For the first time, Anthropic’s Claude appears to have overtaken OpenAI’s ChatGPT in US business adoption, according to Ramp spending data cited by VentureBeat. The result is more than a rivalry score: it is an early glimpse of how companies may choose their AI partners when reliability, coding performance and cost matter more than consumer fame.
Imagine a software engineer arriving at work on a Monday morning to find that her team’s AI assistant has already reviewed the weekend’s pull requests, identified a security issue, drafted tests for an unfamiliar service and prepared a migration plan for a database that nobody wants to touch.
By lunchtime, the assistant has helped another team summarize a procurement contract. In the afternoon, it has turned a spreadsheet of customer complaints into a list of product priorities. None of this looks like a spectacular demonstration. There is no robot walking through an office or machine announcing that it has surpassed human intelligence.
Instead, the AI is scattered through ordinary work: inside a code editor, a browser tab, a customer-support console and a finance dashboard. It is useful because it is present at the right moment.
That is the kind of environment in which Anthropic’s Claude appears to be gaining ground.
Data from corporate-spending platform Ramp, cited by VentureBeat, reportedly shows Claude surpassing OpenAI’s ChatGPT in US business adoption for the first time. The measure is not a universal census of every company using an AI system, and it does not settle the question of which model is technically superior. But it offers an important signal: the assistant with the greatest consumer visibility is not necessarily the one companies are using most deeply in their operations.
The shift also reveals an uncomfortable paradox for Anthropic. Claude’s momentum is closely associated with Claude Code and other demanding professional workloads, particularly in software development, finance and professional services. Those workloads can create intense demand for tokens, computing capacity and premium access. The same customers helping Claude win enterprise attention may also be the customers most motivated to control its cost.
That leaves Anthropic with a difficult question. Can it turn a surge of business usage into a durable platform advantage, or will customers eventually distribute their workloads across ChatGPT, open models and cheaper specialist tools?
The answer will depend less on a leaderboard victory than on what happens after the first month of adoption.
The enterprise market is not the consumer market
Consumer attention and business adoption overlap, but they are not the same contest.
ChatGPT became a cultural reference point before many companies had established formal AI strategies. Its name became shorthand for generative AI itself. Employees brought it into the workplace independently, often using personal accounts to draft emails, explain code or summarize documents before information-technology departments had approved an official tool.
That early lead matters. Familiarity reduces friction, and a familiar product is easier to recommend inside an organization. ChatGPT also benefits from a broad product ecosystem, a large user base and OpenAI’s extensive partnerships. In many offices, “use ChatGPT” is the default instruction when somebody wants to experiment with an AI assistant.
But business adoption is shaped by a different set of questions.
Can a model handle a long, complicated document without losing the thread? Can it produce code that is not merely plausible, but safe to merge? Can an organization manage user permissions, data retention and audit logs? Can a finance department predict its monthly bill? Can an information-technology team integrate the system into existing software without rebuilding its workflow every few weeks?
A product can dominate public conversation and still lose a particular enterprise use case to a quieter competitor.
Ramp’s data is useful precisely because it observes spending behavior rather than asking people which brand they recognize. Corporate card and expense data can reveal when companies move from casual experimentation to paid deployment. It can show which services appear in business budgets, even when employees use multiple tools.
At the same time, the data should not be treated as a complete measurement of the AI market. Ramp’s customers are not every US business. Spending does not equal usage, and the number of companies paying for a product does not reveal how extensively each company uses it. Some organizations may buy a single subscription for a small team, while others may connect an AI service to thousands of employees or millions of lines of code.
The reported crossover is therefore best understood as a directional signal. It suggests that Claude has become a serious default choice for parts of the business market, not that ChatGPT has suddenly disappeared from corporate life.
Why Claude Code has become a strategic asset
Claude’s business momentum appears closely connected to Claude Code, Anthropic’s tool for working with software repositories and development environments.
Coding is an unusually powerful entry point for enterprise AI. Software teams already work in digital systems, their output can be inspected and tested, and the economic value of saving an engineer time is relatively easy to estimate. If an assistant helps a developer resolve a difficult bug, update outdated documentation or create a first version of a routine feature, the benefit can be measured in hours and project velocity.
The experience is also more intimate than a chatbot conversation. A coding assistant can inspect a repository, navigate files, understand dependencies, run commands, respond to test failures and revise its work. It becomes less like a search box and more like a junior collaborator who is available at any hour.
That does not mean the collaborator is always correct. It means the tool is placed inside a workflow where even imperfect assistance can be valuable if a human remains responsible for review.
