Anthropic’s stealth enterprise coup: How Claude 3.7 is becoming the coding agent of choice

Whereas client consideration has centered on the generative AI battles between OpenAI and Google, Anthropic has executed a disciplined enterprise technique centered on coding — doubtlessly probably the most priceless enterprise AI use case. The outcomes have gotten more and more clear: Claude is positioning itself because the LLM that issues most for companies.

The proof? Anthropic’s Claude 3.7 Sonnet, launched simply two weeks in the past, set new benchmark information for coding efficiency. Concurrently, the corporate launched Claude Code, a command-line AI agent that helps builders construct functions sooner. In the meantime, Cursor — an AI-powered code editor that defaults to Anthropic’s Claude mannequin — has surged to a reported $100 million in annual recurring income in simply 12 months. 

Anthropic’s deliberate concentrate on coding comes as enterprises more and more acknowledge the ability of AI coding brokers, which allow each seasoned builders and non-coders to construct functions with unprecedented velocity and effectivity. “Anthropic continues to come back out on prime,” stated Guillermo Rauch, CEO of Vercel, one other fast-growing firm that lets builders, together with non-coders, deploy front-end functions. Final 12 months, Vercel switched its lead coding mannequin from OpenAI’s GPT to Anthropic’s Claude after evaluating the fashions’ efficiency on key coding duties.

Claude 3.7: Setting new benchmarks for AI coding

Launched February 24, Claude 3.7 Sonnet leads on practically all coding benchmarks. It scored a formidable 70.3% on the revered SWE-bench benchmark, which measures an agent’s software program improvement expertise, handily outperforming nearest opponents OpenAI’s o1 (48.9%) and DeepSeek-R1 (49.2%). It additionally outperforms opponents on agentic duties.

Developer communities have shortly verified these ends in real-world testing. Reddit threads evaluating Claude 3.7 with Grok 3, the newly launched mannequin from Elon Musk’s xAI, constantly favor Anthropic’s mannequin for coding duties. “Primarily based on what I’ve examined, Claude 3.7 appears to be the very best for writing code (no less than for me),” stated a prime commenter. (Replace: Even Manus, the brand new Chinese language multi-purpose agent that took the world by storm earlier this week, when it launched saying it was higher than Open AI’s Deep Analysis and different autonomous duties, was largely constructed on Claude.)

Alongside the three.7 Sonnet launch, Anthropic launched Claude Code, an AI coding agent that works immediately by means of the command line. This enhances the corporate’s October launch of Pc Use, which permits Claude to work together with a person’s pc, together with utilizing a browser to look the online, opening functions, and inputting textual content.

Most notable is what Anthropic hasn’t performed. In contrast to opponents that rush to match one another feature-for-feature, the corporate hasn’t even bothered to combine internet search performance into its app — a fundamental characteristic most customers anticipate. This calculated omission indicators that Anthropic isn’t competing for basic shoppers however is laser-focused on the enterprise market, the place coding capabilities ship a lot increased ROI than search.

Palms-on with Claude’s coding capabilities

To check the real-world capabilities of those coding brokers, I experimented with constructing a database to retailer VentureBeat articles utilizing three totally different approaches: Claude 3.7 Sonnet by means of Anthropic’s app; Cursor’s coding agent; and Claude Code.

Utilizing Claude 3.7 immediately by means of Anthropic’s app, I discovered the answer supplied outstanding steerage for a non-coder like myself. It really useful a number of choices, from very strong options utilizing issues like PostgreSQL database, to simpler, light-weight ones like utilizing Airtable. I selected the light-weight answer, and Claude methodically walked me by means of methods to pull articles from the VentureBeat API into Airtable utilizing Make.com for connections. The method took about two hours, together with some authentication challenges, however resulted in a useful system. You might say that as an alternative of doing the entire code for me, it confirmed me a grasp plan on how to do it.

Cursor, which defaults to Claude’s fashions, is a full-fledged code editor and was extra wanting to automate the method. Nonetheless, it required permission at each step, making a considerably tedious workflow.

Claude Code supplied yet one more strategy, operating immediately within the terminal and utilizing SQLite to create an area database that pulled articles from our RSS feed. This answer was less complicated and extra dependable by way of getting me to my finish purpose, however undoubtedly much less strong and feature-rich than the Airtable implementation. I’m now understanding the character of those tradeoffs, and know that the coding agent I decide actually is determined by the precise mission.

The important thing perception: Whilst a non-developer, I used to be in a position to construct useful database functions utilizing all three approaches — one thing that will have been unthinkable only a 12 months in the past. They usually all relied on Claude underneath the hood.

For a extra detailed evaluate of how to do that so-called “vibe coding,” the place you depend on brokers to code issues whereas not doing any coding your self, learn this nice piece by developer Simon Willison printed yesterday. The method may be very buggy, and irritating at occasions, however with the correct concessions to this, you’ll be able to go a good distance.

