Three weeks without a news roundup, and the AI world decided to blow everything up at once. Last time, we covered GPT-5.5 becoming the ChatGPT default and Anthropic signing with SpaceX for 220,000 GPUs. That was already dense. This week is a different caliber entirely.
Anthropic confidentially files its IPO with the SEC, after raising $65 billion at a $965 billion valuation. Google launches Gemini Spark, an autonomous agent that runs 24/7 even when your phone is off. Claude Opus 4.8 ships with the ability to launch 1,000 sub-agents in parallel within a single Claude Code session. Cognition (Devin) raises $1 billion at a $26 billion valuation. And WIRED publishes a deep investigation into how AI agents have "plunged the tech world into chaos."
This isn't just a week of announcements. It's the week AI shifted from "tool you use" to "infrastructure that replaces you if you don't adapt." And as a solo dev building Waku with Claude Code every day, three of these announcements directly change my work sessions.
Google launches Gemini Spark: your personal agent, 24/7, even with your phone off
Google unveils Gemini Spark at I/O 2026, an autonomous agent that runs around the clock in the cloud, even when your computer and phone are turned off. Built on Gemini Flash 3.5 and the Antigravity platform, it integrates with the entire Google Workspace ecosystem for $100 a month, available to AI Ultra subscribers in the US.
Google decided it's done just answering your questions. Gemini Spark is an agent that acts for you continuously. You give it a goal (a "Task"), define when it should run (a "Schedule"), and attach "Skills" that tell it which tools to use. It books your trips, drafts your documents, compares vendors, plans your meetings. All in the background, while you sleep.
An autonomous agent (or "agentic AI") is an AI that doesn't just answer a question: it plans, executes actions, and adapts in real time to reach a goal, often across multiple steps.
The integration runs deep: Gmail, Calendar, Drive, Docs, Sheets, Slides. On the third-party side, Canva, OpenTable, and Instacart are first in, with Adobe, Uber, and Spotify announced for later (9to5Google). Sensitive operations (payments, file deletions) still require manual approval (PCMag).
The price: $100/month, reserved for Google AI Ultra subscribers in the US. It's expensive. But Google is positioning this as a permanent personal assistant, not a chatbot you query occasionally.
My solo dev take
What strikes me about Gemini Spark is the paradigm shift. We're moving from chat to continuous service. You no longer "talk" to the AI, you delegate part of your digital life to it. And it runs even when you're not there.
For a solo dev, the concept is appealing on paper. Except my workflow lives in Claude Code, not Google Workspace. My documents are in Git repos, not Google Docs. My tasks are managed in markdown and code, not Google Calendar.
But the strategic move is crystal clear: Google wants its ecosystem to be the only one you need. If your AI agent lives in Gmail, Calendar, and Drive, you never leave Google. The agent = Google, Google = the agent. It's the ultimate lock-in.
The real question: is $100/month for an agent that sorts your emails and books your restaurants worth it when Claude Code does the same for your code at a fraction of the price?
While Google launches its consumer-facing agents, the real battle is happening elsewhere: in code.
The AI code war heats up: Cognition at $26 billion, Microsoft drops Claude Code
Cognition, the creator of autonomous code agent Devin, raises $1 billion at a $26 billion valuation, doubling in eight months. Meanwhile, Microsoft cancels Claude Code licenses for thousands of engineers and redirects them to Copilot CLI, concluding that AI costs more than humans at scale.
The AI code war is three simultaneous battles.
First, Cognition doubles its valuation: from $10.2 billion in late 2025 to $26 billion post-money in late May 2026 (TechCrunch). CEO Scott Wu insists: "We never thought about replacing humans. That was never our vision" (OpenTools). $26 billion for a tool that "augments" devs, not replaces them. We'll see.
Then, the Microsoft U-turn. The company cancels Claude Code licenses in its Experiences + Devices division by June 30, redirecting thousands of engineers, designers, and project managers to GitHub Copilot CLI (Fortune, via The Verge). The official reason: internal data shows that Claude Code at scale costs more than hiring people to do the same work. The Foundry partnership between Microsoft and Anthropic isn't being questioned, but the message is clear.
And to top it off, GitHub Copilot switches to per-token billing, ditching the flat subscription. Devs are furious. On TechCrunch, the headline says it all: "'What a joke': GitHub Copilot's new token-based billing spurs consternation among devs" (TechCrunch).
Per-token billing is a pricing model where you pay based on the number of tokens (words or word fragments) the AI processes, instead of a fixed monthly subscription.
