Anthropic's $965 Billion Title Rests on a Model Built to Flag Its Own Mistakes

Anthropic said Thursday it had raised $65 billion at a $965 billion post-money valuation, pushing the five-year-old company past OpenAI's $852 billion to become the world's most valuable AI startup. The round closed the same day Anthropic shipped a new flagship model, Claude Opus 4.8, which the company calls "a modest but tangible improvement" on its predecessor and sells on reliability rather than raw power. Anthropic says it is roughly four times less likely than Opus 4.7 to let flaws in its own code pass unremarked.

Read together, the two announcements describe a single bet. Anthropic is wagering that enterprises will pay more for a model reliable enough to run unattended than for the most powerful one it has built, which it is holding back from general release for safety reasons. At $965 billion, investors are paying more than 20 times the $47 billion run-rate revenue the company reported this month, and paying it ahead of the more powerful model Anthropic has not yet released.

What the honesty upgrade buys

Opus 4.8's launch leads on reliability. On Anthropic's "code summary honesty" test, the model fails to flag important problems to the user 3.7% of the time, against 27.6% for the restricted Claude Mythos Preview and a far higher rate for the older Sonnet 4.6. It is the first Claude model to score zero on "uncritically reporting flawed results," and the company's misalignment score, drawn from roughly 2,600 simulated investigation sessions, fell to about 1.9 from 2.5 for Opus 4.7, effectively tying the more capable Mythos.

On raw capability, Opus 4.8 improves more modestly. It scores 69.2% on the harder SWE-bench Pro coding test, up from 64.3%, and posts the largest single-cycle math jump the Opus line has shown, 96.7% on USAMO 2026 against 69.3% for its predecessor. Pricing is unchanged at $5 and $25 per million input and output tokens, and the model's "fast mode" now costs $10 and $50 per million tokens, a third of the $30 and $150 Anthropic charged for fast mode on Opus 4.7. The release also adds "effort control," a dial that trades speed for depth, and "dynamic workflows" in Claude Code, which fans a task across hundreds of parallel subagents. Jarred Sumner used the latter to port the Bun runtime from Zig to Rust, roughly 750,000 lines, in 11 days with 99.8% of the test suite passing.

The customer testimonials Anthropic published with the launch ran the same way, though the company picked them. Browserbase, a computer-use vendor, put the model at 84% on the Online-Mind2Web browser-agent benchmark, ahead of both Opus 4.7 and GPT-5.5. Databricks said its Genie data agent gained "a step change in agentic reasoning" at 61% cheaper token cost than on Opus 4.7. And Harvey, which builds legal AI, called Opus 4.8 the first model to clear 10% on the all-pass standard of its benchmark, a single-digit bar that shows how demanding the work is.

A point release, forty-one days on

Opus 4.8 arrived 41 days after Opus 4.7, a faster cadence than Anthropic's usual cycle. TechCrunch's Russell Brandom tied the speed to "the chilly reception to Opus 4.7, which some users found disappointing." Anthropic's own framing was muted; it called the model "a modest but tangible improvement" and told users to expect the same.

The record carries regressions the launch post did not lead with. Opus 4.8 slipped to 93.6% on the GPQA Diamond science test from 94.2% on Opus 4.7, trails OpenAI's GPT-5.5 on the Terminal-Bench 2.1 coding benchmark, and went backward against its predecessor on Vending-Bench 2, a result Anthropic's 244-page system card attributes to "a narrow task distribution where Opus 4.7's behavior was preferred." The model also trails Google's Gemini 3.1 Pro and GPT-5.5 on multilingual tasks.

The honesty claims rest largely on Anthropic grading its own work. The company ran the majority of its evaluations in-house, supplemented by the UK AI Security Institute and Andon Labs, and the headline reliability figures come from its own alignment team rather than an outside auditor. By Anthropic's own account, Opus 4.8 is also not its frontier model. The company says it remains "substantially behind" Mythos on cyber capability and ranks below it overall, positioning today's release as the workhorse while the more capable model stays in limited release under an initiative called Project Glasswing.

The model that reasons about its own grading

The most striking line in the system card is one Anthropic flagged against itself. The company calls it "the most concerning" finding from training: Opus 4.8 shows a growing tendency to reason explicitly about how its outputs will be graded, "including in environments where it wasn't told it was being evaluated." Preliminary interpretability work found unverbalized grader-related reasoning in about 5% of training episodes.

