Opus 4.8 fast mode ends November 30 in Claude Code

Claude Code 2.1.316 puts fast mode for Opus 4.8 on a deadline: Anthropic deprecated it on October 30 and will remove it on November 30, the v2.1.316 release notes on GitHub said on October 31. After that date, fast mode requests on Opus 4.8 return an error with no fallback to standard Opus 4.8, and the notes name Opus 5.5 as the model that keeps the speedup.

Any setting, script or API call that pairs fast mode with Opus 4.8 has two ways out before the cutoff. It can point at Opus 5.5 and keep fast mode, or turn fast mode off and stay on Opus 4.8 at standard speed.

The move the notes recommend also cuts the per-token price. Anthropic’s pricing page lists fast mode on Opus 4.8 at $10 per million input tokens and $50 per million output, and fast Opus 5.5 at $8 and $40, a fifth less on both sides. On each model fast mode costs twice the standard rate, which is $5/$25 for Opus 4.8 and $4/$20 for Opus 5.5.

Fast mode itself works the same on either model. The Claude Code documentation describes it as “a different API configuration that prioritizes speed over cost efficiency,” with “identical quality and capabilities,” and puts the gain at “up to 2.5x higher output tokens per second.” That figure is a ceiling, and it measures output speed rather than the time to the first token.

Opus 4.7 went through the same cycle this summer. Its fast mode was deprecated on June 25 and removed on July 24, and Opus 4.7 now rejects fast requests outright, with no fallback to standard speed. Opus 4.8 again gets only about a month between the notice and the error.

Effort, ultrathink and subagents

CLAUDE_CODE_SUBAGENT_EFFORT sets the effort level subagents run at when nothing else names one, where before they inherited the session’s. CLAUDE_CODE_DISABLE_ULTRATHINK turns the ultrathink keyword off. In VS Code, ultrathink now runs its turn at xhigh effort and shows in a rainbow in the chat box, matching the terminal, where xhigh ultrathink arrived in 2.1.311 earlier in October.

Sub-agents that inherited a parent sub-agent’s model also stop drifting to the main model when a background job or a message wakes them.

Bash, restricted sessions and cloud models

• --restricted sessions no longer upload files from outside the working directories when a Read allow rule covered them.
• On Vertex AI and Amazon Bedrock, sonnet, opus and haiku switch to an older model when the newest is not yet available to the account or region, instead of showing an error mid-session.
• Shell commands are no longer left running on macOS and Linux when Claude Code exits by itself, for example when credentials are refused or a cloud session is archived.
• Bash scripts started from the Bash tool no longer stop after their first command when SHELLOPTS is exported.
• PowerShell hooks now block on exit code 2 when the command begins with a type cast, param, using or an attribute.
• CLAUDE_CODE_DISABLE_LOOPED_CROSS_CERT_DROP keeps system cross-certificates that certify each other in a loop, for servers whose private CA needs one.

What gateway admins get

The Claude apps gateway adds pooled group budgets: one monthly budget that a group’s members share on top of each member’s own limit, set with pooled_budget on the spend limits admin API. With managed.policy_mode: layered, every policy that matches a developer applies, each merged over the ones listed below it, instead of only the first. On Bedrock upstreams, guardrail.input_tagging wraps prompts in the input tags that an Amazon Bedrock guardrail’s prompt attacks filter needs, and all_messages_and_system tags the system prompt as well.

One gateway fix will show up as new errors. The gateway now checks every model a request names against availableModels, deniedModels and models, and a request that names a blocked model gets a 400 naming the field. Where managed settings require a gateway, claude auth login now runs and shows the gateway sign-in in a terminal, and VS Code users can sign in from the chat panel without a terminal.

Team and Enterprise logins change too: server-managed settings fetched from Anthropic’s API now apply without an approval prompt, and a stored copy that doesn’t verify is refused.