Investor updates used to be the bane of Aakash Shah's month. Now, he says, it's "just an afternoon with coffee." Shah founded allergy tech startup Wyndly in 2020, and would spend hours weaving together data from disconnected sources. As a self-proclaimed "hobby coder," Shah, the company's CEO, found himself trying to program his own analysis system, to no avail. Then, earlier this year, Shah's lead engineer gave him a tip: download Claude Code, a new AI-powered software engineering tool from San Francisco-based startup Anthropic, and just ask it to conduct the analysis he needed.
Within minutes of typing his first prompt, Shah was fixing problems that had previously taken him hours, and messaging with the tool as if it were a human analyst. Stunned, he recalls thinking that Claude "fundamentally is doing something that was effectively impossible to do before."
Shah isn't alone. Over the past year, many entrepreneurs have tapped Claude, Anthropic's AI model, to write code and launch features, in effect using the AI assistant as a technical co-founder. As Claude has become ever more sophisticated, so have the tasks it can handle—spawning a whole new way of developing software known as vibe coding, a term coined by renowned AI scientist and OpenAI co-founder Andrej Karpathy. It's what happens when you describe what you want the AI to build in plain language and watch it write the code for you—no programming knowledge required.
Dario Amodei, Anthropic's CEO and co-founder, isn't the biggest fan of "vibe coding" as a term ("it's kind of woo-woo," he tells Inc.), but its impact on both Anthropic and the tech industry is undeniable—it has sparked the establishment of a new tech ecosystem. Now, AI coding companies such as StackBlitz, Lovable, Replit, and Cursor are some of Anthropic's highest-profile customers, and Claude is playing an integral role in their trajectories.
Anthropic will end 2025 having become one of the fastest-growing software companies in history. In less than two years, its annual run-rate revenue has exploded from $87 million to $7 billion as of October 2025, and the company projects it will hit $9 billion by year's end. Along the way, it has transformed the way many businesses develop code. To understand how the company got so far so fast, it helps to wind back the clock to December 2020, when Amodei and six colleagues at OpenAI had a Jerry Maguire moment.
The group that eventually launched Claude included some of the figures most instrumental in the creation of GPT-3, a breakthrough large language model that set off an explosion of interest in the artificial intelligence industry. For most machine learning researchers, shipping GPT-3 would be a career high point. But for these seven, led by Amodei, who was OpenAI's research head, it was just the beginning of a much more ambitious project.
While at OpenAI, Amodei's team researched what's known as scaling laws, a principle describing how AI models typically get better when trained on more computing power and data. But he felt that OpenAI's leaders weren't "particularly thoughtful" about responsibly scaling and deploying new AI models, and was unconvinced that the company "would do it right in the future."
So Amodei, along with six others— Daniela Amodei (his sister), Tom Brown, Jared Kaplan, Sam McCandlish, Chris Olah, and Jack Clark—decided to quit OpenAI. In May 2021, the founders unveiled Anthropic, a safety-focused AI lab that would build state-of-the-art models.
But Amodei had an even more specific priority: AI-powered coding. When he got his hands on GPT-3 in 2019, "one of the first things I noticed was that it could write code," he says—but not particularly well. Back then, the best GPT-3 could manage was a single line of Python, but Amodei saw a path for AI to graduate from writing lines of code to doing much more. "I was like, it's gonna go all the way," Amodei says. "This thing's gonna code better than humans in a few years."
Unlike OpenAI, which has found massive success offering ChatGPT as a personal companion, Anthropic is growing quickly on the strength of enterprise customers. As Amodei anticipated, Claude's coding prowess has proved lucrative.
So what makes Claude, which Anthropic introduced to the world in March 2023, so good at coding? Amodei will offer only a circular explanation, saying that it is "good at code because Anthropic researchers and engineers focus a lot on making it good at code."
Alex Albert, Anthropic's dev relations lead, posits a theory. "Code is the easiest domain to verify," he says, in part because the software development process itself is one big feedback loop; developers write code, run it, look at the errors they get back, and make adjustments.
Feedback loops are crucial in shaping Claude's personality, says Amanda Askell, a researcher often credited with developing Claude's character—which Anthropic describes as "helpful, honest, and harmless." Through a process called reinforcement learning with AI feedback, Claude gains personality traits by generating prompts, responding to them, and judging which responses best demonstrate the desired characteristics.
"What I find different about Claude is a kind of capacity for self-reflection and self-awareness," says Askell. Because of all this, in Askell's opinion, coding with Claude feels more like "play" than with other models. "If you take good character, good instruction-following, and good code," she says, "I think those all combine into something really nice."
In June 2024, Anthropic brought AI-powered coding to the next level with Claude Sonnet 3.5, a model that could develop entire small applications with just a single prompt. (Currently, Claude comes in three sizes—the small and cheap Haiku, the middle-level Sonnet, and the large and expensive Opus.)
Claude Sonnet 3.5 acted as rocket fuel for San Francisco-based StackBlitz. In 2024, co-founder Eric Simons was on his last legs. He and childhood friend Albert Pai launched the company in 2017 with a vision of developing a web-based platform for writing and running code, essentially "what Canva and Figma did for design, but for web development," Simons says. But getting developers to use that environment proved harder than expected. By mid-2024, they were nearly out of money.
Simons and Pai were preparing to wind down the company, but wanted to go out in a blaze of glory. Witnessing the rapid progress of AI models, the pair plotted one last pivot: creating a ChatGPT-like platform that enabled users to create software by asking an AI model to do it for them. After a false start with OpenAI, they got early access to Claude Sonnet 3.5.
As soon as Simons and Pai slotted the model into their coding platform, the results were obvious. "Holy heck," Simons remembers thinking, "this is way better than anything I've seen before. We can build a product around this."
