Anthropic: Version Number Recedes, Embedding Reigns Supreme

OpenAI's Codex is making rapid progress, and it seems that Anthropic can no longer sit still.

This month, there has been a sudden increase in product news from Anthropic. Developers can say on X that the logo of Opus 4.8 has appeared in the Google Vertex backend; three months ago, in the code leaked during an npm update of Claude Code, there was the mention of "sonnet-4-8"; the Mythos model, which has long been unavailable to the public, also briefly appeared in the Claude interface.

It's hard to say that this isn't a defensive measure by Anthropic. The consecutive appearance of these three pieces of news is quite a coincidence. However, upon closer inspection, both Opus 4.8 and Sonnet 4.8 are just rumors on social media and have not been verified by mainstream media. The only reliable progress is that Mythos, which was released in April, is moving from a small-scale trial to a larger-scale productization.

Rather than focusing on the existence of a new version, what's more worth noting is the fundamental change happening at Anthropic - the company's annual revenue has increased from approximately $9 billion at the end of 2025 to over $44 billion in May 2026.

At this scale, the core competitiveness is no longer "whose model is smarter", but rather who can integrate AI more deeply into the daily operations of enterprises, making it difficult for customers to switch. Clearly, Anthropic is considering this proposition.

Rumors, Traces, and a Verified Clear Line

Currently, there is only one post on X about Opus 4.8. Although Anthropic did have the habit of being first discovered by developers in the Vertex backend and then making an official announcement, it doesn't mean that 4.8 is a certainty. Before the official statement, it's just a reasonable assumption.

Sonnet 4.8 originated from a development error. On March 31st, when Claude Code updated its npm package, the developer forgot to exclude the.map file from the configuration file, and a 59.8MB code mapping file was leaked to the public repository. It contained the mention of "sonnet-4-8", but the person who reported it marked it as "unconfirmed".

The appearance of a model name in the client code could have many possibilities: it could be a placeholder, a long-term development branch, or just some uncleaned residue. Using this to claim that "Anthropic skipped 4.7 and went straight to 4.8" is weak evidence and may not be reliable.

However, Mythos is different. Its progress is real. On April 7, 2026, Anthropic officially released the Mythos Preview, opening it to 12 giants such as AWS, Apple, Cisco, Microsoft, and NVIDIA, as well as more than 40 key open-source organizations through Project Glasswing on an invitation-only basis. Anthropic provided a special call quota of $100 million for partners and set up a special fund of $4 million for donations and support to open-source security organizations.

The brief appearance of the Mythos entrance in the Claude interface in May was not a "first leak", but rather an already released product expanding into more scenarios. Its commercialization path has also been determined: $25 per million input tokens and $125 per million output tokens, about five times that of Opus 4.6, targeting the high-end security vertical market.

Mythos: The Commercial Turn of Security Strategy

Among the three pieces of news, Mythos best illustrates Anthropic's strategic shift. Previously, Anthropic was always worried that this network security model would be used by both good and bad people, so it was not made available to the public. In official tests, Mythos discovered a 27-year-old remote crash vulnerability in OpenBSD and a 16-year-old bug in FFmpeg that had evaded 5 million automated tests.

This ability to discover old vulnerabilities does exceed that of ordinary models, which is also the reason why Anthropic was hesitant to release it before.

The launch of Project Glasswing means that Anthropic has changed its mind. It has brought 12 giants into the access network, granting permissions only to invited partners, disclosing vulnerabilities according to regulations, and directly pricing its security capabilities for sale.

After Cloudflare's participation, it was mentioned that even without an additional security layer, Mythos would reject some dangerous requests on its own, but this rejection was unstable. The result of the same task might be different in a different context. This shows that relying solely on the model's own judgment of security is unreliable, and access control, real-time monitoring, and multi-party collaboration are also needed for constraints.

From a commercial perspective, the pricing of Mythos reveals Anthropic's real intentions. It no longer regards AI security as a pure cost or a brand halo, but rather as a high-threshold revenue category. Enterprise security teams are willing to pay $125 per million output tokens, which shows that "security first" can not only make money but also command a higher price.

Anthropic has hired Wilson Sonsini to prepare for an IPO and is now in the pre-IPO quiet period. Security is no longer a burden on commercialization but may instead become the most unique profit growth point in the IPO story. The pricing logic of Mythos is actually a microcosm of Anthropic's overall revenue explosion.