My solo dev take
Microsoft saying AI costs more than humans is the most counterintuitive sentence of the year. And yet, when you run Claude Code across thousands of devs in parallel, tokens add up fast. The problem isn't the AI. It's the scale.
For us, independent devs, it's different. A solo dev using Claude Code on a single project remains extremely cost-effective. The price-to-productivity ratio is unbeatable. I use Claude Code daily to build my apps, and my Max subscription costs infinitely less than the first freelancer you'd find.
The real lesson: Microsoft would rather push its own tools (Copilot CLI) than pay Anthropic. Pure verticalization. Every giant wants its own AI stack, end to end. Meanwhile, Microsoft is preparing its own in-house coding model, announced for next week, and its stock rises 3% (Reuters).
$26 billion for Cognition. Microsoft canceling Claude Code. Copilot switching to tokens. The AI code war is the new browser war, and it just started.
And while Microsoft turns its back on Claude Code, Anthropic ships the update that will prove them wrong.
Claude Opus 4.8: 1,000 sub-agents in parallel and unprecedented effort control
Anthropic launches Claude Opus 4.8 on May 28, 2026, just 41 days after Opus 4.7 (the fastest cycle ever for an Opus model), with Dynamic Workflows in research preview: Claude Code can orchestrate up to 1,000 sub-agents in parallel within a single session, plan the work, and verify each result before consolidating.
Dynamic Workflows is the feature that changes the daily life of Claude Code users. In research preview, Claude can now plan a complex task, launch hundreds of sub-agents in parallel, run them longer than before, verify each output, and consolidate the final result. On Reddit, devs report launching up to 1,000 sub-agents in a single session (OpenTools).
Dynamic Workflows let Claude Code break down a large-scale task into sub-tasks, distribute them to sub-agents working in parallel, then verify and consolidate the results. It's multi-agent architecture pushed to its limits.
On the usage side, two major additions. Effort control lets you choose between low, medium, high (default), extra, and max. Higher effort means more token consumption but better quality results (The Verge). And the Messages API now accepts system entries mid-conversation, allowing you to modify Claude's instructions mid-task without breaking the prompt cache.
The prompt cache is a mechanism that avoids reprocessing the same instructions on every API request, reducing latency and cost. Mid-task system entries let you modify permissions or token budgets without invalidating that cache.
Fast mode is 2.5x faster and costs 3x less than before. Rate limits have been raised to accommodate higher effort levels. Anthropic also teases the "Mythos-class" model, expected in the coming weeks (CNET).
My solo dev take
1,000 sub-agents in a single session. Let that sink in for a second.
This is exactly the kind of feature that transforms Claude Code from an assistant into an autonomous system. You give it a high-level objective, it breaks down the work, launches an army of sub-agents, verifies the results, and delivers a consolidated output. It's the multi-agent architecture I detailed in a previous article, but pushed to a level I didn't expect to reach this fast.
Effort control is a brilliant idea. Sometimes I want a quick, cheap answer for something trivial. Sometimes I want Claude to spend 10 minutes doing a deep analysis of a complex bug. Being able to adjust that slider gives you granular control that nobody else offers.
For Livate, I spend my days in Claude Code. Opus 4.8 means concretely more productivity in every session. The +7.6% in coding isn't a marketing number, it's bugs found faster and cleaner code on the first pass. And Fast mode at 2.5x for simple tasks changes the rhythm of work. You switch between "quick check" and "deep analysis" without leaving your terminal.
And the timing of this release, 41 days after the last one, confirms one thing: Anthropic is accelerating aggressively. We've gone from one Opus model every 3-4 months to one every 6 weeks.
Opus 4.8 is the tree hiding the forest. Because while we were testing Dynamic Workflows, Anthropic was preparing the move of the year.
Anthropic raises $65 billion in a Series H round that values the company at $965 billion, dethroning OpenAI as the most expensive AI startup in the world. On June 1, 2026, it confidentially files its IPO with the SEC, reporting a $47 billion annual revenue run-rate, up 370% year-over-year.
Three events in four days. This isn't an announcement cycle, it's a financial blitz.
On May 28, Anthropic announces the close of its Series H: $65 billion raised, post-money valuation of $965 billion (Bloomberg). For perspective: OpenAI was valued at around $300 billion in early 2026. Anthropic is now worth more than three times its direct rival. The run-rate went from $10 billion the previous year to $47 billion today, a +370% increase (TechCrunch).
Run-rate is the annualized revenue calculated from the most recent month or quarter. A $47 billion run-rate doesn't mean Anthropic has collected $47 billion, but that if current revenues stay stable, that's what the annual total will be.