In October 2024, Simons and Pai launched Bolt.new, a browser-based platform for agentic coding powered by Claude. (An AI model can act as an agent when it's given the tools it needs to take actions like running code.) It turned out that the virtual environments StackBlitz spent years perfecting were exactly what Claude needed to operate effectively, allowing it to write, execute, and iterate on code in real time.
Within four weeks, Bolt catapulted StackBlitz's annual recurring revenue from under $1 million to over $4 million. By the end of month two, the company's ARR had reached $20 million entirely from word-of-mouth growth. In January, the company announced that it had raised $105.5 million.
Today, StackBlitz's ARR is over $40 million, with much of that coming from nontechnical execs, product managers, and designers, all of whom use Bolt to develop software prototypes that can be fleshed out by an engineering team.
David Hershey, a member of Anthropic's applied AI team, says StackBlitz's overnight success was a "crazy inflection point." As part of the applied AI team, Hershey helps Anthropic's enterprise and business customers optimize their Claude-centric apps and features. StackBlitz's turnaround is Hershey's "favorite canonical example" of Claude's ability to solve seemingly impossible problems.
Claude can act as a force multiplier for small teams as well. Kevin Tamura, co-founder and CEO of sales outreach automation platform Strama, says Claude is an "integral part" of his two-person startup. Using Claude Code, Anthropic's own AI-powered coding tool, Tamura built a stable of agents, each of which specializes in a different aspect of the business, including a go-to-market agent and a communications agent.
"It's like I've got four or five different people working for me," Tamura, a former Salesforce account executive, says of his Claude-powered agents. Each of these agents has access to the relevant context to get its work done. Claude is a "critical member of our team," he adds.
Claude has also been helpful at legal AI unicorn Harvey, where it's adept at analyzing information from complicated credit agreements, according to Winston Weinberg, co-founder and CEO.
Anton Osika, founder and CEO of Sweden-based vibe-coding company Lovable, likens Claude to "the quietest person in the room, but the one you trust most when things go sideways." (Not everyone trusts Anthropic entirely: In September, the company agreed to pay $1.5 billion to settle a lawsuit from a group of authors and publishers after a judge ruled that the company downloaded more than seven million books from illegal file-sharing sites for training data.)
Still, Osika says Claude Sonnet 3.5 was the first model to be both good and fast enough to put nontechnical users in what he calls a "vibe-coding loop." In this loop, he says, "you're in the product, making changes, reacting in real time. It feels more like jamming with a creative partner than giving instructions to a machine."
With Claude's help, Lovable scaled rapidly in 2025, and in November passed $200 million in ARR. Amodei says he didn't specifically predict that vibe coding would take off as an industry, but was confident that AI models would quickly become roughly as capable as an "eager intern," opening up a world of possibilities.
But Claude's long-term dominance in the coding space isn't assured, and it doesn't come cheap, either. Companies like Anthropic and OpenAI charge for direct access to their tech via API fees, and in practice, these APIs function like a utility that bills based on how much you use a resource. That charge is determined by the number of tokens processed and generated by a model.
Tokens are measurable units of computational work that add up as an AI model reads and writes text. Claude Sonnet 4.5, Anthropic's most recent midsize model, costs $15 for every million tokens generated. GPT-5, OpenAI's current flagship model, is much cheaper in comparison, at $10 per million tokens generated.
This makes Claude more expensive than the competition, especially in token-heavy use cases such as coding. "A significant part" of the money paid by Lovable's customers goes directly toward paying those API fees, confirms Osika via email, "but we're OK with that for now. The real priority is giving our users value over the long term."
Osika expects the cost of frontier AI to slowly come down (OpenAI cut the price of one of its models by 80 percent in June), but Michele Catasta, president and head of AI at Replit, doesn't believe state-of-the-art AI will be getting cheaper anytime soon.
Replit was founded in 2016, and for its first eight years, had a small, dedicated user base of software engineers. But in 2024, it pivoted to focus on nontechnical customers. Claude Sonnet 3.5 enabled the creation of Replit Agent, a vibe-coding tool that skyrocketed Replit's ARR from $2.8 million to over $250 million.
Anthropic has largely resisted cutting prices. For most of 2025, the company focused its efforts on releasing ever-better Claudes at its three price points. The company's theory is that over time large, state-of-the-art Claudes will pass their capabilities onto smaller, cheaper Claudes. But in November, the company released its latest large-size model, Claude Opus 4.5, which is 67 percent cheaper than the previous Opus model.
Despite cost concerns, vibe-coding companies keep doubling down on Claude. StackBlitz, Lovable, and Replit have all upgraded their platforms to primarily use Claude Sonnet 4.5, and have also integrated the larger Claude Opus 4.5. As these vibe-coding startups explode, they fuel ever more growth at Anthropic. Claude Code has to date generated over $500 million in run-rate revenue and is hugely popular with software developers, startups, and tech companies. For entrepreneurs thinking of building a vibe-coding platform, Amodei has a key piece of advice: "Find something that's complementary to the models, something that the models a year from now won't have."
Of course, Anthropic's rivals aren't sitting still. In April, OpenAI launched Codex, its own take on a Claude Code-like software engineering agent. Since then, OpenAI has been relentlessly improving the coding abilities of its models; the latest, GPT-5.1, recently took the top spot on a new vibe-coding benchmark developed by startup Vals AI.
Wyndly CEO Shah says that while he is sticking with Claude, friends who serve as enterprise technical consultants believe OpenAI is "slowly inching out" Anthropic. But benchmarks may not tell the whole story. The vibes of working with a specific platform matter, too. Shah observes that the future of AI coding could be more about individual user preference than sheer capability. Claude appeared first and won loyal users. "Once someone has something that works," Shah says, "they're not going to switch."
Read on Inc.com →