From Revenue Explosion to Agentic Embedding

Since 2026, the company's revenue curve has been more telling than any version number. Dario Amodei revealed at the Code with Claude conference that the company initially thought its annual revenue would increase by 10 times, but in the first quarter, the actual year-on-year growth rate reached 80 times. The annualized revenue climbed from approximately $9 billion at the end of 2025 to over $30 billion in April and further exceeded $44 billion in May.

The annual revenue of Claude Code alone has exceeded $2.5 billion. There are more than 1,000 enterprise customers with annual payments of over $1 million, compared to only about 500 in February. Doubling within two months shows that large customers are not just trying out a new toy but are making Claude an indispensable daily tool.

This reveals the fundamental shift in Anthropic's core competitiveness: from "leading in model performance" to "deep embedding of tools and workflows". Enterprises spend millions of dollars on Claude each year not for a smarter chatbot but for an intelligent infrastructure that can be integrated into the daily processes of development, operation, and security auditing. Code review, automated testing, and vulnerability scanning - these processes are being deeply intervened by AI.

So at the Code with Claude conference in early May, Anthropic did not release a new model but instead talked about the next stage of AI agents. This "next stage" is rapidly taking shape. In late May, it was reported that Anthropic is testing a brand-new "Memory Files" system - Claude will automatically take classified notes during chats and can be called up when needed, and users can edit or delete them at any time.

Compared with the "classic memory" mode in the past, which compressed all information into a single summary and caused old memories to be overwritten by new ones, the file memory is more like a built-in personal Wiki-style structured memory library, breaking the capacity limit and significantly improving accuracy.

With the background automatic sorting function called "Dreams", Claude can automatically organize memories, merge duplicates, and replace outdated information during idle time. After enterprises such as Netflix and Rakuten were among the first to access it, the document verification speed increased by 30%, and the first-time processing error rate decreased by 97%.

These memory upgrades are not isolated functions but are laying the foundation for the always-on background agent platform Conway - an AI agent that can be triggered by external signals and run in the background around the clock.

Although Conway is still in the internal testing stage and has not been officially released, it, along with Memory Files and Dreams, all point in the same direction: Claude is transforming from a "question-and-answer" chat tool into a digital colleague that "actively remembers, runs continuously, and is deeply embedded in the workflow".

A few more percentage points in benchmark tests are becoming less and less persuasive to enterprise decision-makers. Continuing to focus on version numbers will only lead to homogeneous competition. The real moat lies in whether AI can become an indispensable part of the enterprise workflow.

Meanwhile, Google announced at I/O 2026 that Gemini 3.5 Pro will be launched in June, and there are market rumors that OpenAI's next-generation model may be updated in June. The competition is intensifying, and Anthropic has chosen to pursue differentiation: Opus maintains flagship performance, Sonnet offers cost-effectiveness, and Mythos opens up the high-threshold vertical field of security.

By taking these three-pronged approaches, the goal is one - to transform Anthropic from a "better chatbot" into "the intelligent infrastructure that enterprises cannot do without" before the IPO.

Version Numbers Take a Backseat, and Embedding Depth Is the Endgame

There may be many product releases in June, but it won't be a dramatic moment of "ASI showdown" or "a fierce battle among the three giants". The signals released by Anthropic in May are just a normal spillover during the product layout period of a super unicorn.

What really determines the value of Anthropic is not whether Opus 4.8 or Sonnet 4.8 has been leaked, but whether the security capabilities of Mythos can command a continuous high price and whether the AI agent of Claude Code can become an indispensable tool in the daily operations of enterprises.

As the industry competition shifts from "whose model is smarter" to "whose AI is more irreplaceable", version numbers will become increasingly unimportant.

Enterprises won't switch suppliers just because Opus has been upgraded from 4.7 to 4.8, but they will continue to pay because Claude Code has been integrated into the entire development pipeline, Mythos has become the default part of security auditing, and Memory Files have recorded all the project contexts of the team. Embedding depth and commercial stickiness are the ultimate measurement criteria.

This article is from the WeChat official account "AI Contrarian", author: Changqing. Republished by 36Kr with permission.