Claude has developed a reputation among many developers for handling large amounts of context and producing useful results across complex coding tasks. Those perceptions are difficult to reduce to a single benchmark. In practice, engineers notice whether an assistant can understand an unfamiliar codebase, preserve the conventions of an existing project and make changes without creating a trail of new problems.
A model that writes a brilliant function but misunderstands the architecture may be less useful than one that makes conservative, consistent changes across several files. The winning quality is often not theatrical intelligence. It is continuity.
This is why coding can create unusually strong product loyalty. Once an assistant has been connected to a team’s repository, issue tracker and development habits, replacing it is not as simple as canceling a chat subscription. Developers form preferences around how a tool explains a change, how it handles failed tests and how much supervision it requires. Over time, those preferences can spread through a department.
Claude Code also gives Anthropic a route into broader organizational adoption. A company may begin with a handful of developers. Then product managers ask for help understanding technical constraints. Security teams examine generated code and documentation. Operations groups use similar models to automate repetitive analysis. The original coding deployment becomes a demonstration of what an AI system can do with company context.
In that sense, Claude Code is not just another product feature. It is a bridge from model capability to organizational habit.
The less visible workloads driving demand
Software development attracts attention because it is easy to picture. But Claude’s appeal to businesses may also come from workflows that are less visible to the public.
Professional-services firms work through large volumes of language: proposals, contracts, due-diligence materials, research, client correspondence and internal knowledge. Financial organizations face similar demands, combined with strict requirements around precision, confidentiality and traceability. Consultants need to compare documents, identify patterns and prepare briefings. Legal teams need to locate clauses and summarize risks. Analysts need to move from raw information to a defensible recommendation.
In these settings, the model’s value is not necessarily that it produces a final answer without supervision. Its value is that it helps a skilled person reach a final answer faster.
A lawyer may ask an assistant to map differences between two versions of a contract before reviewing the changes herself. A financial analyst may use it to turn a long earnings transcript into questions for an executive call. A consultant may create a first draft of a market assessment, then spend more time challenging its assumptions rather than formatting its pages.
The interface changes the rhythm of knowledge work. Instead of opening a blank document and facing a wall of information, an employee starts with a structured draft, a comparison table or a list of unanswered questions. The person’s role shifts toward judgment, verification and communication.
Claude’s ability to work with substantial context can be particularly important here. Businesses rarely deal in short, clean prompts. Their materials are messy: scanned documents, contradictory spreadsheets, internal acronyms, half-finished reports and email threads that contain essential facts only indirectly.
An assistant that performs well on a single question may still fail when asked to maintain a coherent understanding across a large collection of material. Enterprise customers discover this difference quickly. A prototype can look impressive in a demonstration; a production workflow must survive exceptions.
That is why business adoption can favor a model that feels steady rather than flashy. Employees remember the time an assistant confidently invented a policy or misunderstood a critical number. They also remember the tool that quietly handled a difficult task without forcing them to start over.
The token bill behind the success story
The most important challenge facing Anthropic may be hidden inside its success.
AI services are not ordinary software subscriptions. A conventional business application can add users at relatively predictable cost. An AI assistant can become more expensive as users ask longer questions, attach more files, run more steps and request increasingly complex outputs. Agentic tools can consume even more resources because they may inspect documents, call other systems, execute actions and revise their work repeatedly.
Claude Code is an example of a workload that can be highly valuable and computationally demanding. A developer may ask it to understand an entire repository, make changes across many files, run tests, diagnose failures and attempt another solution. The request is not one exchange. It is a chain of model interactions, often involving large volumes of input and output.
If thousands of developers use the product this way, demand can rise rapidly.
That creates a tension between customer enthusiasm and provider economics. Anthropic must make Claude capable enough to justify premium prices while keeping inference costs under control. It must also maintain fast response times when users are working against a deadline. A developer waiting several minutes for an answer may abandon a tool, even if its eventual output is excellent.
Capacity pressure can appear to users as rate limits, slower performance or restrictions on high-intensity features. These limits are understandable from a provider’s perspective. They are also disruptive inside a business. A team that builds an important workflow around an AI assistant does not want to discover that its access changes during a product launch or a sudden demand spike.
Companies are likely to ask increasingly precise questions about their AI bills:
- How many tokens did each department use?
- Which workflows generated measurable savings?
- Did the model reduce engineering time, or simply add another review burden?
- What happens to cost when the company doubles its usage?
- Can the provider guarantee capacity during critical periods?
- Is a cheaper model good enough for routine tasks?
The final question is especially important. Most organizations do not need their most powerful model for every request. A model that is excellent at understanding a complex architecture may be wasteful for classifying support tickets or extracting fields from a standard form.
