Claude Code: Anthropic’s $2.5B Lesson in Killing Bad Ideas
Most companies don’t have an innovation problem. They have a cowardice problem: nobody is allowed to kill a bad idea fast enough.
Most companies don’t have an innovation problem. They have a cowardice problem: nobody is allowed to kill a bad idea fast enough.
Anthropic has built a roughly 20-person team around the opposite idea. Ben Mann’s Labs group expects most product bets to die — and one of its biggest successes, Claude Code, is now running at more than $2.5 billion in annualised revenue. That is not a creativity story. It is a management story.
Anthropic has made failure part of the job
The interesting leadership news out of Anthropic is not another model release or another silly valuation headline. It is the company’s operating system for turning frontier technology into products people will actually pay for.
Business Insider reported this week that Anthropic Labs is a rotating group of about 20 people led by co-founder Ben Mann. The team operates like an internal startup: it tests product ideas in short cycles, reviews them roughly every two weeks, kills or reshapes the weak ones, and hands the winners to standalone teams once they have real momentum.
Mann’s stated success rate is only around 20% to 30%. Good. That is what a real experimental function should look like.
If every internal project succeeds, you have not built an innovation engine. You have built a committee that only approves safe bets after the answer is obvious. That might preserve the careers of middle managers, but it will not build the next category-defining product.
Claude Code is the proof point. Anthropic said in January that the product went from research preview to a billion-dollar product in six months. By March, the company said its run-rate revenue had grown beyond $2.5 billion, more than doubling since the start of 2026.
That is a serious commercial outcome from an idea that began as an experiment, not a five-year planning deck with 47 stakeholders and a bloke from finance asking whether the button should be blue.
Ben Mann and Mike Krieger changed the structure, not just the org chart
Anthropic expanded Labs in January 2026. Mike Krieger, Instagram’s co-founder and Anthropic’s chief product officer at the time, moved into Labs to build alongside Mann. Ami Vora took the product leadership role, working with CTO Rahul Patil on scaling the products that had already earned their place in the market.
That split is worth studying.
Most businesses mash two completely different jobs together: discovering the next thing and operating the current thing. Then they act surprised when neither gets done particularly well.
Discovery needs speed, technical proximity, tolerance for ambiguity and a licence to look foolish for a while. Scaling needs reliability, customer support, security, commercial discipline, hiring, documentation and the boring machinery that makes customers trust you with their money.
Put the same executive team in charge of both, measured by the same quarterly targets, and the scaling work wins every time. It is measurable. It is safer. It can be explained in a board meeting without anyone breaking into a sweat.
Anthropic’s answer is structural separation. Labs gets room to explore what new model capabilities can make possible. The core product organisation gets the job of making the good stuff dependable for millions of users and enterprise customers.
This is not glamorous. It is simply adult management.
The company has also made a crucial decision that plenty of founders miss: successful experiments should graduate. They should not remain pets of the innovation team forever, protected from commercial reality because the founder likes them.
A prototype is not a product. A product is not a business. And a business is not a durable advantage until it can survive handover from the people who dreamed it up.
The $2.5B number changes the conversation
A $2.5 billion run-rate product does more than make for a nice press release. It changes the power dynamics inside a company.
Claude Code gives Anthropic a commercial engine tied directly to a high-frequency, painful customer problem: software development. The more useful it becomes, the more deeply it can embed itself in the work of engineering teams. That is a far stronger position than being merely another clever chatbot people use when they are bored in a meeting.
It also explains why Labs matters as Anthropic approaches a potential public listing. Reuters reported last week that the company was expected to make its IPO prospectus public later in September, with marketing potentially beginning in mid-October and a listing targeted for just before the November U.S. midterm elections. Those plans can change, obviously. IPO calendars are written in pencil until they are not.
But public investors will ask a brutal and fair question: can Anthropic keep making new revenue engines, or did it get lucky once?
