ANTHROPIC OFFICIALLY CLOSES $65 BILLION, SURPASSES OPENAI, AND STARTS WALKING TOWARD WALL STREET
Look, a couple of days ago we told you about Anthropic being in talks to raise at a $900 billion valuation. That was the rumor stage. Yesterday it became real. The deal closed. $65 billion raised. $965 billion post-money valuation. And the sentence buried in the TechCrunch article that you should not skim past: this is, in the words of people close to the deal, likely Anthropic’s last private fundraising round before going public.
They are going to the stock market. That is the actual news inside the news.
Let me put the money in perspective because $65 billion is not a number the human brain handles well. That is roughly the entire GDP of Luxembourg raised by a company that is five years old and started as a safety-focused AI research lab that most people outside San Francisco had never heard of two years ago. The round was co-led by Altimeter, Dragoneer, Greenoaks, Sequoia, Capital Group, Coatue, and D1 Capital Partners. Institutional investors including Baillie Gifford, Blackstone, Brookfield, Fidelity, and DST Global also participated. Samsung, SK Hynix, and Micron came in on the hardware infrastructure side. Five billion of the total was money Amazon had already committed back in April and is now officially counted in this round.
The number that matters more than any of that is $47 billion. That is Anthropic’s annualized revenue run rate as of this month. For reference, they were at a $30 billion run rate just a few months ago. The Wall Street Journal is reporting that the company expects a 130 percent revenue surge that will push it to its first operating profit. Growing from $30 billion to $47 billion in run rate while also closing the biggest private AI round in history is not a company struggling to find its footing.
Now the competitive piece, because this is where it actually gets interesting. OpenAI raised $122 billion in March at an $852 billion post-money valuation. At the time, that made OpenAI the single most valuable private AI company on earth. As of Wednesday, that is no longer true. Anthropic at $965 billion now sits above OpenAI at $852 billion. The company that Dario Amodei left OpenAI to build has officially outvalued the company he left. That is a sentence worth rereading.
The IPO timeline is the thing to watch from here. No date has been set. Nothing has been officially confirmed. But the framing of this round as the last before a public offering is not accidental phrasing. You do not raise $65 billion and stay private forever. You do not bring in Blackstone, Fidelity, and Brookfield without giving them a path to liquidity. The whole architecture of this round is designed to set up a public offering, and the question is not whether but when. Meanwhile, OpenAI and xAI are also heading for public markets. Three AI giants going public in some proximity to each other is going to be one of the stranger chapters in financial history.
ANTHROPIC ALSO DROPPED A NEW CLAUDE YESTERDAY AND IT CAN BOSS AROUND HUNDREDS OF AI AGENTS SIMULTANEOUSLY
On the exact same day Anthropic closed its $65 billion fundraise, the company released Opus 4.8, the newest version of its most powerful publicly available model. The timing is not a coincidence. This is what a well-run company looks like when it wants to send a message to the market: we have the money and we have the product.
Opus 4.8 replaces Opus 4.7, which launched 41 days ago. If you are not embedded in AI development, 41 days between major model releases is fast. For reference, the current Sonnet and Haiku models from Anthropic are three and seven months old. The acceleration matters because it signals something about the pressure the company is operating under. OpenAI shipped updates to Codex, Google pushed an improved Gemini Flash release, and Opus 4.7 got a reception from developers that was described, charitably, as cool. The competitive pressure forced a faster turnaround than Anthropic typically runs.
So what is actually new in this one? Two things worth paying real attention to.
First, the model has been specifically tuned to handle uncertainty better. Anthropic says Opus 4.8 is more likely to flag when it is not confident about something and less likely to make claims it cannot back up. Bridgewater Associates, the hedge fund, was one of the early testers. Their feedback was that the biggest improvement they noticed was the model proactively raising problems with inputs and outputs rather than leaving the user to catch errors downstream. In professional environments where mistakes cost money, that is not a minor improvement. The model that tells you it is not sure before you rely on its answer is fundamentally more useful than the model that sounds confident and is wrong.
Second, Anthropic launched Dynamic Workflows alongside the model, currently in research preview. This is the feature that breaks new ground. Dynamic Workflows lets Opus 4.8 manage complex tasks across hundreds of parallel subagents running at the same time. The example Anthropic gave was Claude Code handling a codebase migration spanning hundreds of thousands of lines of code from start to a finished, merged pull request, with the existing test suite serving as its quality bar.
To say that in plain terms: this is Claude running an army of smaller AI workers, coordinating what they do, checking the results, and delivering a finished product at a scale that a human engineering team would take weeks to complete. That is not incremental progress. That is a genuine shift in what software development teams will look like a year from now.
The other notable piece in the release is a hint about Mythos, Anthropic’s restricted cybersecurity model. The company said it expects to bring Mythos-class capabilities to all customers in the coming weeks. The most powerful and most carefully guarded AI Anthropic has built is about to get a much wider release. The question of how responsibly the industry actually handles that is one that will answer itself fairly quickly.
