Reported by Synced
Editors: Aeneas, Hao Kun
[Synced Briefing] Just now, Google's internal AI has been exposed! Dubbed "Agent Smith," the tool exploded in popularity the moment it launched, with employees swarming to use it and crashing the servers. Now, secret new models from the "Big Three" tech giants have all surfaced simultaneously. Let the global AI war rage on!
Is The Matrix coming true at Google?
Today, Google's own cutting-edge internal model has been revealed.
Breaking news from Business Insider reports that this AI tool, named Agent Smith, has become wildly popular within the company!
It can automatically write code, call internal systems, and complete various daily tasks.
The most outrageous part? You don't even need to turn on your computer—just pull out your phone, and you can remotely command it to work!
Yes, it's named after that classic villain from The Matrix, the one in the black trench coat who could replicate himself infinitely.
Google giving its AI tool this name carries a touch of dark humor: "If you can't beat them, join them."
Just how hot is Agent Smith? In a word, it's so popular that Google had to impose limits.
According to insiders, the moment Agent Smith went live internally, employees flocked to it, completely overwhelming the servers!
Consequently, Google had to urgently tighten access permissions.
This scenario of "a company's own tool crashing its own servers" might be an accident elsewhere, but at Google, it's a true microcosm of the 2026 AI race.
Interestingly, a professor from the Wharton School couldn't help but吐槽 (roast) the names of these recently leaked models:
"I beg you labs, the only thing worse than naming a model GPT-5.5-xhigh-Codex-nano is naming it Agent Smith or Mythos!"
Because the name "Mythos" sounds way too Cthulhu-esque.
And the name "Agent Smith" was probably chosen by some Google employee who just watched The Matrix...
In comparison, OpenAI's taxonomic naming convention seems terrifyingly serious.
How Powerful Is Agent Smith, Really?
According to foreign media reports, Agent Smith is built upon Google's Antigravity platform, launched last November.
Antigravity itself is already an "Agent-first" development platform, supporting AI in autonomously planning, executing, and verifying complete coding tasks.
Agent Smith takes this a step further.
It's not just a coding assistant; it's a digital employee capable of operating independently.
It is understood that Agent Smith has several key capabilities that made it an instant hit within Google:
First, Asynchronous Operation. You assign it a task, close your laptop, and leave. It quietly finishes the job in the background. No need to stare at the screen waiting for results.
Second, Mobile Control. Through Google's internal chat platform, employees can directly issue commands to Agent Smith and check progress via their phones. Whether on the commute or during lunch, you can command the AI to work anytime, anywhere.
Third, Deep Integration with Internal Systems. It can retrieve employee profiles and pull internal documents, saving the massive amount of time previously spent on manual searching.
Fourth, Extremely High Autonomy. Compared to Google's previous AI coding assistants, Agent Smith can plan task flows, execute them, and verify results on its own. Engineers only need to audit at key nodes.
According to Google employees, for some software engineers, Agent Smith has become a true productivity multiplier.
Currently, over 100 developers are involved in expanding Agent Smith.
It can not only generate code but also simulate user scenarios, automatically test for software bugs, and directly feed discovered vulnerabilities back to the corresponding teams.
Google's Strategic Bet
The explosion of Agent Smith is no accident.
Behind it lies a company-wide AI Agent strategy from top to bottom at Google.
Sergey Brin, Google's co-founder, returned from semi-retirement in 2023 and now spends almost every day frantically coding on the front lines.
Hassabis, head of DeepMind, personally confirmed that Brin has been deeply involved in technical development, "coding personally."
In early March this year, Brin attended an internal all-hands meeting and clearly stated: AI Agents will be Google's most important direction this year.
Brin hinted at the meeting that Google is developing a tool similar to "OpenClaw." Amusingly, employees in the audience were whispering: "Isn't he just talking about Smith?"
An even more interesting moment occurred at the meeting: Google business executive Philipp Schindler joked that he can now instantly tell whether the person replying to messages opposite him is Brin himself or Brin's Agent.
This signals that the future Google might be a massive-scale hive model coordinated by tens of thousands of AI Agents!
On the execution level, Google's attitude is even more radical.
CEO Sundar Pichai told employees last year that they must embrace AI because competitors are doing the same.
By 2026, this advice has turned into a hard requirement: The usage of AI tools has been included in annual performance reviews!
Furthermore, Google internally launched "Project EAT," driving the standardized application of AI from the bottom up.
"No AI usage, no promotion or raise"—this is the real atmosphere in Silicon Valley in 2026.
Three Explosions in One Week
If Agent Smith is just Google's internal story, what follows is a collective earthquake for the entire AI industry.