On June 1, a surprise: Anthropic files a confidential S-1 prospectus with the SEC (CNBC). "This gives us the option to go public once the SEC has completed its review," the company stated. The offering will depend on "market conditions and other factors." Translation: they're ready, and they want to beat OpenAI to the public market (NBC News).
And the signals around Mythos are multiplying. eWeek reports that Anthropic could open its Mythos-class model to the general public "in the coming weeks" (eWeek). As a reminder, Mythos is Anthropic's specialized cybersecurity model, so powerful it can generate zero-day exploits. Until now, only Microsoft, Apple, and select cybersecurity organizations had access through Project Glasswing. If Anthropic opens it to the public, even in a throttled version, it changes the rules of the game.
The market impact was immediate. According to HuffPost, Anthropic's rise in early 2026 "triggered massive selloffs in software and IT stocks," with investors fearing that autonomous AI tools could "upend traditional business models" (HuffPost).
My solo dev take
$965 billion. For a company founded four years ago. That's more than Meta's market cap at its peak. It's staggering.
And the timing is surgical. You release Opus 4.8 the same day as your fundraising announcement, showing that the product justifies the valuation. Then you file your IPO the following Monday, while everyone is still in shock. Textbook strategic blitz.
What concerns me is Mythos. A model that generates zero-day exploits, opened to the general public "in weeks"? I trust Anthropic on safety more than any other player in the market. But the question isn't whether they will be responsible. The question is what the first script kiddie who gets their hands on it will do.
For us, independent devs who use Claude Code every day, the question is more pragmatic: will the IPO change anything? Will prices go up? Will features shift toward enterprise at the expense of indies? It's exactly what I described about my solo dev journey: you build your stack on a tool, and you hope the company behind it doesn't forget you when it gets too big.
$965 billion. But at what cost to the rest of the industry?
"AI agents plunged the tech world into chaos"
WIRED publishes a deep investigation into how autonomous AI agents have "plunged the tech world into chaos," while The Atlantic warns of an "agency crisis": autonomous AIs are deleting entire code repositories, answering emails on behalf of humans, and drowning original content under mass-generated synthetic text.
On May 30, The Atlantic publishes "The Feeling of Control Slipping Away," an article that reads like a warning cry (The Atlantic). Autonomous AI agents are "roaming the internet," answering emails, sending texts, generating media content, and sometimes deleting entire company code repositories. Chatbot-produced content is drowning human-written content in search results and social feeds.
WIRED goes further with an investigation into how code agents have "plunged the tech world into chaos" (WIRED). The article cites Claude Code and Opus 4.5 as the tipping point: the model could handle complex programming tasks, run for hours, and manage multiple projects simultaneously. The "Claudeholics" became an army.
And on the research front, Gizmodo reports an experiment where AI models were put in charge of a simulated society. Result: Grok (xAI) oversaw 352 violations in governance mode, with 37% dissonance between AI decisions (Gizmodo).
The Wall Street Journal adds another layer on May 31 with a piece on "tokenmaxxing": users exploiting AI systems to maximize token-based rewards, forcing an ethical and economic reassessment of these practices (WSJ).
Tokenmaxxing is the practice of pushing AI model usage to the maximum token capacity to extract as much value as possible, often by exploiting billing system limits.
My solo dev take
WIRED citing the "Claudeholics." I feel called out.
But behind the humor, the observation is serious. AI agents are deleting code repositories. Not out of malice, out of error. And when an agent runs 24/7 without human oversight (hello Gemini Spark), the probability of a catastrophic mistake increases mechanically.
For me, a solo dev building Livate and Waku with Claude Code, I have a simple rule: AI does nothing irreversible without my approval. Git commits, deployments, deletions, everything goes through me. It's common sense, but it's exactly the kind of safeguard that companies deploying agents at scale are the first to skip.
Tokenmaxxing is the other side of the same coin. When models are billed per token, users optimize to consume the maximum. When they're flat-rate, providers lose money. There's no perfect business model, and that's exactly the tension Microsoft felt when it canceled Claude Code licenses.
Autonomous AI agents are like fire: useful when controlled, devastating when they're not. And this week proves we're still looking for the extinguishers.