THE STARTUP THAT BUILT DEVIN THE AI PROGRAMMER JUST TRIPLED ITS VALUATION TO $26 BILLION IN EIGHT MONTHS
Cognition, the company behind Devin the autonomous AI software engineer, closed over $1 billion in new funding this week at a $26 billion post-money valuation. Eight months ago, in September, this same company was valued at $10.2 billion after a $400 million round. That means Cognition added roughly $16 billion in paper value in less time than it takes to have a baby.
Let that number settle for a moment. From $10 billion to $26 billion without a headline product launch, without a major acquisition, without breaking any public records on benchmarks. The valuation went up because investor demand for AI coding companies went up, and because when you look at the early revenue data for platforms in this space it is becoming very hard to deny that something significant is happening.
The round was co-led by Lux Capital, General Catalyst, and 8VC. Existing investors including Elad Gil, Founders Fund, and Soma Capital came back in. New money arrived from Ribbit Capital, Atreides, and Layer Global. When the people who already own the equity are willing to write larger checks at three times their entry price, and a completely new group of sophisticated investors joins alongside them, it typically means the internal revenue figures made everyone involved nervous that waiting would be a mistake.
Devin, for those who have not been following closely, is the product that showed up on a lot of front pages when Cognition first introduced it back in 2024. The original pitch was simple and, depending on your career, either exciting or slightly unsettling: an AI that could autonomously complete software engineering tasks from start to finish without needing a human to guide each step. The early demos were impressive enough to kick off a real conversation about what junior developer careers look like in five years. Devin was imperfect at launch. It is better now. The gap between what it could do a year ago and what it can handle today is measurable and meaningful.
The broader context here is that the AI coding market has become extremely crowded and extremely expensive in a short period. Cursor sits in the billions based on its growth among professional developers. Replit is at $9 billion. GitHub Copilot has the advantage of being baked into the most widely used developer platform on earth. Windsurf, which Google acquired, adds another well-funded player. Every major company wants to own the layer where developers touch AI, and right now an enormous amount of money is chasing that position.
What the Cognition raise signals is that investors believe autonomous AI engineers represent a different category than AI coding assistants. A coding assistant helps you write better code faster. An autonomous agent does the entire job while you step away from the keyboard. If that distinction actually holds in production environments and not just in venture pitch decks, then the company leading that race at a $26 billion valuation starts to look less like speculation and more like an early position on a very large market. The test is whether Devin becomes genuinely reliable enough that enterprises trust it with real work at scale. That answer is still being written.
VISA JUST GAVE AI AGENTS A PAYMENT CREDENTIAL AND THE ERA OF THE SPENDING BOT HAS QUIETLY STARTED
This story slid under the radar on a busy news day, but it is worth stopping to think through what it actually means because the implications are bigger than the headline suggests.
Visa has made an undisclosed investment in Replit, the AI coding platform, and the two companies are actively developing something called Visa’s Trusted Agent Protocol. The concept is this: a system that allows AI agents to securely identify themselves when they make payments on a user’s behalf, verifying the agent’s intent and relevant customer context so that a transaction completed by a bot rather than a human can be authorized and trusted by the payment network.
Put plainly, Visa is building the plumbing that allows AI agents to have payment credentials. They are making it possible for the bots to buy things.
Replit is the platform where millions of developers build and ship applications, and increasingly where non-technical people are using AI to create software without writing any code themselves. The company went from a $3 billion valuation to $9 billion in under six months, which tells you how fast the market for AI-assisted development is growing. More than 1,000 Visa employees have been using Replit internally for prototyping, which is a fairly common origin story for corporate investment deals. You use the product, the product works, you write the check.
But the deeper story is not the Replit-Visa partnership on its own. It is the race to own the payment infrastructure of what people are calling the agentic economy. Right now, AI agents can browse the web, write code, send emails, schedule meetings, and coordinate complex multi-step tasks. What they largely cannot do yet is spend money in a verified, trusted, and accountable way. That is the missing piece. Once an AI agent can make a verified payment, the range of things it can accomplish on your behalf expands dramatically. An agent that can book flights, pay for hotel rooms, order supplies, and renew subscriptions without you touching a keyboard is a fundamentally different product than one that just drafts the email asking someone else to do those things.
Robinhood has started letting users deploy agents for trading. Google is pushing agents for shopping. Now Visa is ensuring that when those agents show up at checkout, there is a trusted system to verify who they are and what they intend before the transaction goes through. The Trusted Agent Protocol is specifically designed to prevent a bot from making purchases you never authorized, or being manipulated by a malicious website into spending your money.
None of this is live in a finished product yet. Visa was explicit that the joint work with Replit is in an exploratory stage. But the direction is unmistakable, and the companies getting into position now are not doing it for fun. Visa already processes trillions in payments every year. If AI agents become active participants in the economy, which is the explicit bet that multiple trillion-dollar companies are making simultaneously, Visa has an obvious interest in making sure those agents transact through its network. This is not a sideshow. This is Visa planting its flag in the next version of commerce before the map has been drawn.