All three aces of Anthropic have been revealed. Mythos 1 makes its first appearance, and Opus 4.8 is exposed.
All three of Anthropic's "killer apps" have been exposed at once!
Just today, developers accidentally discovered a new model identifier, claude-opus-4.8, in the Google Vertex backend.
After more than a month, another major version update of Claude is on the horizon!
Almost on the same day, the "security-specific model" codenamed Mythos 1 briefly appeared in the Claude interface.
Moreover, new strings pointing to Claude Code and Claude Security were added to the source code.
In addition, the previously widely circulated 510,000 lines of leaked code spoiled Claude Sonnet 4.8 in advance.
The code also confirmed that it will skip version 4.7 directly and may be released shockingly in mid-June.
Sonnet 4.8 is expected to inherit the visual upgrade of Opus 4.7, with a visual accuracy of over 98% in UI design drafts and architecture diagrams.
Moreover, the code it generates will be cleaner and more concise, and it also supports advanced reasoning.
It's not hard to see that Anthropic is advancing on three fronts and going all out.
After all, on the eve of the IPO, OpenAI is still pressing hard: the next-generation GPP-5.6 has emerged, with extremely powerful front-end explosive power.
In addition, Google's Gemini 3.5 Pro will also enter the game in June.
The ultimate ASI battle in Silicon Valley is on the verge of breaking out, and the "close combat" among the three giants has just truly begun.
Opus 4.8 Appears, All Cards Are Exposed
In April, Claude Opus 4.7 had a major update and defeated the world's top-notch Gemini 3.1 Pro and GPT-5.4.
In just over a month, Anthropic has trained Opus 4.8 internally.
Developer can was the first to break the news: In the model list of the Google Vertex AI platform, the identifier "claude-opus-4.8" clearly appeared.
As soon as the news came out, the whole network was in an uproar.
By the way, Anthropic's previous two major releases -
Opus 4.6 and Opus 4.7 were both discovered in advance in the backend of Vertex AI and then officially announced.
This means that Opus 4.8, which is already in testing, is not far from its official release, and many people expect to see it next month.
510,000 Lines of Source Code Reveal "Heavenly Secrets", Sonnet 4.8 Skips 4.7
The exposure of Opus 4.8 is exciting, but the information about Sonnet 4.8 was actually leaked earlier.
As early as March 31st, when Anthropic pushed an npm update (version v2.1.88) for Claude Code, it made a suffocatingly low-level mistake -
Someone forgot to add the line ".map" to the .npmignore file.
The absence of this one-line configuration caused a 59.8MB source map containing 512,000 lines of TypeScript code and 1,900 internal files to be completely pushed to the npm public repository.
Boris Cherny, the father of Claude Code, confirmed afterwards that this was a "common developer error".
But the amount of information exposed by this "common mistake" is the largest internal leak in Anthropic's history.
And in the leaked code, the key information was confirmed -
In the unreleased keyword filter, references to Sonnet 4.8 and Opus 4.7 appeared, but there was no trace of Sonnet 4.7.
This directly confirms one thing: Anthropic plans to skip Sonnet 4.7 and go straight to 4.8.
According to the leaked information and community analysis, Sonnet 4.8 is expected to bring four major upgrades:
Soaring Visual Ability
Inheriting the visual upgrade of Opus 4.7, the recognition accuracy of UI mockups and complex architecture diagrams is expected to exceed 98%.
Opus 4.7 has already achieved a visual accuracy of 98.5%, and Sonnet 4.8 will bring this ability to a more affordable model level.
Significant Improvement in Programming Ability
Cleaner one-time code generation and more precise "instruction following".
New "X high" Reasoning Level
A new reasoning intensity setting that can enhance logical reasoning ability without significantly increasing the generation time.
Updated Tokenizer
The bad news is that the token consumption will increase by about 30%. For the same prompt, the new tokenizer will use more tokens.
Can Mythos 1 Be Used?
Anthropic Suddenly Changes Its Tune
Even more powerful than Claude 4.8 is that the legendary "too dangerous" Claude Mythos is really coming!
Just yesterday, Anthropic released its first report, revealing all the results of Claude Mythos in the past month -
In key software, 100 million high-risk vulnerabilities were all detected.
Just one day later, the AI testing tracking platform TestingCatalog reported:
Some users briefly saw the "Mythos 1" model option in the Claude interface.
Although it disappeared quickly, the new strings in the source code leaked key information: Claude Code, Claude Security.
This means that Anthropic is upgrading Mythos from a restricted security research tool to a professional product line for developers.
The model identifier is "claude-mythos-1-preview", clearly pointing to two major application scenarios: code generation and security.
Previously, Anthropic's stance was always: Mythos is too dangerous and will not be publicly released.
But in the latest Project Glasswing update, Anthropic's attitude is subtly changing, with a thought-provoking statement -
Looking forward to a general release
This change in wording means that it's only a matter of time before Mythos reaches the general public.
Meanwhile, the productization work of Claude Security is also accelerating.
A brand-new security dashboard is being built, which will display discovered vulnerabilities, 7-day and 30-day historical charts, and more in-depth classification analysis results. Currently, it is only available to enterprise customers.
This means that Anthropic is transforming Mythos from a "research preview" into a commercializable security product matrix:
- Claude Code + Mythos: A security programming assistant for developers
- Claude Security + Mythos: An automated vulnerability discovery and repair platform for enterprises
Both code generation and code security need to be grasped firmly.
Fighting on Three Fronts, the ASI Competition Accelerates
This wave of consecutive exposures has completely torn off the last veil of kindness in the AI competition.
Advancing on three fronts and going all out.
Anthropic's full hand of cards this time sends an extremely clear signal to the entire industry:
The speed of AI evolution has completely deviated from the linear narrative.
What's most terrifying is the commercialized Mythos 1.
As a former "security no-go zone", its lifting of restrictions means that AI has the ability to subvert the offensive and defensive pattern of the online world.
This is precisely the inevitable path towards the ASI stage.
The birth of superintelligence is not just about piling up computing power and parameters, but about whether it can have absolute self-discipline and security evolution ability while having destructive power.
Anthropic is using the double-helix structure of "code + security" to fasten the last safety lock for the violent aesthetics leading to ASI.
With GPP-5.6 and Gemini 3.5 Pro pressing hard, in this battle of the gods, who can be the first to open the door to ASI?
Reference Materials:
https://x.com/kimmonismus/status/2058226072596971694
https://x.com/pankajkumar_dev/status/2057832457655959664
https://www.testingcatalog.com/anthropic-prepares-mythos-1-for-claude-code-and-claude-security/
This article is from the WeChat official account "New Intelligence Yuan", author: ASI Revelation, editor: Peach David. Republished by 36Kr with permission.