Changelog — June 30, 2026

Claude Sonnet 5 is now the default model in Claude Code. Two releases landed:

• v2.1.197 — Claude Sonnet 5 (release): Sonnet 5 becomes the default model, with a native 1M-token context window and promotional pricing of $2/$10 per Mtok through August 31. Update to v2.1.197 for access. If you pin a specific model in your config, you'll keep it — but new sessions default to Sonnet 5.
• v2.1.196 — org defaults + MCP security fix (release): admins can now set an organization default model in the org console (surfaced as "Org default" or "Role default" in /model when you haven't chosen one). Security: claude mcp list / get no longer spawn .mcp.json servers that a repo self-approved via a committed .claude/settings.json — untrusted workspaces now show ⏸ Pending approval. Also: readable default session names, Cmd/Ctrl-click file attachments that reveal in Finder/Explorer, a fix so waking a background job no longer deletes its conversation and re-runs the original prompt, and a rate-limit warning/telemetry over-counting fix.

All Claude Code releases → Affects: /openclaw/, /openclaw/configuration/, /openclaw/security/, /openclaw/cost-optimisation/, /claude-cowork/

The Claude apps and Claude API release notes both refreshed on June 30, coinciding with the Claude Sonnet 5 launch that also reached Claude Code. Because these are hosted release-notes pages (change-detected, not parsed feed entries), we flag the update rather than reproduce each line — see the Claude apps notes and Claude API notes for specifics, and the Claude Code entry above for the model change most Cowork users will feel.

Claude apps release notes → Affects: /claude-cowork/, /claude-cowork/pricing/, /claude-cowork/vs-api/

A pre-release focused on local model serving and Agent Manager model handling:

• Local Anaconda Desktop model server (#11714): connect to a local Anaconda Desktop text-generation model server from the CLI or VS Code.
• Parakeet speech-to-text (#11693): support for NVIDIA Parakeet TDT 0.6B v3 as a speech-to-text input.
• Agent Manager model discovery (#11786): Agent Manager chat tools can discover available models and reasoning variants by model name, then start each session with the chosen model and reasoning effort — the provider for a named model is resolved automatically, preferring the provider behind your current default and falling back to the Kilo Gateway.
• Per-worktree sandbox toggle (#11689): a sandbox toggle in the Agent Manager "New Worktree" modal so each worktree session can start sandboxed.

Release notes → Affects: /kilocode/, /kilocode/models/, /kilocode/orchestrator/, /kilocode/security/

NemoClaw shipped release docs for both v0.0.70 and v0.0.71 today (#6067, #6070), catching the public docs up on the last several days of work. The reader-facing changes:

• Host-mediated gateway restart (#5874): a new supported command to restart NemoClaw-managed OpenClaw and Hermes gateways, with automatic recovery routed through the same authenticated controller/supervisor path.
• uninstall --destroy-user-data (#5784): nemoclaw uninstall --yes is non-destructive by design (it preserves ~/.nemoclaw/rebuild-backups/, backups/, and sandboxes.json). A new --destroy-user-data flag purges that preserved state for a full clean uninstall.
• policy-list preset provenance + Balanced-tier change (#5797): policy-list now annotates each active preset with its inferred source, and the default Balanced tier no longer grants weather egress — a default Balanced OpenClaw onboarding now applies a tighter preset set. Relatedly, --policy-tier=restricted no longer applies openclaw-pricing (#5798).
• Shields auto-relock warning (#5951): after shields auto-restore, a subsequent host nemoclaw agent call could fail with OpenClaw's cryptic missing scope: operator.write; NemoClaw now emits actionable recovery context instead.
• Faster warm rebuilds (#6034): Docker layer reuse is preserved for identical warm sandbox builds by keeping NEMOCLAW_BUILD_ID stable in the stock OpenClaw/Hermes Dockerfiles.
• OpenShell support bumped 0.0.44 → 0.0.71 (#5596): checksum-pinned install, authenticated Docker-driver gateway startup, and fail-closed recovery.
• Docs additions: clarified local compatible endpoints (serve raw GGUF through an OpenAI-compatible server like llama.cpp; the Ollama option expects model tags, not .gguf paths — #5505), a new inference model task-fit guide (#5527), a model capability audit matrix (#5528), and an explicit nemoclaw onboard quickstart step for hosts where the installer doesn't auto-launch the wizard (#6001).

Commit log → Affects: /nemoclaw/, /nemoclaw/setup/, /nemoclaw/policy/, /nemoclaw/switching-providers/