June was the month Anthropic learned that launching the best model in the field does not protect you from the government. Fable 5 shipped June 9 with the highest published score on SWE-Bench Pro, eleven points clear of anything else in the market. Four days later, Commerce Secretary Howard Lutnick's directive arrived: suspend all access for any foreign national, anywhere on Earth, including Anthropic's own employees. Anthropic took both Fable 5 and Mythos 5 offline for every customer worldwide.
The suspension held through the month. The rest of the product line kept running. On June 30, Sonnet 5 became the default model for every Free and Pro user on the platform, no ceremony, just a routing change. The model that replaced the flagship as the everyday default is agentic, carries a one-million-token context window, and costs $2 per million input tokens at introductory pricing.
In between: a confidential S-1 on file with the SEC, a Nobel laureate crossing the street from DeepMind, 28 Claude Code releases, a Slack integration, a formal accusation against Alibaba, and a 162-minute outage that took every Claude product down simultaneously. Not a slow month.
Fable 5 posted 80.3% on SWE-Bench Pro, eleven points clear of the nearest competitor. Spatial reasoning nearly tripled versus Opus 4.8. In a Stripe evaluation, Fable 5 compressed two-plus months of engineering work on a 50-million-line Ruby migration into a single day. Model ID: claude-fable-5. Pricing: $10 per million input tokens, $50 per million output. Free on paid plans through June 22.
The companion model, Mythos 5, is the same architecture with some safety filters lifted, deployed through Project Glasswing for authorized cyberdefenders and infrastructure providers. The public gets Fable. The cyberdefense community gets Mythos.
Four days later, at 5:21 PM ET on June 12, Commerce Secretary Howard Lutnick's letter reached Dario Amodei. The order: suspend all access to Fable 5 and Mythos 5 for any foreign national, inside or outside the United States, including Anthropic's own employees. Anthropic took both models offline for every customer worldwide. The trigger, per Axios: a claimed jailbreak of Mythos by another company, raising concerns about its cybersecurity capabilities being misused. Anthropic disputed the characterization and said it was working to restore access.
Seventy-two hours from general availability to global suspension. No advance notice to Anthropic. A restriction scope broad enough to cover the company's own international staff. At press time, both models remain offline for foreign nationals. All other Claude models (Sonnet, Haiku, Opus 4.7, Opus 4.8) remained unaffected throughout.
Claude Sonnet 5 launched June 30 and became the default model for every Claude Free and Pro user the same day. Not a preview, not a limited rollout, not a phased release. Default, as of today, for everyone.
The architecture targets agentic work. It plans, runs tools (browsers, terminals, code), and operates autonomously at a level that until recently required Opus-class compute and Opus-class pricing. On agentic benchmarks, Sonnet 5 runs close to Opus 4.8. The context window is one million tokens, native. Introductory pricing through August 31: $2 per million input tokens, $10 per million output. Standard pricing after that: $3/$15.
The same day: system card published, Claude Science beta launched for life-sciences researchers, and Claude went generally available in Microsoft Foundry on Azure, running on NVIDIA GB300 Blackwell Ultra GPUs. With AWS Bedrock and Google Cloud already covered, Anthropic now has enterprise-grade GA deployment across all three major cloud hyperscalers.
claude-sonnet-5)Monday, June 1. Anthropic submitted a confidential draft registration statement on Form S-1 to the U.S. Securities and Exchange Commission. The filing initiates the formal IPO process. Post-money valuation at filing: approximately $965 billion, following the $65 billion Series H closed the week before. Annualized revenue run rate, per CNBC: $47 billion. Target listing window: as early as October 2026. Wilson Sonsini, which handled Google's 2004 public debut, is reportedly engaged as legal counsel.
The filing is confidential. Revenue detail, margin structure, and risk factors stay private until Anthropic elects to disclose them. What is public today: the direction, the scale, and the timing. Reported by CNBC, NPR, and Yahoo Finance.
The day after the filing, Anthropic expanded Project Glasswing to 150 new critical-infrastructure organizations across 15-plus countries, bringing total partners to approximately 200. The IPO and the classified cybersecurity program are not unrelated: the pitch to the public market requires a story about defensible moats, and government relationships are part of that story.
The cadence is the story: one new build every 1.07 days. What changed in those builds covers a range from single-line UX wording fixes (v2.1.185: "Waiting for API response" instead of "No response from API") to architectural expansions that didn't exist six months ago.
The headline architectural change was v2.1.172 (June 10): sub-agents can now spawn their own sub-agents, up to five levels deep. Combine that with the fallback model config from v2.1.166 (up to three backup models tried in order on primary-model overload) and what you have is a system that can route, retry, and recurse without human intervention at each decision point. The v2.1.179 permission system added Tool(param:value) syntax with wildcard support, giving organizations granular control over what agents are allowed to touch.
The largest single release was v2.1.183 (June 17): /config key=value for runtime configuration, streaming of long paragraphs line-by-line, auto-retry on API drops during thinking, and over thirty UI and window-management fixes, including a 120ms-per-launch startup regression and a 15-second startup block on degraded networks. That one release would be a meaningful quarterly update at most software companies.
