We’ve deprecated fast mode for Claude Opus 4.8, with removal on November 30, 2026
Requests to claude-opus-4-8 with speed "fast" will error after removal, and Claude Opus 5.5 supports fast mode with no beta header.
Every Anthropic update in one feed — release notes, features and changes, straight from the source.
59 updates
Requests to claude-opus-4-8 with speed "fast" will error after removal, and Claude Opus 5.5 supports fast mode with no beta header.
Token usage and cost reporting for Console and Enterprise organizations moves to new endpoints, while the older usage and cost APIs stay supported as legacy.
Messages and Token Counting requests on the first-party Claude API can now be four times larger, while partner platforms keep their own limits.
Anthropic filed an S-1 with the SEC for a proposed Nasdaq IPO under the ticker ANT; the share count and price range are not set yet.
Google Cloud customers can now use the full Claude Platform, including Managed Agents and code execution, billed through their Google Cloud account.
Anthropic runs the Claude API inside a Google Cloud project, with IAM for access, Marketplace for billing, and fast mode for Opus 5.5.
Anthropic released Claude API version 2026-10-21 and plans more frequent versions. The 2023-06-01 version is not deprecated.
Python, TypeScript, Go and other SDKs plus the ant CLI move to new major versions, now under Apache 2.0, and user profiles gain enforcement fields.
Claude Enterprise owners can send each governed prompt to their own security server for an allow or deny verdict before the model runs.
Enterprise admins can read and change Claude Tag's Slack channel and workspace settings with an Admin API key, in beta.
Claude Enterprise admins can read members' saved Instructions for Claude and Cowork Global instructions through a beta Compliance API endpoint.
Opus 5.5 fast mode needs no beta header or access request, and every usage tier now gets a default rate limit.
Opus 5.5 fast mode is available on Azure-hosted Claude deployments in Microsoft Foundry, which support Global Standard and US Data Zone Standard.
SDK tracing supports Python, TypeScript, C#, Go, Java, PHP and Ruby, while admins configure API exports in the Claude Console.
The beta lets agents run multi-agent programs in the background, with progress events and session budgets to track and cap the work.
The beta is available to Claude Enterprise organizations and works with their existing Compliance Access Key.
Developers supply their own automation methods, while the SDK handles the tool loop, browser policies and approval callback.
Max subscribers receive $100 or $200 USD a month, while Team credits are pooled and capped at $500 USD.
The model has a 1M token context window, up to 128,000 output tokens and configurable thinking effort.
Cache reads now cost 0.05x the base input price, while cache writes and all other prices remain unchanged.
Blocked hosts produce fetch errors and disappear from search results; unrestricted networking and self-hosted environments are unaffected.
Separate fields distinguish code execution tool support from the ability of running code to call other tools.
Both model endpoints report support for turning thinking off, though an incompatible effort level can still cause a request to fail.
Developers can group models in a model picker without parsing IDs, with Claude Opus 4.5 and 4.6 both reporting opus.
Retirement on the Claude API is set for November 30, 2026, and Anthropic recommends moving to Claude Sonnet 5.5.
Admin API methods leave beta, and Claude Sonnet 4.5 users must migrate before its November 30 retirement on the Claude API.
Sonnet 5.5 costs $2/$10 per million tokens and has a 1M context window, but five changes can break code written for Claude Sonnet 5.
Refusals in the bio, frontier_llm, and reasoning_extraction categories are now billed even when they arrive before any output.
The same September 24 update also removes file and document names from Activity Feed entries and exits beta for Microsoft 365 session logging.
Cache diagnostics no longer needs a beta header; opting in just requires adding a diagnostics object to a Messages request.