At HumanX in San Francisco, the room felt electric. And at the center of it all was Anthropic – arguably the hottest company in tech right now – represented by none other than Mike Krieger, Instagram co-founder turned Anthropic Labs co-lead. Moderated by Bloomberg’s Ed Ludlow, the session was part career retrospective, part product philosophy masterclass, and part rare window into how the world’s most talked-about AI lab actually operates day to day.
Anthropic’s momentum is staggering. Revenue has rocketed from a $9 billion run rate at the end of last year to $30 billion – a trajectory that has Wall Street, Silicon Valley, and every startup founder in between paying very close attention. And at the heart of that growth sits Anthropic Labs, an internal incubator governed by a deliberately ruthless two-week review cycle. Every project gets scrutinized – progress or die. No pet projects, no bloated roadmaps quietly consuming engineering bandwidth. “The opportunity cost of working even two extra weeks on the wrong project is pretty high,” Krieger said. It’s this discipline that birthed Claude Code, a product that reached a billion dollars in revenue within just six months of launch, and the Model Context Protocol, now an open standard being adopted across the entire industry.
What makes Anthropic’s rise even more remarkable is that it’s being built on a foundation of productive failure as much as breakout success. Early computer use prototypes were broken and unshippable – but they created a testing harness that meant every new model got immediately benchmarked against real tasks. When Claude 3.7 finally cracked it, the team knew instantly. That’s not luck. That’s infrastructure for serendipity, and it’s a flywheel that keeps accelerating.
Then there’s “Mythos” – the news that sent ripples through the HumanX crowd. Anthropic has developed a new model so powerful that it dramatically outperforms Opus on both coding and offensive cybersecurity benchmarks. Yet rather than releasing it to the public and capitalizing on the hype, Anthropic is quietly deploying it first to a curated group of cybersecurity researchers and open source maintainers. The deliberate, staged rollout is a direct and bold reflection of Anthropic’s safety-first identity – a company willing to leave money on the table in the short term to get the long game right. “It’s not the revenue-optimizing move,” Krieger acknowledged, “but it’s absolutely the right one.” In a landscape crowded with AI companies racing recklessly to ship, that kind of restraint is itself a competitive differentiator.
Internally, Anthropic’s culture of “ant-fooding” – their playful spin on dogfooding – means over 2,000 employees are continuously stress-testing products before public launch. The Claude Code team even runs a dedicated on-call rotation purely to process the flood of internal feedback. It’s product development at a pace and signal quality that most companies can only envy, and it helps explain why Anthropic keeps shipping products that actually stick.
But what’s consuming Krieger most right now is a deceptively simple, deeply human question: how do you extend genuine AI autonomy to everyone? He described his own nightly ritual of queuing Claude for eight-hour uninterrupted work sessions – waking up to completed tasks and a cheerful morning summary. Founders are rewriting entire software stacks using Claude’s agent SDK, slashing code volume dramatically while multiplying capability. Enterprises are deploying long-horizon agents running for 30 hours straight. The productivity unlock is real, extraordinary – and still largely inaccessible to most people.
That’s the frontier Anthropic is now racing toward. And with the hottest models, the fastest revenue growth, and a product culture unlike anything else in AI, they may just get there first.
“This is the worst the models will ever be,” Krieger said with quiet confidence. At HumanX, nobody in that room disagreed. For builders paying attention, that’s not a footnote. That’s the headline.
