Claude Code Auto Mode — August Rollout and Current Defaults

Checked 1 October 2026: Anthropic's current Claude Code docs now describe the built-in starting mode by version, interface, session type, provider, and organization policy. The v2.1.283-or-later default for interactive terminal and VS Code sessions supersedes the earlier plan-specific rollout described below. The subscription-limit arithmetic and BAA/retention section retain their 26 September review date and should be read against the source dates stated in those sections.

From 14 August 2026, new Claude Code sessions on Pro, Max, and Team began starting in Auto mode: a classifier screens some tool calls such as shell commands and network requests instead of prompting on every step. That was the initial rollout, not the complete current rule. Current Claude Code documentation says v2.1.283 or later uses Auto as the built-in starting mode for interactive terminal and VS Code sessions, unless settings or organization policy change it. claude -p/Agent SDK, other providers, Desktop, and web sessions have distinct rules; check the current permission-mode documentation and /status for the specific session. Auto classifier overhead is free on Pro, Max, and Team; where server-side checks are available for other sessions, they are also free, but gateways or staged platform/region rollouts can cause billed local-classifier fallback.

This is a setup and risk article for UK buyers who already shortlisted Claude Code, or who are deciding between Claude Code, Cursor, and Copilot for the next quarter. Full product caveats live on the Claude Code harness page and the setup guide.

What Auto mode changes day to day

In Manual-style flows, Claude Code asks before many tool actions. That is safer for unfamiliar repos and production scripts, and slower for routine edits. Auto mode uses a classifier so ordinary read/edit/test loops can proceed with fewer confirmations. The point is less click-fatigue on trusted work, not “the agent may rewrite production without governance.”

You can still switch modes (Shift+Tab / UI). Pinned admin defaults win on Team and Enterprise. If your security policy requires human confirmation for shell, network, or deploy-adjacent tools, keep Manual or Plan for those repos and document the exception path. Auto reduces friction; it does not erase blast radius.

Treat Auto like autopilot on a familiar route: useful when the map is correct, dangerous when the repo is alien or the credentials are over-scoped. Pair Auto with branch protection, required reviews, and a habit of reading diffs before merge — the same discipline you should already use with Cursor Agent or Copilot coding agent.

September update: the classifier moved to the server

Claude Code v2.1.278, released on 19 September 2026, introduced server-side Auto mode checks for additional account types. When eligible, the server performs checks as part of the session's own model requests at no charge. Eligibility depends on platform, region, credentials, and rollout; this is not a universal Enterprise/API guarantee.

What that means in practice:

• Direct connections to the Anthropic API can use free server-side checks when the session meets Anthropic's current version and rollout requirements; /status is the session-level check. The current billing docs note that direct connections require Claude Code v2.1.281 or later for the server-control setting.
• Gateways and proxies are the usual reason it fails. If your gateway strips the safeguards request field or drops the safeguard_results response field, Claude Code falls back to its own classifier and bills it as before. The first time that happens, the session shows a one-off notice. Nothing breaks.
• Check the status. Run /status and look for the Auto mode server row: Enabled means the server is making the calls, Disabled means the session fell back to Claude Code's billed classifier requests. Current docs describe the no-charge path and eligibility in detail.
• Opting out. Teams that know their gateway cannot pass the fields through can set CLAUDE_CODE_AUTO_MODE_SERVER=0. Anthropic describes this as a temporary setting that may be removed.
• Pro, Max and Team never see the notice; their classifier overhead was already free.

Current Anthropic documentation now publishes both the gateway compatibility requirements and the versioned default-mode rules. Do not use the rollout predictions previously reported by third-party gateway documentation as a product deadline; check Anthropic's changelog and current classifier-billing page before scheduling a migration.

The same week, v2.1.277 (18 September) started reading AGENTS.md in projects that have no CLAUDE.md. That matters if several coding agents share one repository.

Weekly limits after the boost

Anthropic's temporary +50% weekly usage boost was extended several times over the summer and finally ran to 13 September 2026. From 14 September, standard weekly limits in Claude Code are permanently 25% higher than the pre-boost baseline for Pro, Max, Team, and seat-based Enterprise plans. Anthropic's own follow-up post put it plainly: compared with the boosted level, that is a 17% reduction (125 ÷ 150 ≈ 0.83).

Then Claude Opus 5.5 launched on 22 September with two more subscription changes: higher five-hour usage limits on Pro, Max, Team, and seat-based Enterprise, and a rate-limit reset that subscribers can save and use when they choose. Anthropic has not published per-plan token figures for the new weekly envelope.

So forecast seats against the +25% level. Do not use the summer's +50% level. If your team already hits weekly walls on Max, Auto will not invent more tokens. It may change how quickly you burn them by removing idle confirmations, so watch usage for a week before buying more seats. The full seat-forecasting maths is in Claude Code limits after 14 September.

