Anthropic raised $65 billion at a near-trillion valuation, shipped a flagship model, and told customers the model it withheld for safety arrives in weeks.
Tuesday, May 26. Anthropic co-founder Chris Olah stood at a Vatican press conference for the launch of Pope Leo XIV's encyclical on artificial intelligence: “Magnifica humanitas: On safeguarding the human person in the time of artificial Intelligence.” Forty-two thousand words. Two hundred and forty-five paragraphs. (anthropic.com, National Catholic Reporter) Pope Leo XIV is the first pope to personally present an encyclical to the world. He chose AI. The Vatican did not invite the CEO. It invited the interpretability researcher.
Wednesday brought a new flagship model. Thursday brought an SDK update cascade. Friday brought a $65 billion Series H at a $965 billion valuation and a timeline on the model Anthropic had classified as too dangerous for general release.
The interpretability researcher was at the Vatican on Tuesday. By Friday, the company was worth nearly a trillion dollars and preparing to ship the model the Vatican's encyclical might have had something to say about. Whether those facts are in tension depends on who you ask. What is clear is that they belong to the same week.
Anthropic closed a $65 billion Series H on May 29, led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia. Post-money valuation: $965 billion, surpassing OpenAI as the most valuable private AI company. (anthropic.com/news/series-h) Run-rate revenue crossed $47 billion, per Fortune and Simon Willison. The revenue trajectory: $9B at year-end 2025, $14B at the Series G close in February, $30B in April, $47B now. Three times the revenue in three months. Claude Code is the stated driver.
In the same announcement window, Bloomberg reported that Anthropic told customers it has made “swift progress” developing safety safeguards for Mythos-class models and plans to release them broadly in “coming weeks.” Mythos has been restricted to Glasswing partners since April, on the grounds that it can identify and exploit vulnerabilities in every major OS and browser when directed. That restriction is close to lifting. No specific date.
These are not separate stories. A company raising capital at a $965 billion valuation is also a company telling its customers that the capability it once classified as too dangerous for general release is coming soon. The revenue validates the demand. The investment validates the trajectory. The safety work, Anthropic says, is done enough. Mythos is not a model with a launch date yet. It is a model with a direction of travel.
Anthropic released Opus 4.8 on May 28. Pricing unchanged: $5/$25 per MTok. Context window: 1M tokens on the Claude API, Amazon Bedrock, and Google Cloud Vertex AI (200k on Microsoft Foundry). Max output: 128k tokens. (anthropic.com/news/claude-opus-4-8)
The headline benchmarks are in math and honesty. USAMO 2026 moved from 69.3% to 96.7%. Opus 4.8 is the first Claude model to score 0% on uncritically accepting and reporting flawed results, and shows more than a tenfold reduction in overconfidence compared to 4.7. SWE-bench Pro: 64.3% to 69.2%. GraphWalks 1M-token F1: 40.3% to 68.1%. Fast mode repriced: 2x standard rate for 2.5x speed, down from 6x previously.
Claude Code v2.1.154 (May 28) introduced Dynamic Workflows as a research preview: describe a large task and Claude proposes splitting execution across tens to hundreds of background subagents running in parallel. (github.com/anthropics/claude-code/releases) Opus 4.8 defaults to high effort. Shell background sessions via ! <command>. Streaming tool execution universally enabled across Bedrock, Vertex, and Foundry. Worktree isolation guard added.
v2.1.152 earlier that week removed Auto mode's opt-in consent requirement and landed /code-review --fix, which applies review findings to the working tree directly. v2.1.157 on Friday added auto-loading plugins from .claude/skills and the EnterWorktree hook for mid-session worktree switching.
The Claude API now accepts role: "system" messages at any position in the messages array after the first user turn, on Opus 4.8. System prompts no longer need to be fixed at conversation start. Enables mid-session updates without restarting. stop_details field is now publicly documented. Python SDK v0.105.0 and TypeScript SDK v0.100.0 shipped support the same day. (anthropic.com/news/claude-opus-4-8)
Anthropic held its inaugural partner webinar on May 27: May product news, CCAF certification updates, and the launch of Anthropic Partner Academy with the Claude Partner Certification alongside it. (anthropic.com/webinars/the-first-broadcast) Global replay ran May 28 at 11:00 AM SGT.
No research paper published this week. No standalone Claude Apps update beyond the effort-control feature attached to the Opus 4.8 launch. The Python SDK was quiet from May 22 (v0.104.1) until Opus 4.8 triggered three releases in 24 hours on May 28. docs.anthropic.com returned HTTP 403 on all four sweeps this week, leaving API release notes, Claude Apps notes, and system prompt docs unverifiable from primary source. Claude Code v2.1.155 does not appear in this week's dailies.
