PALO ALTO, CA — Chinese AI startup DeepSeek unveiled its fourth major model on Friday, promising dramatic improvements in reasoning and agentic capabilities, prompting what multiple sources describe as “a very quiet but very real panic” spreading through Silicon Valley’s open-plan offices like a silent, well-ventilated fog.

The new model, DeepSeek V4, features a 1-million-token context window, a novel Hybrid Attention Architecture, and the ability to autonomously write and deploy code — capabilities that several senior engineers at competing US labs described as “fine,” “completely fine,” and “I’m totally fine.”

“We’ve been saying for two years that these things can’t reason,” said one researcher at a major AI lab, who asked not to be named because he is actively updating his CV. “And technically that’s still true. They just do something indistinguishable from reasoning and the market has stopped caring about the distinction.”

What V4 Can Do

According to DeepSeek’s release notes, the new model is capable of:

  • Autonomously writing, testing, and deploying production code from a single natural-language prompt
  • Maintaining coherent context across conversations the length of a short novel
  • Completing “complex multi-step agentic tasks,” which DeepSeek defines as anything currently billed at $340,000 per year in the Bay Area

The model is also, according to its benchmarks, the highest-performing publicly available AI in fourteen separate categories, which DeepSeek says is “a coincidence” and OpenAI says is “a different set of categories than the ones we use.”

Industry Response

The release triggered the by-now-familiar Silicon Valley ritual: a brief period of denial, followed by an emergency all-hands described in the calendar invite as a “sync,” followed by a company blog post about how this is “good for the ecosystem.”

NVIDIA’s stock ticked up 4% on the news, on the logic that someone will need to run all of this.

Several venture capital firms issued statements expressing confidence in their portfolio companies. One VC, who asked to remain anonymous, said he had spent most of Friday afternoon staring at a Google Doc titled “Differentiation Strategy 2026 DRAFT v7” and had not yet typed anything into it.

“What we have in the US,” he said, after a long pause, “is vibes. And I think vibes still matter.”

The Human Angle

Recruiters in New York and Chicago report a sharp uptick in unsolicited LinkedIn connection requests from Bay Area software engineers this week, with many profiles quietly removing phrases like “building the future of AI” and replacing them with “open to opportunities.”

One engineer at a major model lab confirmed he had spent “a couple of hours, just looking” at graduate programs in accounting. “Not seriously,” he clarified. “I was just curious what the options are. I have options. I’m a person with options.”

His current employer has not yet announced any layoffs. His employer’s employer has not yet announced any layoffs. His employer’s employer’s employer — a sovereign wealth fund in Singapore — declined to comment.

Context

DeepSeek’s first model caused a significant market reaction in early 2025. The second model caused a slightly smaller one. The third caused what analysts described as “concern.” The fourth has caused what analysts describe as “a Tuesday.”

“At some point,” said Dr. Margaux Ellison, a professor of Labour Economics at UC Berkeley, “you stop being disrupted and you start being the new normal. I think we crossed that line sometime around model three. Now it’s just how things are.”

She paused, then added: “I’ve also been looking at the accounting programs. Just to see.”

At press time, DeepSeek had released a brief follow-up statement noting that V4 had written most of its own documentation, which it described as “a feature” and which three US AI labs’ legal teams had begun examining to determine if it was also “a threat.”

CCNN reached out to OpenAI, Google DeepMind, and Anthropic for comment. All three responded within minutes with nearly identical statements, which their respective PR teams insist were written by humans.