Teams building on Claude are actually tracking three separate streams of change, and conflating them is the root of most surprises. The SDK libraries evolve on GitHub. The API and its features evolve in Anthropic's platform documentation. And the models have their own lifecycle of releases, deprecations, and retirements. Each stream has its own home and its own subscription mechanism.
Stream 1: SDK releases on GitHub
All official Anthropic SDKs live under the anthropics organization on GitHub — Python (anthropic-sdk-python), TypeScript (anthropic-sdk-typescript), Java, Go, Ruby, C#, and PHP, plus the ant command-line tool, whose release tarballs are published at github.com/anthropics/anthropic-cli/releases. Each repository's Releases page is the canonical changelog: per-version notes with the tag, date, and change list.
To subscribe without drowning in noise, use GitHub's Watch → Custom → Releases option on just the repositories for languages you ship. That delivers release notifications only — no issues, no PR chatter. GitHub also exposes each repository's releases as an Atom feed, which is handy for piping into a team Slack channel or an internal dashboard. And if you've set up Dependabot or Renovate per the version-pinning article, every new release additionally arrives as a pull request with the notes attached — the changelog delivered to the exact place you'll act on it.
Stream 2: API and feature changes
New API capabilities usually appear in the SDKs as ordinary minor releases ("add support for X"), but the substance — what the feature does, which platforms have it, what beta headers it needs — lives on platform.claude.com. This stream matters doubly for third-party platform users because availability differs by surface: historically, features like the Message Batches API and Files API reached the first-party API and Claude Platform on AWS without being available on Bedrock or Vertex AI. An SDK release note saying "added Files API helpers" does not mean your platform serves that endpoint — check the feature-availability documentation (or our feature matrix) before adopting.
Stream 3: model lifecycle
Model deprecations arrive on their own calendar and break running systems just as thoroughly as an SDK change. Concrete example from Anthropic's current models documentation: claude-opus-4-1-20250805 is deprecated and will be retired on August 5, 2026, with Claude Opus 4.8 as the recommended migration target; retired model IDs simply return errors. The models overview and the official model migration guide on platform.claude.com are the pages to re-check on a cadence — quarterly is a reasonable floor. No SDK update will save a request that names a retired model.
Reading a changelog effectively
When a release lands, triage in this order rather than reading top to bottom:
First, the version number itself. Under semantic versioning, a major bump is the "stop and read everything" signal — see migrating between major SDK versions. A pre-1.0 package (the PHP SDK's README, for instance, shows a ^0.36.0 constraint) deserves major-bump caution even on minor releases.
Second, anything labeled breaking, removed, or deprecated. Deprecations are tomorrow's breaks with advance notice; grep your codebase for the named symbol while the notice is fresh and file the ticket now.
Third, changed defaults and dropped runtime support. The quiet items — a raised minimum language version, a changed default timeout or retry count — are the ones that pass code review and then behave differently under load. The current READMEs set the floor at Python 3.9+, Node.js 20 LTS+, Go 1.24+, Ruby 3.2.0+, PHP 8.1.0+, and Java 8+; a release that raises one of these breaks builds on older toolchains without touching a single line of your code.
Last, the features. Genuinely fun, rarely urgent.
Where to go next
Turn awareness into safe action with version pinning and automated updates, and keep the major-version migration checklist ready for the big ones.