A small, repeatable product-incubation machine is part of the answer. Claude Code is the trophy. The operating model behind it is the asset.
Anthropic has also pushed the Model Context Protocol, or MCP, into the market as an open standard for connecting AI systems to tools and data. In July, Anthropic said MCP had surpassed 400 million monthly SDK downloads. You do not need to be an AI tragic to understand why that matters. Products become much stickier when they sit inside the plumbing rather than merely on top of it.
That said, founders should avoid the lazy conclusion that copying Anthropic’s “Labs” label will produce Anthropic’s outcomes. Slapping a neon sign saying INNOVATION LAB on a spare meeting room is usually corporate cosplay.
The model works only if the team has three things: access to genuine technical change, authority to kill work without political theatre, and a clean path for winners to receive serious resources.
Without those, it is just an expensive arts-and-crafts program.
The overlooked angle: this is a talent-retention machine
The obvious reading is that Labs exists to invent products. The more useful reading is that it also exists to keep exceptional builders from leaving.
The best product people, engineers and researchers do not join a fast-moving company to spend two years maintaining a roadmap created by somebody who left last summer. They want proximity to important problems, quick feedback and the chance to see their work reach customers.
A rotating internal-startup model gives them that without forcing them to quit, raise venture capital, hire a team and spend half their lives begging for cloud credits.
That is especially important in AI, where the underlying capability is moving too quickly for a rigid annual planning cycle. If the model improves sharply, a product opportunity can appear in weeks. A company that needs six layers of approval to test it has already handed the advantage to somebody faster.
But there is a trade-off. Labs can become a prestige club. The core teams can start feeling like maintenance crews while the shiny people get to play with the future. That kills morale and creates a two-class company.
Leadership has to manage that deliberately. The graduation path matters because it brings successful work into the main business. So does rotation: builders need to move in and out, and the people scaling products need status, rewards and influence equal to those inventing them.
If you run a company, do not romanticise experimentation while treating operators as second-class citizens. The operator who turns a scrappy prototype into reliable revenue is not the clean-up crew. They are the reason you get paid.
Don’t copy the team size — copy the rules
Twenty people is not magic. For most businesses, 20 people would be absurdly large. If you run a 35-person company, a 20-person lab is not innovation; it is bankruptcy with beanbags.
The lesson is not to carve out a big separate department. The lesson is to establish hard rules around small bets.
First, give one person clear ownership of each experiment. Not a steering committee. Not a working group. One accountable adult.
Second, define the evidence that will make you continue, change direction or stop. Before the work begins. If you decide the rules after you have fallen in love with the project, you are not evaluating it; you are defending it.
Third, use short review cycles. Anthropic’s roughly fortnightly rhythm is aggressive, but the principle is dead right. A bad idea rarely becomes good because you let it consume another six months of payroll.
Fourth, separate an experiment’s success from the ego of the person running it. A killed project should not mean a failed employee. If it does, everyone will keep dead projects alive and call it resilience.
Finally, make graduation explicit. If an experiment wins, who funds it? Who owns it? When does it stop being experimental? If you cannot answer those questions, you are collecting prototypes, not creating businesses.
What this means for you
Here is something you can use tomorrow: take your three most expensive “strategic initiatives” and ask one question of each — what evidence would make us stop?
If nobody can answer within five minutes, you do not have a strategy. You have a sunk-cost hobby funded by the company.
Then pick one customer problem you keep hearing about but have not solved because it falls outside the current roadmap. Put two sharp people on it for two weeks. Give them a tight budget, direct customer access and permission to return with one of three answers: build more, change it, or kill it.
No slide deck. No theatre. No fake certainty.
Anthropic’s 20-person Labs team matters because it demonstrates a truth founders and executives would rather avoid: speed is not created by telling people to work harder. It is created by designing a company where good ideas can move and bad ideas can die.
The businesses that win the next decade will not be the ones with the most innovation workshops. They will be the ones with the stomach to stop wasting money before everyone else admits the idea was rubbish.