In the same week Agent Smith was exposed:
Anthropic, due to a "configuration error," accidentally leaked all details of its strongest model, "Claude Mythos."
OpenAI cut its flagship product Sora to go all-in on a mysterious model codenamed "Spud."
Three codenames, three bombs, all detonated within one week.
Anthropic: Strongest Model Leaked, Dare Not Release It Themselves
First, the situation at Anthropic is extremely dramatic.
Security researchers discovered that the content management system behind Anthropic's official blog, due to a configuration error, left the door wide open to nearly 3,000 internal files—accessible by anyone.
Fortune magazine retrieved a blog draft from the leak, the content of which is enough to make the entire AI circle explode:
Anthropic has completed a new model named "Claude Mythos."
Currently, Claude has three tiers: the strongest Opus, the mid-range Sonnet, and the lightest Haiku.
Mythos belongs to "Capybara," a brand new fourth tier above Opus—larger, stronger, and more expensive.
In terms of performance, compared to the previously strongest Claude Opus 4.6, Capybara scored significantly higher in software coding, academic reasoning, and cybersecurity tests.
Anthropic subsequently stopped pretending, directly admitting: "This model is real, and it is the strongest thing we have ever built," describing its capabilities as a "leapfrog" jump.
However, Mythos has not been released yet for one reason only: It is too dangerous!
In the leaked files, Anthropic explicitly warned that Mythos is far ahead of any global model in cyberattack capabilities. It can discover and exploit software vulnerabilities at a speed far exceeding the response capacity of defense personnel.
After reviewing the leaked content, Stifel analyst Adam Borg stated bluntly: "Mythos is obviously an order of magnitude more powerful than any frontier model on the market, and its compute consumption is extremely high."
"We believe this could become the ultimate hacker tool."
Coincidentally, on the same day Mythos was leaked, Anthropic was also reported to potentially start an IPO before October 2026.
The existence of Mythos undoubtedly adds a massive weight to the IPO valuation table.
Slide up and down to view
Slide up and down to view
OpenAI: Cutting Sora, All-in on "Spud"
Now looking at OpenAI.
According to The Information: OpenAI has completed pre-training for a new model, internally codenamed "Spud."
Sam Altman told employees internally that this model can "truly accelerate the economy."
Spud's specific parameters have not been made public, nor has its specific performance been announced.
However, employees revealed that Spud possesses an "unprecedented new capability."
Rumors suggest it might be a native multimodal architecture, processing text, audio, and images uniformly from the ground up, rather than stitching together different components like existing models.
To free up compute power for Spud, OpenAI made a decision that shocked the entire industry:
Cutting Sora.
Sora, which launched with great fanfare last September, signed a $1 billion deal with Disney, and was once seen as the benchmark for the AI video track, was cut just like that.
Absurdly, on the very Monday Sora's cancellation was announced, the Disney team was still in a meeting with OpenAI discussing the Sora project.
Who would have thought that just 30 minutes after the meeting adjourned, Disney received the notification: Sora is gone.
The $1 billion partnership, licensing for over 200 Disney characters, Mickey Mouse, Cinderella, Star Wars... all vanished.
The reason is direct: It burns too much money.
Reports state Sora was losing about $15 million daily, with download numbers plummeting consecutively since December 2025.
OpenAI's Application CEO Fidji Simo told employees the company was "spreading itself too thin across too many Apps" and could no longer engage in "side quests."
The GPU power saved by cutting Sora is being fully invested into Spud and the "Super App" currently in development—a unified platform integrating ChatGPT, Codex, and a browser.
Meanwhile, OpenAI's product team has quietly renamed itself "AGI Deployment."
This isn't just a name change; it's a declaration to the world: We believe we have entered the stage of AGI deployment.
AI Giants All-In on the Same Table
Agent Smith, Spud, Mythos—three distinct codenames pointing in the same direction:
The paradigm shift from "Chatbots" to "Autonomous Agents."
Google's Agent Smith transforms AI from a passive tool that waits for your prompt into a digital employee capable of operating independently in the background and proactively completing tasks.
OpenAI going all-in on Spud and the Super App is betting on an all-capability Agent under a unified entry point—chatting, coding, browsing, and executing, all completed in one window.
Anthropic's Mythos pushes model capabilities to a height that even scares themselves, to the point where they have to let the "defense" run first before considering opening it up.
Now, the stakes in this game have far exceeded the scope of "whose model scores higher."
AI companies are betting everything. What happens after they win? They haven't figured that out yet.
But regardless, All-In first!
References:
https://x.com/emollick/status/2037565418970185786
https://www.businessinsider.com/google-agent-smith-employees-ai-driven-coding-2026-3