Claude Code Max Effort Level Setting: Cap AI Spend (2026)

Add maxEffortLevel to your Claude Code settings file and no session on that machine can run above the effort ceiling you set — Anthropic's changelog for v2.1.267 (9 September 2026) describes it as a cap that applies "on every provider, including Bedrock, Vertex and Foundry", while users can still pick a lower level. That single line is the claude code max effort level setting, and if you pay for AI out of your own margin — agencies, solo operators, anyone running Claude Code across client work — it is the closest thing yet to a built-in spend guardrail for effort.

📺 Watch: The NEW Claude + Agent OS is Crazy Good!

🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →

What the Claude Code Max Effort Level Setting Does

Effort levels control how hard Claude Code works on each request — higher levels buy more thinking and more thorough work, and the top xhigh level was introduced as the recommended setting for the hardest coding tasks, per Anthropic's own What's New digests (Week 16 and Week 22, 2026). The catch has always been the bill: effort you did not need is money you did not need to spend. The claude code max effort level setting closes that gap from the top down. Per the official changelog entry, maxEffortLevel "caps the effort level on every provider", and it can be set either top-level or per model under modelSettings. Anthropic's Week 37 What's New digest (7–11 September 2026) confirms the same behaviour in one line: "the maxEffortLevel setting caps the effort level on every provider."

Read that carefully, because the design is a ceiling, not a dial. Anyone using the machine can still drop effort lower for quick tasks — the changelog says users "can still pick a lower level" — but nobody can push it above your cap. That asymmetry is exactly what you want in a team or client context: savings stay available, overspend does not.

Why a Cost Ceiling Beats Policing Usage

If you run Claude Code across a team, a VA, or a fleet of automation machines, you have two options for controlling effort-driven spend. Option one is a written rule — "keep effort at medium unless you ask me" — which survives until the first deadline. Option two is the max effort level setting: a config line that makes the rule physically enforceable, on every provider you route through, whether that is Anthropic directly or Claude models on Bedrock, Vertex or Foundry. For anyone selling AI-powered deliverables, that provider-wide scope is the notable part — your cap follows the account setup you actually run, not just one API endpoint. It is the same logic covered in the GPT Live 1 pricing breakdown and the DeepSeek V4.1 Flash API pricing guide: the profit in AI work is made or lost in the gap between what a task needs and what you paid for it.

If you want AI running your business at a cost you control — the exact configs, prompts and daily tutorials Julian uses — check out the AI Profit Boardroom. Want a personal cost-and-growth plan instead? Book a free SEO strategy session and get it mapped 1-on-1.

How to Set maxEffortLevel

The changelog gives you two placements, and they serve different jobs:

• Top-level cap. Set maxEffortLevel once in your Claude Code settings and it applies across the board. This is the simple case: one ceiling for the whole machine or profile.
• Per-model cap under modelSettings. Set maxEffortLevel for a specific model instead. This lets you run a sensible split — for example, allowing higher effort on a cheaper model while capping effort on premium models where each notch up costs real money.

The pattern worth copying for client work: cap premium models conservatively, leave working models room to stretch, and lift the cap deliberately on the rare project that justifies it. Because it is one settings line, raising the ceiling for a hard week and lowering it again afterwards takes seconds — which means your default state is protected, and expensive states are opt-in. That is the reverse of how most people run it, and it is why their margins leak. If you are still getting oriented in the tool itself, learning Claude Code covers the settings files this lives in, and if you have not yet turned Claude Code into an actual business system, the agentic OS approach to Claude Code is where a capped, predictable setup starts to compound.

Claude Code Max Effort Level Setting for Automation Fleets

The setting gets more valuable the further your usage drifts from a human at a keyboard. Scheduled runs, batch jobs and always-on agents do not notice when a task escalates effort — they just burn what the settings allow, run after run. Putting a max effort level on automation machines converts an open-ended liability into a known unit cost per run, which is the number you actually need before you can price the output. The same principle shows up in the Gauntlet Loop workflow — repeatable AI output only becomes profitable when the cost per cycle is nailed down — and in the guide to using Codex free, which attacks the other side of the equation by pushing tool cost towards zero.

One honest caveat: capping effort caps one driver of spend, not all of them. Token volume, model choice and how many sessions you run still matter — a low-effort session that loops all night can out-spend a high-effort session that finishes in ten minutes. Treat maxEffortLevel as the guardrail it is, alongside model selection — Goldie Bench covers how the current model brains compare in hands-on tests, which is the other half of the cost-per-result decision — rather than a full budget system. For rate-limit and quota behaviour on your plan, Anthropic's own usage documentation remains the source of truth; this setting governs effort, not billing itself.

Who Should Cap Effort — and Where

Wherever you land, set the cap before the surprise invoice rather than after it. The pattern across AI pricing pages — the Abacus AI Smaug models included — is that costs are lumpy and announcements move prices without warning; a ceiling you configured is the one variable that stays exactly where you put it.

The cap also simplifies scaling decisions. When you add a second automation machine or hand a workflow to a VA, the marginal cost question is usually a guess; with a ceiling in place, the worst case per session is a number you chose in advance. That is the mindset behind Agent OS — treat your AI operation as a system with known inputs and outputs, not a collection of ad-hoc chats — and effort caps are the budget arm of that system. A practical starting point if you have never touched effort settings: check what level your sessions actually run at today, cap at that level so nothing regresses upward silently, and revisit after a week of real invoices. Most people find the cap never binds on routine work — which is precisely the evidence that higher levels were costing money without changing outcomes on those tasks. The days it does bind are the days you decide, deliberately, whether the task in front of you is worth the uncapped rate.

Quick Answers on the Max Effort Level Setting

When did it ship? Claude Code v2.1.267, released 9 September 2026, per the official Anthropic changelog.

Does it work outside Anthropic's own API? Yes — the changelog states the cap applies on every provider, naming Bedrock, Vertex and Foundry.

Can people still choose lower effort? Yes. It is a maximum, not a fixed level — the changelog is explicit that users can still pick a lower level.

Does it guarantee a fixed bill? No — it caps the effort dimension of spend. Session count, tokens and model choice still drive the total.

If you want your AI stack making money instead of quietly eating it, the AI Profit Boardroom gives you Julian's full Agent OS, prompt libraries and weekly coaching calls — get inside the AI Profit Boardroom. Prefer to start with a 1-on-1 audit of your setup? Book a free SEO strategy session.

Real wins from inside the AI Profit Boardroom

See all 3,000+ members →

← Back to all posts