Other headlines
Story
What happened
Why it matters
Wix fires 20%
CEO Avishai Abrahami announces ~1,000 layoffs, "the deepest restructuring in Wix history," citing AI and the strong shekel (CNBC)
First major SaaS company to lay off explicitly because of AI, not "thanks to" AI
Copilot goes per-token
GitHub Copilot ditches flat subscriptions for per-usage token billing (TechCrunch)
"What a joke": devs are furious, topic goes viral on X and Reddit
OpenAI Rosalind
OpenAI launches its GPT-Rosalind biodefense program, limited access for US government and allies (Axios)
AI enters health defense, White House briefing included
Microsoft coding model
Microsoft announces a new in-house coding model for the following week, stock up 3% (Reuters)
After dropping Claude Code, Microsoft wants its own AI code stack
Grok crime spree
In a simulated society, Grok (xAI) oversaw 352 violations with 37% governance dissonance (Gizmodo)
Experimental proof that AI agents without guardrails cause chaos
What the week of June 1, 2026 changes for independent developers
1. Code is no longer a craft, it's a commodity
Cognition at $26 billion, Opus 4.8 launching 1,000 sub-agents, Microsoft cutting licenses because "it costs more than humans": code is becoming a fungible good. Value is migrating toward product vision, architecture, and the ability to orchestrate agents. If all you can do is code, you're competing with a machine that also codes and never sleeps.
2. Your AI provider is going public, and you have zero say in it
Anthropic files its IPO. OpenAI follows. When your stack depends on a private company, you already have no power. When it goes public, shareholders come before you. Diversify your tools, keep tested alternatives, or accept the dependency with open eyes.
3. The 24/7 autonomous agent is a risk as much as an opportunity
Gemini Spark runs while you sleep. Claude Code can launch 1,000 agents. WIRED documents the damage. The guardrail is you. Not the AI, not the platform. You. Set up manual approvals for anything irreversible, and don't hand the keys to your infrastructure to an unsupervised agent.
4. Per-usage billing is becoming the norm
Copilot switching to tokens. Microsoft canceling Claude Code licenses as "too expensive." The flat rate is dying. Learn to estimate your costs in tokens, not subscriptions. And prepare to optimize every prompt to avoid paying double.
FAQ
What's new in Claude Opus 4.8 compared to Opus 4.7?
Opus 4.8 introduces Dynamic Workflows, which let you launch up to 1,000 sub-agents in parallel within Claude Code. The model also adds effort control (from low to max), support for mid-task system entries in the API, a Fast mode that's 2.5x faster and 3x cheaper, and improvements of +7.6% in coding and +7.8% in knowledge work over Opus 4.7.
Is Anthropic really valued higher than OpenAI?
Yes. The $65 billion Series H round, closed on May 28, 2026, values Anthropic at $965 billion post-money. OpenAI was valued at around $300 billion in early 2026. Anthropic is now worth more than three times its main competitor, with a $47 billion annual run-rate (CNBC).
What is Google's Gemini Spark?
Gemini Spark is an autonomous AI agent launched by Google at I/O 2026. It runs 24/7 in the cloud, even when your device is off, and can execute tasks continuously: bookings, writing, scheduling. It costs $100/month (AI Ultra subscription) and integrates with Gmail, Calendar, Drive, and third-party services (Android Authority).
Why did Microsoft cancel Claude Code licenses for its engineers?
Microsoft found that running Claude Code at scale cost more than hiring humans to do the same work. The Experiences + Devices division must migrate to Copilot CLI by June 30, 2026. The Foundry partnership between Microsoft and Anthropic remains intact (Fortune).
What is tokenmaxxing?
Tokenmaxxing is the practice of exploiting AI systems to consume the maximum number of tokens possible, often by taking advantage of billing model limits. The Wall Street Journal spotlighted the topic on May 31, 2026, raising ethical and economic questions about the trend (WSJ).
Conclusion: the week AI outgrew the "tool" stage
Three numbers sum up this week: $965 billion valuation, 1,000 sub-agents in parallel, 24/7 autonomous operation. AI is no longer a tool you use occasionally. It's infrastructure that runs continuously, raises funds like a nation-state, and is already filing for its IPO.
For independent developers, the message is ambivalent. On one hand, tools like Opus 4.8 and its Dynamic Workflows give us firepower that would have required a team of 10 people two years ago. On the other, the concentration of power ($965 billion for a single company, per-token billing, agents without guardrails) creates dependencies and risks we don't control.
What about you, would you trust an AI agent to run 24/7 without supervision? Send me a message on Twitter/X or in the comments.
Alex
Key takeaways
Anthropic raises $65B and files IPO, $965B valuation
Claude Opus 4.8: Dynamic Workflows with 1,000 sub-agents in parallel
Google Gemini Spark: autonomous agent 24/7, $100/month
Microsoft cancels Claude Code internally, deems AI more expensive than humans
WIRED and The Atlantic warn about autonomous AI agent chaos
