feature
Messages API supports on-demand conversation compaction
docs.claude.comagent
The Messages API can now compact conversations on demand in beta, returning a signed compaction block that summarizes prior messages. Developers can keep recent turns verbatim after the summary and avoid invalidating prompt caches while managing context windows.
Stacks this affects
Claude ships in 34 tested stacks here.
Full-stack with Lovable
Code · Builder
Cloud IDE with Replit
Code · Builder
UI-first with v0
Code · Builder
AI newsletter
Writing · Newsletter
AI newsletter (Substack)
Writing · Newsletter
AI thumbnails + ad creative
Marketing · Visual
Android app with AI
Code · Mobile
Autonomous research agent
Research · Intelligence
Cheap bulk LLM automation
Code · Backend
Cold outbound
Sales · Outbound
Cold outbound (solo)
Sales · Outbound
iOS + Android with Expo
Code · Mobile
Faceless YouTube
Video · Creator
Faceless YouTube (budget)
Video · Creator
Inbox triage
Productivity · Inbox
Inbox triage (free)
Productivity · Inbox
Indie SaaS
Code · Builder
Indie SaaS (Gemini long-context)
Code · Builder
iOS app with AI
Code · Mobile
LinkedIn ghostwriting
Writing · Agency
LinkedIn (for yourself)
Writing · Personal
Live trend research
Research · Intelligence
Multilingual content factory
Writing · Content
Open-weight LLM stack
Code · Backend
SEO content engine
Writing · SEO
Antigravity (Google)
Code · Workflow
Codex (CLI + cloud)
Code · Workflow
Ship code with AI
Code · Workflow
Cross-platform short-form
Marketing · Distribution
Self-hosted scheduler
Marketing · Distribution
TikTok character
Video · Creator
TikTok character (audio + text)
Video · Creator
UGC ad creative
Marketing · Ads
UGC ads (CapCut-only)
Marketing · Ads
More Claude
featureCache diagnostics graduated from beta on Claude APIreleaseClaude Opus 5.5 launched with 1M context and always-on adaptive thinkingpricingClaude drops from $100 to $20 per monthfeatureClaude Managed Agents permission policies now auto-evaluate tool callsreleaseClaude Fable 5.1 launched with preserved thinking blocks and cheaper cache reads