Bottom line: Claude Code switches to Auto Mode as the default for Pro, Max, and Team users starting August 14, with Anthropic stating that its own testing shows this achieves equal or better safety than manual approval while no longer billing the previous classifier overhead.
Anthropic is making Auto Mode the default in Claude Code for Pro, Max, and Team users, while simultaneously waiving the previous additional costs for the classification logic used in the process. Users who have already set their own default will be asked once whether they want to switch, while those with a pinned setting remain unaffected.
Starting August 14, 2024, new sessions in Claude Code will launch in Auto Mode by default for users on the Pro, Max, and Team plans. Users who have configured a different default themselves will receive a one-time prompt asking whether they want to switch; those who have configured a pinned default are not affected by the change. The classifier used for Auto Mode requires a small number of additional tokens per tool call – Anthropic will now cover these costs itself for Pro, Max, and Team users going forward. For Claude Enterprise, the Claude API, Claude Platform on AWS, Amazon Bedrock, Google Cloud’s Agent Platform, and Microsoft Foundry, Auto Mode remains opt-in for now, but is expected to become the default there as well within a month in collaboration with the cloud partners, again without billing the classifier overhead. Enterprise admins can already set Auto Mode as the default via Managed Settings.
Technically, Auto Mode replaces individual confirmation dialogs with a classifier that evaluates every tool call and specifically blocks irreversible, destructive actions or actions targeting systems outside the user’s own environment. If the classifier blocks an action, Claude will typically look for a safer approach on its own or ask directly for approval. If no agreement can be reached – after three consecutive blocks or twenty within a session – Claude Code automatically falls back to manual approvals. Permission rules continue to take precedence over the classifier, with the exception of allow rules that are defined so broadly that they permit arbitrary code execution.
Anthropic justifies the step with several months of testing, including internal and external red-teaming, prompt injection evaluations, a controlled study with 1,053 paying testers, and an analysis of real production sessions. Across all metrics examined, Auto Mode matched or exceeded manual review. As supporting data for the decision, Anthropic notes that users confirm 97% of permission prompts in Claude Code, while they reject Claude’s plans for approval in 39% of cases – an indication that individual approval requests are often clicked through reflexively rather than actually reviewed. As of June 2026, according to Anthropic, 49.5% of active CLI users have manually set up a Bash allow rule, 5% blanket-allow every shell command, and a further 43% use interpreter rules such as Bash(python:*) or Bash(node:*), which in practice are similarly broad. This share is said to be growing by about 5 percentage points every 5 weeks. In addition, 62% of users have used bypassPermissions or clicked “don’t ask again” for Bash commands, and 25% of interactive sessions start directly in bypass-permissions mode.
For development teams, Auto Mode primarily means longer autonomous runtimes without interruption, which is especially relevant for models designed for long-running tasks such as Claude Opus 5, allowing them to work on larger tasks for several hours at a stretch. According to Anthropic, Auto Mode users among Team and Enterprise customers deliver around 25% more pull requests. Teams at Adobe, Nuro, Gusto, and Garner Health are already using Auto Mode in production as their default, according to Anthropic. Anyone who does not want to adopt the change can continue to configure a different default mode or set a pinned setting.
Source: claude.com · Published August 6, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.