This creates a market for model routing. Companies may send difficult tasks to Claude or ChatGPT while directing simpler work to less expensive systems. The best AI strategy may resemble a power grid, with different sources used according to the task, price and availability.
ChatGPT still has advantages that are difficult to dislodge
Claude’s reported lead should not obscure OpenAI’s strengths.
ChatGPT has a powerful consumer brand, broad awareness and a substantial installed base. Employees already know how to use it, which makes training easier. OpenAI also offers a wide range of capabilities, including text generation, analysis, image-related features and integrations that can appeal to organizations seeking a general-purpose assistant rather than a specialized coding partner.
That breadth matters to CIOs. A company may prefer a platform that can support marketing, sales, research, customer operations and software development under one commercial relationship. Even if another model performs slightly better in one category, consolidating procurement, administration and security review can be attractive.
OpenAI’s position is also reinforced by distribution. AI assistants become more valuable when they are near the tools employees already use. An assistant integrated into a company’s documents, meetings, email, messaging and productivity applications can become part of the daily flow of work. The less often employees must switch windows or copy information between systems, the more likely they are to use it.
This is a crucial distinction between model quality and product quality. An organization may select an assistant not because its raw responses are always best, but because it is accessible inside the systems where decisions already happen.
ChatGPT’s scale can also help it learn from a broad range of use cases, though the relationship between usage volume and product improvement is not automatic. Enterprise customers care about administrative control, privacy commitments and predictable behavior. A large platform can invest in these areas, but it can also become complex as features and policies accumulate.
The competition will therefore not be decided by one model answering isolated questions. It will be decided by which company can make an AI system feel dependable across thousands of small interactions every week.
The open-model pressure valve
Anthropic and OpenAI are not competing only with each other. They are also facing growing interest in open and low-cost models.
Open models offer organizations more control over where and how AI is deployed. A company may run a model in its own environment, use a specialized hosting provider or fine-tune the system for a particular domain. This can reduce dependence on a single vendor and make sensitive data easier to govern, depending on the architecture and the organization’s security practices.
The attraction is strongest when a company’s workload is repetitive and well understood. A customer-service classifier, document extractor or internal search component may not need the most capable general-purpose model. A smaller model can deliver acceptable performance at lower cost and with more predictable throughput.
Open models are not free in the broad sense. Organizations must pay for hardware, engineering, monitoring, maintenance, security and upgrades. Running a model internally can be more difficult than sending a request to a commercial API. The headline price per token does not capture the total operating burden.
Yet open models provide a valuable negotiating tool even when companies do not deploy them everywhere. A CIO can compare a premium model with a lower-cost alternative. An engineering team can reserve the expensive system for difficult cases. A procurement department can resist a contract that creates excessive dependence on one provider.
This is how open technology can influence a market without becoming the dominant option. It changes the outside options available to buyers.
The likely future is not a simple division between closed and open models. Businesses will combine them. A cloud-based frontier model may handle complex reasoning. A smaller hosted model may process routine documents. An internal model may work with especially sensitive information. A specialized coding system may operate inside development environments.
The providers that succeed will need to explain why their premium capability is worth paying for when a lower-cost model can handle a large share of everyday work.
Reliability is becoming the real benchmark
Public model comparisons often focus on a single question: Which system gives the best answer?
Enterprise buyers ask a broader question: Which system produces the lowest total cost of correct work?
That calculation includes more than the model’s output. It includes time spent reviewing mistakes, correcting misunderstandings, managing permissions, integrating APIs and retraining employees after product changes. It includes the cost of downtime and the risk of an incorrect answer reaching a customer or decision-maker.
Reliability is therefore multidimensional.
A reliable model should produce accurate responses, but it should also communicate uncertainty. It should follow instructions consistently, preserve context, respect access controls and avoid taking actions it was not authorized to take. When connected to business systems, it should fail safely.
For coding, reliability means more than passing a benchmark. It means making changes that fit the project, explaining what was modified and exposing assumptions that deserve review. For finance, it means distinguishing sourced figures from estimates. For professional services, it means showing where a conclusion came from and what remains uncertain.
The best enterprise assistant may not be the one that attempts the most. It may be the one that knows when to ask a clarifying question.
That behavior can feel slower in a demonstration. In production, it can prevent expensive errors.
Companies are also learning that reliability depends on workflow design. Giving an AI system unrestricted access to a large repository or database does not automatically create a useful employee. Teams need clear boundaries, human approval points and mechanisms for checking outputs. A weak process can make even a strong model appear unreliable.
This is where the rivalry may move next. The leading providers will compete to offer not just models, but systems for supervising models: permissions, evaluations, logs, testing environments, policy controls and ways to compare outputs across model versions.
The assistant becomes infrastructure. And infrastructure is judged by what happens when conditions are unfavorable.
The danger of vendor lock-in
As AI tools become more embedded in work, switching costs will rise.
A team may build prompts that encode its internal terminology. It may create automated workflows around one provider’s API. Developers may become accustomed to a particular coding assistant. Analysts may store structured outputs in a format designed for one system. Over time, the company accumulates not only data, but habits.
Vendor lock-in can emerge even without a long-term contract. If employees learn to think through one assistant, moving to another may require retraining. If an AI system is connected to internal tools, replacing it may involve a major engineering project. If the vendor changes pricing or access rules, the customer may discover that its alternatives are less ready than expected.
The practical response is not to avoid all providers. It is to design for portability where possible.
Organizations can separate their application logic from a single model API. They can maintain test sets that measure performance on their own work rather than relying only on public benchmarks. They can record prompts, expected outputs and evaluation criteria. They can route requests through an internal layer that makes it easier to change models later.
They can also distinguish between proprietary advantages and replaceable components. A company’s unique value may lie in its data, processes and expert judgment—not in permanently attaching itself to one model.
This is easier said than done. The best tools often become powerful because they are deeply integrated. Portability introduces additional complexity and may reduce the convenience that encouraged adoption in the first place.
CIOs will have to balance both risks: the risk of staying with a provider whose economics become unfavorable and the risk of building an architecture so abstract that employees receive a worse experience.
The winning strategy may be selective commitment. Companies will develop deep relationships with one or two providers for high-value workflows while keeping routine workloads portable and maintaining at least one credible alternative.
What the Claude lead really means
If Ramp’s reported data is a sign of a shift, it does not mean enterprise AI has chosen a permanent winner. It means the market is becoming more discriminating.
In the first phase of generative AI adoption, companies often asked whether they should use AI at all. In the next phase, they are asking which system should perform which task, under what controls and at what price.
Claude’s rise indicates that a focused product can challenge a dominant brand when it solves a pressing professional problem. For developers, that problem may be navigating complex codebases. For finance and professional-services teams, it may be managing dense information without losing important details.
But adoption is only the beginning. Anthropic must convert usage into durable value while managing the cost of serving customers who make unusually heavy demands. It must expand capacity, preserve reliability and give businesses enough control to plan around its service.
OpenAI, meanwhile, can rely on its enormous awareness only if it keeps turning that awareness into effective workplace tools. Distribution opens the door, but performance and trust determine whether companies continue walking through it.
Low-cost and open models will keep pressure on both companies. They will not need to replace frontier systems entirely. They only need to become good enough for enough tasks to force customers to question premium pricing.
The future workplace may contain no single AI champion. A developer could use Claude Code for a difficult refactor, ChatGPT to prepare a project briefing and an open model to classify routine issues. Employees may not even know which model answered a particular request. An orchestration layer could make that choice in the background, based on complexity, sensitivity, speed and cost.
In that world, the visible brand matters less than the invisible service beneath it. People will care whether the answer is ready when needed, whether the result can be trusted and whether the company can afford to use the system every day.
The next phase will be measured in ordinary moments
The most revealing test for Claude will not arrive as another dramatic ranking. It will appear in the routine moments that determine whether employees keep using a tool after the novelty fades.
A developer will ask the assistant to repair a test that failed just before a release. A consultant will upload a complicated client document and look for a contradiction. A financial analyst will question an unexpected number. A manager will need a summary before a meeting. A security officer will inspect the logs to understand what the system accessed.
If Claude handles those moments consistently, its reported business lead could become a foundation for long-term growth. If its popularity produces slow responses, unpredictable limits or bills that rise faster than the value delivered, customers will look elsewhere.
That is the central tension of enterprise AI. The most capable systems invite the most ambitious uses, but ambitious uses consume the most resources. The companies that manage this tension will define the next stage of the industry.
For Anthropic, the opportunity is substantial. Claude has shown that an enterprise challenger can break through the gravitational pull of a household-name product. Claude Code has given the company a particularly strong foothold in software organizations, where successful tools can spread through teams and departments.
The threat is equally clear. A lead built on intense usage can be expensive to maintain. Customers who depend on the service will scrutinize its price and availability more closely, not less. And every successful deployment teaches those customers how much of their work can be moved between providers.
Claude may have won an important round in the US business market. The larger contest is only beginning: to become the AI system that companies can trust not merely for impressive answers, but for the long, repetitive, financially consequential work of modern business.