The Claude Apps Gateway landed June 29 as a self-hosted control plane for enterprise Claude Code deployments on Bedrock and Google Cloud: a single stateless container on Linux, backed by PostgreSQL, with corporate SSO, centrally enforced policy, per-user cost attribution, and daily/weekly/monthly spend limits. With Microsoft Foundry going GA the same week, Anthropic now has enterprise-grade deployment infrastructure across all three major cloud hyperscalers.
John Jumper spent nine years at Google DeepMind. He shared the 2024 Nobel Prize in Chemistry for AlphaFold, the protein-structure prediction model that cracked the protein-folding problem and became the foundational tool for structural biology. He was the youngest chemistry laureate in more than 70 years. Bloomberg reported his departure on June 19. TechCrunch confirmed on June 20. His role at Anthropic has not been formally announced.
The hire connects to a declared strategy. Life sciences is Anthropic's stated second vertical after developer tooling, and a Nobel in hard science signals something that benchmark scores cannot: credibility inside research institutions that distrust AI labs on principle. Claude Science, the beta product for scientific computing launched June 30, connects to over 60 curated data sources across genomics, proteomics, structural biology, and cheminformatics. Anthropic is funding up to 50 research projects at up to $30,000 in credits each, with applications open through July 15.
The same week, Anthropic sent a formal letter to US senators and White House officials accusing Alibaba's Qwen AI lab of running a coordinated campaign to extract Claude's capabilities in software engineering and agentic reasoning via thousands of fraudulent accounts. Alibaba ADRs fell to a 16-month low in Hong Kong following the disclosure. Anthropic escalated through Congress rather than courts, which suggests the complaint is at least partly political. The capabilities named as targets (software engineering, agentic reasoning) are exactly where Claude Code and Sonnet 5 are most directly competitive with Qwen's published benchmarks.
The suspension is unresolved. Fable 5 and Mythos 5 remain offline for all foreign nationals. Anthropic published a statement calling the Commerce Department directive "extraordinarily broad" and says it is working to restore access. The directive is the broadest application of US export control to a commercial AI product on record. No resolution timeline has been given.
The Mythos public promise is on hold. The June 2 Glasswing announcement confirmed that Mythos-class models would reach all customers "in the coming weeks." That announcement predates the suspension by eight days. At press time, there is no floor on the public availability timeline.
Opus 4.7 fast mode removal: July 24. After that date, requests to claude-opus-4-7 with speed: "fast" will return errors. Migrate to claude-opus-4-8 fast mode, which is approximately three times cheaper per request. Opus 4.6 fast mode was already removed June 30; those requests now run at standard speed without error.
Opus 4.1 retires August 5. Formally deprecated June 5. Thirty days remain at press time. Any code referencing claude-opus-4-1 as the model ID will return errors after August 5.
Every confirmed Anthropic release in June 2026. Sourced from the Shipped. daily and weekly digests (June 1 to June 30). Primary sources only. 9 sections.
Two launches, two retirements, one new default. Fable 5 was the best model available on June 9; it was suspended on June 12. Sonnet 5 became the default on June 30.
claude-fable-5. 80.3% SWE-Bench Pro, 11 pts above field. Spatial reasoning nearly 3x vs. Opus 4.8. Safety routing sends ~1 in 20 sessions to Opus 4.8 for high-risk queries (cybersecurity, biology). Pricing: $10/$50 per million tokens. Free on paid plans through June 22, then usage credits.claude-sonnet-4-20250514 now return errors. Migrate to claude-sonnet-4-6.claude-opus-4-20250514 now return errors. Migrate to claude-opus-4-8.claude-sonnet-5. Immediately became the default for all Claude Free and Pro users. 1M-token native context. Agentic performance close to Opus 4.8 on agentic benchmarks. Introductory pricing: $2/$10 per million tokens through August 31, 2026, then $3/$15. System card published the same day.claude-sonnet-5. Update Claude Code to v2.1.197 to get Sonnet 5 as the in-app default. Pricing: $2/$10 per MTok through Aug 31, then $3/$15.Billing restructured, fast-mode deprecated, and a 162-minute outage resolved with no public postmortem.
claude -p usage now draw from a dedicated monthly credit, metered at standard API list prices. No longer bundled with subscription rate-limit pools. Interactive Claude Code terminal sessions and claude.ai chat are unaffected. Pro: $20/mo. Max 5x: $100/mo. Max 20x: $200/mo. Credits reset each billing cycle.claude-opus-4-6 with speed: "fast" now run at standard speed and are billed at standard rates. No error is returned. Migrate fast-mode Opus 4.6 workflows to claude-opus-4-8.claude-opus-4-7 is deprecated effective June 30. After July 24, 2026, requests with speed: "fast" to Opus 4.7 will return an error. Fast mode on Opus 4.8 is approximately three times cheaper per request than prior Opus models.claude-opus-4-7 fast-mode workflows to claude-opus-4-8 before July 24, 2026.28 releases from v2.1.160 to v2.1.197. Selected highlights below. Full changelog at github.com/anthropics/claude-code.
