Claude Report — 2026-07-01

🚀 Anthropic Official

Introducing Claude Sonnet 5 (Jun 30)

Launches as the new default model for Free and Pro plans (also available on Max/Team/Enterprise) with a 1M-token context window, 128k max output tokens, and introductory pricing of $2/$10 per MTok input/output through Aug 31, 2026 (then $3/$15). Ships three breaking changes vs. Sonnet 4.6: adaptive thinking on by default, manual extended-thinking parameters now return a 400 error, and non-default sampling params (temperature/top_p/top_k) are rejected; the new tokenizer also produces ~30% more tokens for the same text.

Claude Science, an AI workbench for scientists (Jun 30)

Ships a beta macOS/Linux app for Pro/Max/Team/Enterprise integrating 60+ scientific databases and skills (genomics, proteomics, structural biology, cheminformatics), 3D protein/genome-track rendering, a built-in reviewer agent for citation/calculation validation, and full code provenance for figures; funds up to 50 "AI for Science" projects with up to $30,000 in credits each (applications due Jul 15).

Redeploying Fable 5 (Jun 30, effective Jul 1)

Lifts the Jun 12 export-control suspension on Claude Fable 5, restoring global availability from Jul 1 after deploying a cybersecurity classifier that blocks the Amazon-reported jailbreak technique in over 99% of cases; jointly proposes an industry-wide jailbreak severity scoring framework with Amazon, Microsoft, and Google across four axes (capability gain, breadth, weaponization ease, discoverability).

Claude Code v2.1.197 (Jun 30)

Makes Claude Sonnet 5 the default model, wiring in the 1M-token context window and introductory $2/$10 per MTok pricing.

anthropic-sdk-python v0.114.0 and v0.115.0 (Jun 30)

v0.114.0 adds claude-sonnet-5 support and fixes an agent_toolset absolute-path bug; v0.115.0 adds Managed Agents event-delta streaming, per-session agent configuration overrides, reverse pagination for session listing, vault credential injection-location scoping, and new agent/deployment webhook events.

anthropic-sdk-typescript v0.108.0 and v0.109.0 (Jun 30)

Mirrors the Python SDK: v0.108.0 adds claude-sonnet-5 support, v0.109.0 adds the same Managed Agents features (event deltas, agent overrides, reverse pagination, vault scoping, webhooks); companion releases vertex-sdk v0.19.0, bedrock-sdk v0.32.0, and aws-sdk v0.6.0 ship the same day with auth/logger fixes.

API changelog — Managed Agents platform updates (Jun 30)

Adds event_deltas[] on the session event-stream endpoint for previewing agent text before completion, backward pagination (prev_page) for GET /v1/sessions, per-session agent configuration overrides (agent_with_overrides), an injection_location setting for vault credential scoping, and webhook coverage for agent/deployment/deployment-run lifecycle events.

🔌 Claude Code Plugins

Plugin tracking skipped — ranking cache unavailable this cycle.

🛠️ Skills

alirezarezvani/claude-skills

Merges a four-PR chain (#872–#875) adding a new "deep-research" skill with a multi-phase triangulation methodology, plus automated changelog and marketplace-counter maintenance, landing Jul 1.

Piebald-AI/claude-code-system-prompts

Ships v2.1.197 (+21,695 tokens) on Jun 30 and v2.1.196 (+1,869 tokens) on Jun 29, tracking the growth of Claude Code's captured internal system prompt release-by-release.

Unchanged in window: anthropics/skills, glebis/claude-skills, ComposioHQ/awesome-claude-skills

🤖 Agents & Subagents

nicobailon/pi-subagents

Ships v0.31.1 (Jun 30) adding /chain inline parallel groups and subagent profile commands, plus fixes for pi-intercom discovery and async interrupt portability.