Enterprise, BAA, ZDR, and Covered Models

Claude Code under Anthropic's BAA requires zero data retention (ZDR). The catch is Covered Models. These are models Anthropic designates for extra safeguards, and they carry a 30-day minimum retention that rules out ZDR. The current list is Claude Fable 5, Fable 5.1, Mythos 5 and Mythos 5.1. Opus is not on it. Opus 5 and the new Opus 5.5 stay ZDR-eligible: Anthropic's launch page says "like previous Opus models, Opus 5.5 is available with zero data retention."

For regulated buyers, that changes the practical choice:

• BAA + ZDR with Opus 5.5 is a normal configuration. Opus 5.5 is also Anthropic's current recommendation for long-running agentic coding.
• Fable 5 / 5.1 (Covered) sits outside ZDR under the standard policy. Anthropic's "Covered Models under Anthropic's BAA" article sets out which configurations can use Covered Models as eligible services, so check it before you promise Fable to a HIPAA workload.
• Transition option. Anthropic is introducing Enterprise Frontier Safeguards (EFS), which lets customers keep retained monitoring data in cloud infrastructure they control. EFS rolls out in phases from fall 2026. Until then, eligible customers can get a time-limited ZDR option for Fable 5 and 5.1 for internal business applications.

Write the model choice into the procurement pack so legal and engineering stop arguing past each other.

Enterprise keeping Manual as its built-in default is intentional caution. Security teams should pilot Auto on non-production repos, define which tool classes always require Manual, and only then set org defaults. “Consumer Pro defaults to Auto” is not a compliance argument.

What to do now

1. Read the setup guide and confirm your team's current default mode.
2. Decide which repos stay Manual forever (infra, payments, secrets-adjacent).
3. If Claude Code runs behind a gateway, check /status for the Auto mode server row and ask the gateway owner to pass safeguards / safeguard_results through untouched.
4. Re-forecast weekly usage at the permanent +25% level, then fold in Opus 5.5's higher five-hour limits.
5. If you need a BAA, write the model choice into the ticket: Opus 5.5 with ZDR, or Fable under the Covered-Model terms and the EFS transition.
6. Compare Auto behaviour against your Cursor or Copilot agent policy so humans get one rulebook, not three.

Common mistakes

• Equating Auto with “safe for production deploys.”
• Sizing annual seats on the summer's +50% boost.
• Assuming Enterprise inherits the Pro/Max/Team default flip automatically.
• Calling Opus a Covered Model. It is not; Fable and Mythos are.
• Leaving a gateway in place that strips the safeguards field, then paying for classifier calls that are now free on a direct connection.
• Forgetting that classifier-free billing still leaves model tokens on the meter.

Bottom line

Auto as default was first announced for Pro, Max, and Team on 14 August 2026; current defaults are version-, surface-, and policy-specific. Server-side classifier checks can remove the additional classifier charge only where Anthropic's documented eligibility and /status checks confirm that path; gateway or rollout fallback can still be billable. Weekly limits settled at +25% over the stated pre-boost baseline, without a published per-plan token quota. Opus 5.5 is ZDR-eligible under Anthropic's stated conditions; confirm account qualification and current terms before relying on that for a BAA workflow. For harness shortlists, start with the AI coding assistant finder.

Sources (checked 2026-10-01 unless dated otherwise)

• Anthropic — Introducing Claude Opus 5.5 (22 Sep 2026: $4/$20, ZDR availability, five-hour limits, saveable rate-limit reset): https://www.anthropic.com/claude-opus-5-5
• Claude Help Center — Covered Models (Fable 5/5.1, Mythos 5/5.1; 30-day retention; EFS; time-limited ZDR for Fable): https://support.claude.com/en/articles/15425695-covered-models
• Anthropic — Claude API pricing (Opus 5.5 current; Opus 5 in the prior-models group): https://platform.claude.com/docs/en/about-claude/pricing
• Anthropic — Claude Code documentation: https://docs.anthropic.com/en/docs/claude-code
• Anthropic — Choose a permission mode (current built-in defaults by version, surface, plan, and session type): https://code.claude.com/docs/en/permission-modes — accessed 2026-10-01
• Anthropic — Auto mode classifier request charges (eligibility, gateway behavior, /status, and fallback billing): https://code.claude.com/docs/en/auto-mode-classifier-billing — accessed 2026-10-01
• Hacker News thread quoting @ClaudeDevs (+25% from 14 Sep; “17% reduction” compared with the boost): https://news.ycombinator.com/item?id=49491631
• The Economic Times — boost ends 13 Sep, +25% from 14 Sep (31 Aug 2026): https://m.economictimes.com/ai/ai-insights/claude-code-users-to-get-17-less-weekly-usage-from-september-14-as-anthropic-ends-50-boost/articleshow/133647006.cms
• MIXED — Claude Code v2.1.278 server-side classifier and v2.1.277 AGENTS.md support (21 Sep 2026): https://mixed-news.com/en/claude-code-stops-charging-auto-mode-safety-checks-agents-md/
• LiteLLM — Claude Code server-side auto mode through LiteLLM (safeguards contract; reported 25 Sep and 23 Oct dates): https://docs.litellm.ai/blog/claude-code-server-side-auto-mode