that works at the infrastructure level.

Key Behaviour Details

A few specifics matter for production deployments:

Sessions pause, they do not terminate. This distinction is significant. A terminated session loses its context and requires a restart. A paused session sits idle with all its history intact, waiting for a budget update. For long-running tasks that are mid-way through meaningful work, this is the difference between a cost guardrail and a data loss event.

The cap cannot be added retroactively. It must be set when the session is created. If a session is created without a budget, no budget can be added later. Teams should design their session creation logic with this in mind from the start.

Once removed, a budget cannot be re-added. If a cap is cleared from a session, that session runs without limits for the rest of its life. The recommendation is to change the cap rather than remove it if ongoing cost control is needed.

Multiagent sessions share a single budget. In a multiagent setup where one orchestrator spawns subagents, all of their consumption draws from the same cap. There are no per-thread limits within a session.

What This Means for Business

For any business running Claude agents in production, this feature changes the risk calculation.

Previously, the only way to prevent runaway agent costs was to monitor usage after the fact and hope no individual session caused significant damage before it was noticed. That approach works in low-volume environments but becomes fragile quickly as agent usage scales.

With session budgets, the cost ceiling is part of the deployment configuration. A customer support agent that should cost no more than a dollar per conversation can be capped at that amount. An overnight data processing job can be bounded at a number that reflects its expected scope. A one-off research agent can be capped at an experimental budget without requiring manual supervision.

This also changes how businesses can think about AI cost as a line item. Instead of a variable expense that depends on what agents decide to do, it becomes a controllable parameter set by the team deploying the agents.

The deeper implication is about trust. Enterprises that have been waiting to see whether AI agent costs are manageable before committing to broader deployments now have a concrete answer. The spend can be bounded at the session level, per run, before the work starts.

For organisations building on Claude through Enterprise DNA’s AI services, this kind of predictable cost structure is what makes moving from pilot projects to production systems feasible at scale.

---

Anthropic’s full release notes are available at platform.claude.com. The session budgets documentation is at platform.claude.com/docs/en/managed-agents/budgets.

If you are evaluating AI agent deployment and want to understand how to structure cost controls and governance for your organisation, book a session with the Enterprise DNA team.