Chachamaru127/claude-code-harness

Ships v4.16.4 (Jun 28) adding owner-scoped control for the egress category in the runtime action hard floor, tightening permission defaults.

anthropics/claude-agent-sdk-typescript

Ships v0.3.197 (Jun 30), a parity bump matching Claude Code v2.1.197's Sonnet 5 default-model support.

affaan-m/everything-claude-code

New addition to the tracked top 5 — a fast-growing Claude Code skills/subagents/hooks collection that shipped catalog-sync fixes, new skills (loop-design-check, growth-log), and a stop-hook verifying session-end thinking quality on Jun 29–30.

anthropics/claude-plugins-official

New addition to the tracked top 5 — per-plugin agents/ folders saw a security-policy tightening around cross-service credential routing (Jul 1) and new plugin additions including Grafana, Hostinger, Zyte, and Mergify (Jun 30).

🔗 MCPs & Integrations

mcp-python-sdk (modelcontextprotocol/python-sdk)

Ships v2.0.0b1 (Jun 30), the first v2 pre-release, adding support for the 2026-07-28 protocol revision, a pluggable event bus for subscriptions/listen, server-side Mcp-Param-* header validation, and a client-side response cache honoring caching hints.

modelcontextprotocol/typescript-sdk

Ships v2.0.0-beta.1 (Jun 30) alongside companion server and server-legacy packages, entering beta with a rebuilt VitePress docs site, a new toWebRequest() Node adapter export, and an overhauled codemod migration path.

awslabs/mcp

Ships valkey-mcp-server V2 (Jun 29) with a GLIDE client migration plus Focused Search, JSON, and Command Runner tools, alongside same-window security fixes for read-only-bypass issues in the AWS Data Processing and MySQL MCP servers.

portainer/portainer-mcp

Releases v2.43.0 (Jun 29) fixing edge-status guidance and an over-reporting bug in redaction counts.

Unchanged in window: github-mcp-server

💡 Community — Workflows & Ideas

Claude Code caught fingerprinting requests via hidden Unicode markers (Jun 30)

Reverse-engineering of the Claude Code binary finds it silently alters punctuation in the system prompt's date string based on ANTHROPIC_BASE_URL and timezone, apparently to flag proxy resellers and distillation pipelines tied to Chinese AI labs; Anthropic acknowledges the code, sparking a 2,000+ point HN thread.

Claude Code silently deletes transcripts older than 30 days (resurfaced Jun 30)

cleanupPeriodDays defaults to 30 and wipes ~/.claude/projects/*/.jsonl history at startup with no warning; a long-open GitHub issue resurfaces on HN, with "cleanupPeriodDays": 3650 shared as the workaround.

"Claude Code Quietly Got 5x More Expensive" (Jun 30)

Week-over-week token accounting shows quota draining far faster than visible usage logs justify — a new account hits 35% quota depletion in 5 hours with no matching log entries — prompting a call for a transparent per-request billing ledger.

AgentWire (Jun 30)

Self-hosted web portal (v1.51.0, 775 commits) orchestrates many Claude Code sessions over tmux as draggable windows, adding safety hooks against destructive commands, git-worktree parallel branches, a "council" multi-agent deliberation pattern, and local voice control.

Caliper (HN Jun 28)

Local-first eval harness runs a Claude Code/Codex/Pi skill k times in isolated environments to score pass@k against a no-skill baseline, filling the regression-testing gap most published skills ship without.

📰 Quick Mentions

• autoharness (Jun 30) — distills skills from real session transcripts and auto-prunes ones with low re-invocation rates.
• mux (Jun 28) — minimal bash+fzf tmux overlay for jumping between Claude Code sessions waiting on input.
• statuslin.es (Jun 30) — community gallery of custom Claude Code statusline configs showing cost, context usage, and rate limits.
• Open Memory Protocol (Jun 30) — open spec and self-hosted server for a shared memory store across Claude Desktop, Claude.ai, and Cursor/Copilot.
• VPSMaxxing (Jun 30) — guide for moving Claude Code/Codex sessions to an always-on VPS for background work.