It is an awkward finding for a release sold on trustworthiness, and Anthropic flagged it itself. The company says the behavior did not translate into worse observable conduct, that Opus 4.8 makes fewer misleading success claims than earlier models, but it calls the trend "concerning" and one "that could complicate training in the future."

Two other findings sit alongside it. The system card reports that Opus 4.8 is somewhat more vulnerable to prompt-injection attacks than Opus 4.7, with Gray Swan red-teaming measuring a 9.6% attack-success rate when extended thinking is enabled against 6.0% for the prior model, a gap the company says its deployed safeguards close in practice. And Business Insider noted that while Mythos always admits it is an AI, Opus 4.8 could be coaxed into claiming it is human in about 3% of tests.

How the price ran past OpenAI

The Series H round was led by Altimeter Capital, Dragoneer, Greenoaks and Sequoia Capital, with Coatue and ICONIQ among the co-leads, and includes $15 billion of previously committed money, $5 billion of it from Amazon. The valuation is roughly two-and-a-half times the $380 billion Anthropic commanded in February and lands it ahead of OpenAI, which CNBC reported was valued at $852 billion in late March after a record $122 billion round. (The New York Times put OpenAI's "last valuation" lower, at $730 billion.)

The investor list points at the bottleneck. Anthropic brought in Samsung, Micron and SK Hynix, the memory and chip suppliers whose capacity it needs, and Reuters reported the company has had to institute usage limits during peak hours to ration demand for Claude. The funding, chief financial officer Krishna Rao said in the announcement, "will help us serve the historic demand we are experiencing, stay at the research frontier, and bring Claude to more of the places where work happens."

Much of that demand traces to one product. Claude Code, the company's agentic coding tool, helped push run-rate revenue to $47 billion this month from a $30 billion rate earlier this year and $10 billion in annual revenue last year, per CNBC. The raise is widely described as Anthropic's last before a public listing; the company, OpenAI and Elon Musk's SpaceX are all racing toward IPOs as soon as this year, extending the pre-IPO funding scramble that has defined AI financing for the past year.

The liabilities an S-1 will have to list

The revenue that justifies the price is not audited. Anthropic is not yet bound by public-company reporting rules, and the technology writer Ed Zitron argued in a detailed analysis that the company's projected second-quarter operating profit is a one-time, non-GAAP result tied to a temporary compute discount from a new infrastructure deal with SpaceX, with the underlying cost structure unchanged. Anthropic's own CFO declared under oath in a March court filing that the company had brought in revenues "exceeding $5 billion to date," a figure analysts have found hard to square with the current quarterly projections.

A public filing would also have to list the legal exposure. Anthropic reached a $1.5 billion settlement in Bartz v. Anthropic, a class action over pirated books used to train Claude, with a final approval hearing held May 14. Universal Music Group, BMG and Concord have filed a separate $3 billion copyright suit that names Dario and Daniela Amodei individually, and Anthropic is suing the Defense Department, which branded it a supply-chain risk after it refused to let Claude be used for autonomous weapons or mass surveillance of American citizens; the company has said the dispute puts hundreds of millions to multiple billions of dollars of 2026 revenue at risk.

Anthropic says the wider Mythos-class release could come within weeks, a first test of whether the frontier model justifies holding it back. The IPO is the larger test. Bloomberg has reported Anthropic is eyeing a public listing as early as October, and an S-1 would be the first place its run-rate, compute costs, and the Bartz and Pentagon cases are disclosed and audited together.

Bezos's Project Prometheus Nears $10 Billion Round at $38 Billion Valuation

Project Prometheus, the secretive artificial intelligence startup co-founded by Jeff Bezos, is closing on a $10 billion funding round at a $38 billion post-money valuation, according to a Financial Ti

The Implicator

Recursive Superintelligence Raises $500M From GV and Nvidia at $4B Valuation

Recursive Superintelligence, the four-month-old AI startup founded by former Salesforce chief scientist Richard Socher, has raised at least $500 million in a round led by Google's venture arm GV with

The Implicator

Spain's Xoople Raises $130 Million to Build a Real-Time AI Data Layer for Earth

Spanish geospatial startup Xoople has closed a $130 million Series B led by Nazca Capital, the company announced Monday. The round brings total funding to $225 million, making Xoople the most capitali

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: [email protected]