fallbackModel configuration for up to three alternative models tried in order when primary model is overloaded. Glob pattern support in deny rule tool positions. Cross-session message relay hardened: relayed messages no longer carry user authority.Tool(param:value) syntax for permission rules to match tool input parameters with wildcard support (e.g., Agent(model:opus) to block Opus subagents). Nested .claude/skills directory discovery. /doctor to verify setup status. Bug fixes for subagent transcript viewing, authentication/session management, MCP server-level specs in subagent disallowedTools, vim mode undo, and VSCode CJK IME candidate window./config key=value sets any setting from the prompt in interactive, -p, and Remote Control modes. sandbox.allowAppleEvents opt-in for macOS. CLAUDE_CLIENT_PRESENCE_FILE suppresses mobile push notifications. Streaming of long paragraphs now line-by-line. API connection drops mid-thinking now auto-retry. Subagent panel auto-hides idle agents. Bun runtime upgraded to 1.4. 30+ fixes including a 120ms startup regression and a 15-second startup block on degraded networks.claude mcp login <name> and claude mcp logout <name>: MCP server authentication on the command line without opening the interactive menu. --no-browser flag for SSH sessions. iTerm2 teammate mode added. AWS credential refresh surfaces in /login. Agent gets a compaction reminder when MEMORY.md approaches its size limit./rewind restores conversation context from before /clear was run. Streaming CPU usage drops ~37% through text update coalescing at 100ms intervals. v2.1.191: Twelve bug fixes and two performance improvements. Superseded by v2.1.191 the same afternoon.autoMode.classifyAllShell setting routes all Bash and PowerShell commands through the classifier, not just arbitrary-code-execution patterns. New claude_code.assistant_response OTEL event. Three backgrounding bugs fixed that had been leaving phantom agents in sessions after completion.CLAUDE_CODE_DISABLE_MOUSE_CLICKS=1 disables click-and-drag in fullscreen while preserving scroll.claude update or reinstall.Claude Tag in Slack, Claude Science for researchers, and GA in Microsoft Foundry completing the three-hyperscaler footprint.
Moved from v0.105.2 (May 29) to v0.115.0 (June 30). Key additions: Opus 4.1 deprecated, retired model IDs removed, system.message streaming, and Managed Agents event delta streaming.
$defs preserved when schema root is a $ref.claude-sonnet-4-20250514 and claude-opus-4-20250514 from the SDK's type definitions, coordinated with today's model retirements.system.message streaming events: clients can receive streamed system-level events mid-conversation. New refusal category in the stop reason taxonomy. User Profile ID support in request headers. Fix: x-stainless-helper header now uses a single source of truth.20260318 web fetch and support tools. Async count_tokens bug fix. Token counting accepts user profile IDs.claude-sonnet-5. Agent toolset allows absolute paths that resolve inside workdir. v0.115.0: Managed Agents event delta streaming, agent overrides, reverse pagination, vault credential injection scoping, agent and deployment webhook events.Client middleware landed across all platforms in early June. The pattern for June: parity with Python SDK releases, then a burst of platform-specific updates on June 30.
stop_details now carries through beta message_delta accumulation. JSON number parsing fixed for scientific notation. Vertex, Foundry, Bedrock, and AWS platform SDKs all ship the same pair of changes: Opus 4.1 marked deprecated, client middleware added.claude-sonnet-4-20250514 and claude-opus-4-20250514 from TypeScript SDK type definitions, coordinated with today's model retirements.system.message streaming events, new refusal category, and User Profile ID in request headers. Fixes x-stainless-helper header. Parallel release to Python SDK v0.112.0.20260318 web fetch and support tools. Symlink hops bounded. Glob confinement check before match resolution. BatchCreateParams.Request.params restored. Token counting accepts user profile IDs.claude-sonnet-5. Agent-toolset absolute path fix. v0.109.0: Managed Agents event delta streaming, agent overrides, reverse pagination, vault credential injection scoping, agent and deployment webhook events.Two public data releases and one policy analysis. Anthropic is now publishing behavioral research alongside product launches.
The IPO filing, Glasswing expansion, a Nobel laureate hire, the government directive, an accusation, and a new partner network.
Two model IDs retired in June; two more deprecations active for July. Check your production model strings now.
claude-opus-4-8 before August 5; requests to Opus 4.1 after that date will return errors.claude-sonnet-4-20250514 and claude-opus-4-20250514 reached end of life June 15 (deprecated April 14 with 60-day notice). Requests to both model IDs now return errors. Migrate to claude-sonnet-4-6 and claude-opus-4-8 respectively.claude-opus-4-7 is deprecated effective June 30. After July 24, 2026, requests with speed: "fast" to Opus 4.7 will return an error, not silently degrade. Opus 4.6 fast mode was already removed June 30; those requests now run at standard speed. Fast mode on Opus 4.8 is approximately three times cheaper than on prior Opus models.claude-opus-4-8 with speed: "fast".Daily digest at 9 PM ET. Weekly magazine every Friday morning. Six labs, one feed. No spam, one-click unsubscribe.