The technique: Why coding is Anthropic’s enterprise play

Anthropic’s singular concentrate on coding capabilities isn’t unintended. In keeping with projections reportedly leaked to The Info, Anthropic goals to achieve $34.5 billion in income by 2027 — an 86-fold enhance from present ranges. Roughly 67% of this projected income would come from API enterprise, with enterprise coding functions as the first driver. Whereas Anthropic hasn’t launched actual numbers for its income up to now, it stated its coding income surged 1,000% during the last quarter of 2024. Final week, Anthropic introduced it had raised $3.5 billion extra in funding at a $61.5 billion valuation.

This coding guess is supported by Anthropic’s personal Financial Index, which discovered that 37.2% of queries despatched to Claude had been within the “pc and mathematical” class, primarily overlaying software program engineering duties like code modification, debugging and community troubleshooting.

Anthropic seems to be marching to its personal beat — at a time when opponents are distracted, speeding to cowl each enterprise and client markets with characteristic parity. OpenAI’s lead is bolstered from its early client recognition and utilization, and it’s caught attempting to serve each common customers and companies with a number of fashions and performance. Google is chasing this development too, attempting to have considered one of all the pieces.

Anthropic’s comparatively disciplined technique extends to its product choices. As an alternative of chasing client market share, the corporate has prioritized enterprise options like GitHub integration, audit logs, customizable permissions and domain-specific safety controls. Six months in the past, it launched a large 500,000-token context window for builders, whereas Google restricted its 1-million-token window to personal testers. The result’s a complete coding-focused providing that enterprises are more and more adopting.

The corporate not too long ago launched options permitting non-coders to publish AI-created functions inside their organizations, and simply final week upgraded its console with enhanced collaboration capabilities, together with shareable prompts and templates. This democratization displays a kind of Trojan Horse technique: First allow builders to construct highly effective foundations, then develop entry to the broader enterprise workforce, together with up into the company suite.

The coding agent ecosystem: Cursor and past

Maybe probably the most telling signal of Anthropic’s success is the explosive progress of Cursor, an AI code editor that reportedly has 360,000 customers, with greater than 40,000 of them paying clients, after simply 12 months — making it probably the quickest SaaS firm to achieve that milestone.

Cursor’s success is inextricably linked to Claude. “You’ve obtained to assume their primary buyer is Cursor,” famous Sam Witteveen, cofounder of Purple Dragon, an unbiased developer of AI brokers. “Most individuals on [Cursor] had been utilizing the Claude Sonnet mannequin — the three.5 fashions — already. And now it appears everybody’s simply migrating over to three.7.”

The connection between Anthropic and its ecosystem extends past particular person firms like Cursor. In November, Anthropic launched its Mannequin Context Protocol (MCP) as an open customary, permitting builders to construct instruments that work together with Claude fashions. The usual is being broadly adopted by builders.

“By launching this as an open protocol, they’re kind of saying, ‘Hey, everybody, have at it,’” defined Witteveen. “You may develop no matter you need that matches this protocol. We’re going to help this protocol.”

This strategy creates a virtuous cycle: Builders construct instruments for Claude, which makes Claude extra priceless to enterprises, which drives extra adoption, which attracts extra builders.

The competitors: Microsoft, OpenAI, Google and open supply

Whereas Anthropic has discovered its focus, opponents are pursuing totally different methods with various outcomes.

Microsoft maintains vital momentum by means of its GitHub Copilot, which has 1.3 million paid customers and has been adopted by greater than 77,000 organizations in roughly two years. Corporations like Honeywell, State Road, TD Financial institution Group and Levi’s are amongst its customers. This widespread adoption stems largely from Microsoft’s current enterprise relationships and its first-mover benefit, whereby it invested early into OpenAI and used that firm’s fashions to energy Copilot.

Nonetheless, even Microsoft has acknowledged Anthropic’s power. In October, it allowed GitHub Copilot customers to decide on Anthropic’s fashions as a substitute for OpenAI. And OpenAI’s latest fashions — o1 and the newer o3, which emphasize reasoning by means of prolonged pondering — haven’t demonstrated explicit strengths in coding or agentic duties.

Google has made its personal play by not too long ago making its Code Help free, however this transfer appears extra defensive than strategic.

The open supply motion is one other vital pressure on this panorama. Meta’s Llama fashions have gained substantial enterprise traction, with main firms like AT&T, DoorDash and Goldman Sachs deploying Llama-based fashions for varied functions. The open-source strategy gives enterprises larger management, customization choices and value advantages that closed fashions can’t match, as VentureBeat reported final 12 months.

Relatively than seeing this as a direct menace, Anthropic seems to be positioning itself as complementary to open supply. Enterprise clients can use Claude alongside open-source fashions relying on particular wants, a hybrid strategy that maximizes the strengths of every.

Actually, most enterprise firms of scale I’ve talked with over the previous a number of months are explicitly multimodal, in that their AI workflows permit them to make use of no matter mannequin is finest for a given case. Intuit was an early instance of an organization that had guess on OpenAI as a default for its tax return functions, however then final 12 months switched to Claude as a result of it was superior in some instances. The ache of switching led Intuit to create an AI orchestration framework that allowed switching between fashions to be far more seamless, as Nhung Ho, Intuit’s VP of AI, advised VentureBeat on the time.

