Releasesby GlyphMid-conversation tool changes enter beta on the Claude API
Add or remove tools between turns while preserving the prompt cache on Claude Fable 5, Mythos 5, Opus 4.8 and Opus 5, using the mid-conversation-tool-changes-2026-07-01 beta header.
Also shipped Jul 24, 2026Anthropic in July 2026 →
Skill and plugin sharing in CoCo (Public Preview)
Skill and plugin sharing is now available in public preview. Share, discover, and govern CoCo skills and plugins across your Snowflake account. Publish them as Cortex Extension objects that teammates can find and install, with administrator certification, access control, and usage telemetry. For more information, see Share skills and plugins .
Governed tags system table is in Beta
The system.tags.governed_tags system table is in Beta . This table contains a record for each governed tag key in your account, including customer-created and system-generated tags. Use it to programmatically determine which tags in your account are governed. See Governed tags system table reference .
Git CLI commands in Git folders are now in Public Preview
Git CLI commands in Git folders are now in Public Preview . Git folders that you create from the UI automatically get Git CLI access when your workspace is eligible, so you can run standard Git commands such as git stash , interactive rebase, pre-commit hooks, Git submodules, and Large File Storage (LFS) directly from a Databricks terminal.
Diagnose and fix Lakebase issues with Genie, using telemetry in Unity Catalog (Beta)
Lakebase adds AI-assisted troubleshooting: database issues and early warnings are surfaced to you with Insights , and you can then investigate and fix them conversationally with Genie . Both draw on database telemetry captured to Unity Catalog. See Observability .
Knowledge Catalog governance workflows for data product access in preview
Governance workflows add automated controls for data product access management through a request-review mechanism.
Vercel WAF for Blob is now in beta
The Vercel WAF can now protect a Vercel Blob store. The same rules that guard your deployments (deny, challenge, rate limit) now apply to blob traffic with no changes to your code, blob URLs, or @vercel/blob . Every blob is already served through Vercel's CDN , so protection is a switch on the store, not a new proxy.
More from Claude PlatformEvery Claude Platform release →
Claude Opus 5 launches with a 1M token context window at $5 / $25 per MTok
Claude Opus 5 improves on Opus 4.8 with a 1M token context window, 128k max output and thinking on by default, at the same $5 / $25 per MTok price. It is available on the Claude API, Bedrock, Google Cloud and Microsoft Foundry.
Claude Opus 5 only allows disabling thinking at effort high or below
Disabling thinking with effort xhigh or max returns a 400 error on Claude Opus 5, a breaking change from Opus 4.8 that integrations should check before migrating.
Effort is the primary control for Claude Opus 5
Claude Opus 5 supports the full effort ladder of low, medium, high, xhigh and max, with max intended for capability-critical work.
Claude API fallbacks parameter adds a default mode for recommended fallbacks
The default mode applies Anthropic's recommended fallback models by refusal category. Server-side fallback is in beta and this mode requires the server-side-fallback-2026-07-01 header.
Fast mode removed for Claude Opus 4.7
Requests to claude-opus-4-7 with speed fast now return an error rather than falling back to standard speed. The model stays available at standard speed; migrate to Opus 5 or Opus 4.8 to keep fast mode.
Claude Managed Agents agents accept an effort level in model configuration
Pass effort inside the model object when creating a Claude Managed Agents agent to set its effort level.
Claude Managed Agents webhooks cover environment and memory store lifecycle
Four environment.* and three memory_store.* webhook event types let you react to environment and memory store lifecycle changes without polling.
Claude Managed Agents sessions can be seeded with initial events
POST /v1/sessions accepts initial_events with up to 50 user.message and user.define_outcome events, and a non-empty list starts the agent loop in the same call.