The Vatican heard from the interpretability researcher on Tuesday.
The company cleared $965 billion by Friday.
The model they called too dangerous for general release is arriving in weeks.
A 1:1 mirror of every Anthropic release for the week of 2026-05-25.
| Date | Area | Title | One-liner |
|---|---|---|---|
| 2026-05-28 | MODEL | Claude Opus 4.8 | New flagship. $5/$25 per MTok unchanged. USAMO 69.3% to 96.7%. First Claude model to score 0% on uncritically reporting flawed results. Overconfidence more than 10x lower than 4.7. |
| 2026-05-28 | API | Mid-conversation system messages | role: "system" accepted after first user turn. Opus 4.8 only. stop_details now publicly documented. |
| 2026-05-27 | CODE | Claude Code v2.1.152 | /code-review --fix applies findings to working tree. /simplify aliases to it. Auto mode opt-in consent removed. MessageDisplay hook event. disallowed-tools frontmatter support. |
| 2026-05-28 | CODE | Claude Code v2.1.153 | Status line commands receive COLUMNS/LINES env vars. macOS background agents surface as “Claude Code” in Privacy and Security. /model saves as default. MCP reconnect-loop and OAuth bugfixes. |
| 2026-05-28 | CODE | Claude Code v2.1.154 | Dynamic Workflows (research preview): parallel subagent execution for large tasks. Opus 4.8 defaults to high effort. Lean system prompt now default for new models. Shell background sessions via ! <cmd>. |
| 2026-05-29 | CODE | Claude Code v2.1.156 | Hotfix: Opus 4.8 thinking blocks were being modified in Claude Code, causing API errors. Fixed. |
| 2026-05-29 | CODE | Claude Code v2.1.157 | Plugins in .claude/skills load automatically. claude plugin init <name> scaffolds plugins. EnterWorktree mid-session switch. WSL image/screenshot paste and Windows Explorer drag-and-drop. |
| 2026-05-28 | APPS | Effort control on claude.ai | Per-conversation effort level selector. Ships with Opus 4.8. High effort is the default for Opus 4.8. |
| 2026-05-26 | SDK-TS | TypeScript SDK v0.98.1 | Fixes directory prefix preservation in skills.versions.create uploads. Trusted Publishing. Renames private-sandbox-worker example to self-hosted-sandbox-worker. |
| 2026-05-27 | SDK-TS | TypeScript SDK v0.99.0 | Custom file size caps. Streaming fix: stop_details now carried through message_delta accumulation. |
| 2026-05-28 | SDK-PY | anthropic-sdk-python v0.105.0–2 | v0.105.0: claude-opus-4-8, mid-conversation system blocks, usage.output_tokens_details. v0.105.1: Trusted Publishing. v0.105.2: minor patch. Three releases in 24 hours. |
| 2026-05-28 | SDK-TS | @anthropic-ai/sdk v0.100.0–1 | v0.100.0: claude-opus-4-8, mid-conversation system blocks, output_tokens_details. v0.100.1: fixes streaming encrypted_content on beta compaction blocks. |
| 2026-05-25 | NEWS | Chris Olah at Vatican, “Magnifica humanitas” | Anthropic co-founder speaks at launch of Pope Leo XIV's 42,000-word, 245-paragraph AI encyclical. First pope to personally present an encyclical. Olah's remarks published on anthropic.com. |
| 2026-05-27 | NEWS | The First Broadcast | Anthropic's inaugural partner webinar. May product news, CCAF cert updates. Partner Academy and Claude Partner Certification launched. |
| 2026-05-28 | NEWS | Opus 4.8 GA for GitHub Copilot | Opus 4.8 available for all GitHub Copilot users in Chat, inline suggestions, and agent mode. Previously limited preview. |
| 2026-05-29 | NEWS | Series H: $65B at $965B valuation | Led by Altimeter, Dragoneer, Greenoaks, Sequoia. $47B run-rate revenue. Trajectory: $9B (Dec 2025), $14B (Feb), $30B (Apr), $47B (May). Surpasses OpenAI as most valuable private AI company. |
| 2026-05-29 | NEWS | Mythos-level models in “coming weeks” | Anthropic told customers: “swift progress” on safety safeguards, Mythos-class capabilities coming to all customers in “coming weeks.” Previously restricted to Glasswing partners. No specific date. |
Daily digest at 9 PM ET. Weekly magazine every Friday morning. Six labs, one feed. No spam, one-click unsubscribe.