Most different enterprise firms have since adopted the same apply. They use no matter mannequin is finest for the precise case, pulling in fashions with easy API calls. In some instances, an open-source mannequin like Llama may work effectively, however in others — for instance, in calculations the place accuracy is vital — Claude is the selection, Intuit’s Ho defined at VentureBeat’s VB Remodel occasion final 12 months. 

Over the previous couple of days, I’ve been attending the HumanX convention in Las Vegas, the place a whole bunch of builders gathered to speak about AI. Claude comes up virtually at all times every time the subject of brokers or coding is raised. Over lunch yesterday, Julianne Averill, managing director at Danforth Advisors, which advises life science firms, stated her firm had discovered Claude superior for a lot of such duties, together with constructing structured evaluation tables.

Vercel CEO Guillermo Rauch, one other attendee, stated his firm, which has surpassed $100 million in annual income, selected Claude final 12 months as its default mannequin to assist builders code after doing rigorous evaluations of all fashions. “3.7 is king,” Rauch advised VentureBeat. He agreed it’s vital to supply builders a alternative of fashions, because the breakneck tempo of advances means there can’t be loyalty to a single mannequin. However whereas Vercel’s V0 product, which lets customers generate internet person interfaces (UIs) utilizing natural-language prompts, gives that alternative, it has to select a default mannequin to assist customers throughout their preliminary ideation and reasoning part. That mannequin is Claude Sonnet. “You want the architect mannequin that’s able to reasoning and does the lion’s share of code technology,” he stated. “A major chunk of our pipeline is powered by Anthropic Sonnet.” Adobe, Chick-Fil-A and Mattress Bathtub and Past are Vercel’s clients. 

Nonetheless Rauch cautioned that fluidity within the LLM race stays, and the lead mannequin may change at any time. Vercel experimented with China’s DeepSeek, he stated, however discovered it fell simply wanting matching Claude’s Sonnet. Equally, he stated, Alibaba’s Qwen mannequin has gotten excellent.

Enterprise implications: Making the shift to coding brokers

For enterprise decision-makers, this quickly evolving panorama presents each alternatives and challenges.

Safety stays a prime concern, however a latest unbiased report discovered Claude 3.7 Sonnet to be probably the most safe mannequin but — the one one examined that proved “jailbreak-proof.” This safety stance, mixed with Anthropic’s backing from each Google and Amazon (and integration into AWS Bedrock), positions it effectively for enterprise adoption.

The rise of coding brokers isn’t simply altering how functions are constructed — it’s democratizing the method. In keeping with GitHub, 92% of U.S.-based builders at enterprise firms had been already utilizing AI-powered coding instruments at work 18 months in the past. That quantity has seemingly grown considerably since then.

“The problem that persons are having [because of] not being a coder is absolutely that they don’t know quite a lot of the terminology. They don’t know finest practices,” defined Witteveen. AI coding brokers more and more bridge this hole, permitting technical and non-technical crew members to collaborate extra successfully.

For enterprise adoption, Witteveen recommends a balanced strategy: “It’s the stability between safety and experimentation in the meanwhile. Clearly, on the developer aspect, persons are beginning to construct real-world apps with these things.”

For a deeper exploration of those points, take a look at my latest YouTube video dialog with Witteveen, the place we take a deep dive into the state of coding brokers and what they imply for enterprise AI technique.

Wanting forward: the way forward for enterprise coding

The rise of AI coding brokers indicators a elementary shift in enterprise software program improvement. When used successfully, these instruments don’t substitute builders however rework their roles, permitting them to concentrate on structure and innovation quite than implementation particulars.

Anthropic’s disciplined strategy in focusing particularly on coding capabilities whereas opponents chase a number of priorities seems to be paying dividends for the corporate. By the top of 2025, we might look again on this era because the second when AI coding brokers grew to become important enterprise instruments — with Claude main the way in which.

For technical decision-makers, the message is obvious: Begin experimenting with these instruments now or danger falling behind opponents who’re already utilizing them to speed up improvement cycles dramatically. This second echoes the early days of the iPhone revolution, when firms initially tried to dam “unsanctioned” gadgets from their company networks, solely to finally embrace BYOD insurance policies as worker demand grew to become overwhelming. Some firms VentureBeat has talked with, like Honeywell, have not too long ago equally tried to close down “rogue” use of AI coding instruments not accepted by IT. 

Talking Monday on the HumanX convention, James Reggio, the CTO of Brex, an organization that gives bank cards and different monetary companies to small and mid-sized enterprises, stated his firm initially additionally tried to implement a top-down strategy to AI mannequin choice, in an effort to achieve perfection. However the firm confronted revolt amongst its developer workers, and shortly realized this was futile. It determined to permit customers to experiment freely. Good firms are already organising safe sandbox environments to permit managed experimentation. Organizations that create clear guardrails whereas encouraging innovation will profit from each worker enthusiasm and insights about how these instruments can finest serve their distinctive wants — positioning themselves forward of opponents who resist change. And Anthropic’s Claude, no less than for now, is a giant beneficiary of this motion.

Watch my video with developer Sam Witteveen right here for a full deep dive into the coding agent development: