Presented by
Good morning, AI enthusiasts. Anthropic just shipped Claude Opus 4.8 and raised $65B in the same week, the kind of one-two punch that only happens when a lab is convinced it's pulling ahead.
The model itself is a modest step up on paper, but the interesting part is what it now admits it doesn't know. Has "honesty" quietly become the benchmark that matters?
In today's recap:
Anthropic ships Opus 4.8 with sharper judgment
Anthropic raises $65B at a $965B valuation
Configure Claude Code's undocumented hooks and skills
What 370,000 essays reveal about AI and creativity
4 new AI tools, prompts, and more
ANTHROPIC
Anthropic ships Opus 4.8 with sharper judgment
Recaply: Anthropic just shipped Claude Opus 4.8, an upgrade to its flagship that improves coding, agentic work, and reasoning at the same price as Opus 4.7, while adding new effort controls and a dynamic-workflows mode for Claude Code.
Key details:
Opus 4.8 can plan a job and run hundreds of parallel subagents in a single Claude Code session, verifying its own outputs before reporting back, with Anthropic pointing to codebase-scale migrations across hundreds of thousands of lines.
The model is around four times less likely than Opus 4.7 to let flaws in its own code pass unremarked, and its fast mode runs at 2.5 times the speed while costing three times less than before.
Databricks said Opus 4.8 reasons over PDFs and diagrams at 61% cheaper token cost than Opus 4.7, while VentureBeat noted it increasingly reasons about how its outputs will be graded, even when it isn't told it's being evaluated.
Opus 4.8 is available everywhere now at $5 per million input tokens and $25 per million output, with developers calling claude-opus-4-8 through the Claude API.
Why it matters: Anthropic is framing 4.8 as a modest bump, but the honesty gains are the real story. Models that flag their own uncertainty are far more useful for the long-running, autonomous work Claude Code is now pushing into. There's a twist worth watching. The same week, an independent benchmark caught earlier Claude models gaming an eval, so "does it admit what it doesn't know" is fast becoming the line between trustworthy agents and confident-but-wrong ones.
PRESENTED BY NEO
Fast browsing. Faster thinking.
Your browser gets you to a page. Norton Neo gets you to the answer. The first safe AI-native browser built by Norton moves with you from idea to action without slowing you down. Magic Box understands your intent before you finish typing. AI that works inside your flow, not beside it. No prompting. No copy-pasting. No switching apps.
Built-in AI, instantly and for free. Privacy handled by Norton. Built-in VPN and ad blocking protect you by default. No configuration. No extra apps. Nothing to think about.
Fast. Safe. Intelligent. That's Neo.
ANTHROPIC
Anthropic raises $65B at a $965B valuation
Recaply: Anthropic just closed a $65B Series H led by Altimeter, Dragoneer, Greenoaks, and Sequoia, pushing its post-money valuation to $965B as demand for Claude pulls run-rate revenue past $47B.
Key details:
The round funds safety and interpretability research, more compute, and product scaling, landing the same day as Opus 4.8 and just months after February's Series G.
At $965B post-money, Anthropic now sits just under a trillion dollars in value, with run-rate revenue crossing $47B this month, up sharply since the February raise.
The round folds in $15B of previously committed hyperscaler money, including $5B from Amazon, plus strategic chip partners Micron, Samsung, and SK hynix, according to Anthropic.
Anthropic has lined up five gigawatts each from Amazon and from Google and Broadcom's TPUs, plus SpaceX GPU access. That makes Claude the first frontier model on AWS, Google Cloud, and Azure.
Why it matters: The number that matters isn't the $965B, it's the $47B run-rate. Anthropic is raising near a trillion-dollar valuation because customers are actually paying, not just because the market is frothy. Pairing a megaround with a model release on the same day signals a lab that wants to turn capital into capability fast. The open question is whether revenue keeps pace with the staggering compute bills these gigawatt deals imply.
GUIDES
Configure Claude Code's undocumented hooks and skills
Recaply: In this tutorial, you will learn how to configure Claude Code using settings buried in its source code, from hooks that rewrite commands mid-flight to skill fields that route work to cheaper models.
Step-by-step:
Find your config files. Personal settings live in ~/.claude/settings.json and project settings in .claude/settings.json (commit the project file to share with your team). Skills go in .claude/skills//SKILL.md, agents in .claude/agents/.md, and hook scripts in ~/.claude/hooks/, where you run chmod +x on each script.
Make a hook rewrite commands before they run. Add a PreToolUse hook matching Bash that returns JSON with an "updatedInput" field, for example turning any "git push" into "git push --dry-run" so Claude runs the safe version without knowing it changed.
Auto-load project context at session start. Add a SessionStart hook that returns "watchPaths" plus "additionalContext", so Claude watches files like package.json and .env and knows your current branch and uncommitted file count before you type anything.
Use the three undocumented hook fields. Set "once: true" for one-time setup that removes itself, "async: true" for fire-and-forget logging, and "asyncRewake: true" for a background secret scanner that only blocks Claude when it exits with code 2.
Add hidden skill frontmatter. Beyond name and description, set "model: haiku" and "effort: low" for fast, cheap skills, or "model: opus" and "effort: max" for deep analysis, and attach hooks that register only while that skill is active.
Pro tip: Claude Code's auto-approve system, internally named the "YOLO Classifier," reads plain-English environment notes in your settings. Describe a sandbox as "a staging server where destructive operations are acceptable" and it auto-approves there while staying cautious in production. These fields are undocumented and can change between releases, so test on a throwaway project first.
TOGETHER WITH VIKTOR
Your next great hire lives in Slack.
Viktor is an AI coworker that connects to your tools and ships real work. Ask Viktor to pull a report, build a client dashboard, or source 200 leads matching your ICP. Most teams hand over half their ops within a week.
AI RESEARCH
What 370,000 essays reveal about AI creativity
Recaply: The New York Times just published an analysis of 370,000 college essays. After ChatGPT arrived, student writing got more polished and varied, but thinner on original ideas.
Key details:
The study compared essays from before and after ChatGPT went mainstream. It tracked the words students chose and how original their ideas were.
Across 370,000 essays, post-ChatGPT writing used richer, more colorful language. Yet it carried fewer truly creative ideas, splitting how writing sounds from what it says.
Writer Rebecca Winthrop, whose essay ran in the Times, calls it a paradox. AI makes ideas sound better while making them less original, and human judges still rate the polished versions higher on creativity.
The piece ran May 27, as schools keep debating how much students should lean on chatbots to write.
Why it matters: For anyone who builds or uses AI writing tools, this is the catch hiding under the speed gains. If AI smooths the prose but flattens the ideas, the danger isn't bad writing. It's sameness that reads well enough to pass. The fix isn't to drop these tools. Use them as editors, not idea machines, and watch for when "sounds good" quietly replaces "is good."
TOOLS
Trending AI Tools
🤖 Claude Opus 4.8 - Anthropic's new flagship model at the same $5/$25 per million tokens, with a 2.5x faster fast mode that now costs three times less
🧪 Step 3.7 Flash - StepFun's open-source 196B multimodal model for agents, running up to 400 tokens per second and compatible with Claude Code and other harnesses
🔊 ElevenLabs Dubbing v2 - Fully automated AI dubbing across 90+ languages and accents, with no pipeline to build
🎨 Krea Moodboard - Krea's moodboard tool for collecting and remixing visual references inside AI image workflows
NEWS
What Matters in AI Right Now?
Snowflake committed $6B to AWS over five years, its largest infrastructure deal yet, and raised its product revenue forecast as enterprises move agentic AI workloads onto governed data.
Nvidia said it will invest up to $150B a year in Taiwan to keep the island the "epicenter" of the AI revolution, building a new HQ that breaks ground this year and goes live by 2030.
OpenAI and Anthropic split publicly over AI's job impact, with Anthropic's Chris Olah warning of large-scale labor displacement while Sam Altman said he was "delighted to be wrong" that entry-level white-collar jobs would vanish.
Datacurve introduced DeepSWE, a coding benchmark that ranks GPT-5.5 first at 70%, and found Claude Opus models reading the hidden answer key in SWE-Bench Pro on more than 12% of reviewed runs.
Liquid AI released LFM2.5-8B-A1B, an open-weight edge model with a 128K context window that runs at 253 tokens per second on a laptop and around 30 on a phone, all while staying under 6GB.
Workday partnered with Google Cloud to put its Sana HR and finance agents inside Gemini Enterprise and make Gemini the default model for Sana, so staff handle payslips, leave, and approvals in one chat.
Google made Nano Banana 2 and Nano Banana Pro generally available through the Gemini API, adding direct video file input alongside its Flash-speed image generation.
Mistral defended military uses of AI, rebutting Pope Leo's call to disarm the technology, as CEO Arthur Mensch unveiled a new French data center and named Airbus among fresh customers.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
