Why Companies Are Rethinking Their Dependence on OpenAI And Anthropic
OpenAI and Anthropic became the default choices for many companies before most had worked out what they actually needed from artificial intelligence. Teams began with a few ChatGPT or Claude subscriptions, developers connected the models to internal applications, and experimental projects gradually became part of daily operations. The initial decision appeared reversible: another model could always be introduced later.
That assumption is now being tested. As AI moves from occasional assistance into customer service, software development, research, document processing and operational workflows, the choice of provider affects far more than output quality. It shapes costs, data architecture, internal skills and the company’s ability to change direction. What began as a convenient technology purchase can become a strategic dependency.
The concern is not that OpenAI or Anthropic have suddenly become unsuitable. Both continue to produce some of the strongest general-purpose models available. The problem is that companies have often allowed one provider to become embedded across too many processes without establishing what would happen if prices, performance, contractual conditions or business priorities changed.
AI Costs Become Harder To Predict At Scale
A monthly subscription gives employees a relatively clear price. The economics change when a company begins using models through an application programming interface, particularly when AI is built into products or automated workflows.
Charges may depend on the amount of information sent to the model, the length of the response, the model selected and the frequency with which the process runs. A task that appears inexpensive during a pilot can become costly when it is repeated thousands of times a day. Adding longer documents, more detailed instructions or additional verification steps increases the bill further.
The difficulty is not limited to the price of individual tokens. AI systems require supporting infrastructure, monitoring, security controls, integration work and human review. Companies may also pay several technology providers for overlapping capabilities without having a clear view of which model is handling which task.
This is prompting finance and technology teams to ask questions that were often absent during the experimental phase. What does it cost to resolve one client request? Is the model generating enough value to justify its operating cost? Would a smaller model complete the same task satisfactorily? How much human correction is still required?
A capable model may produce an impressive answer, but that does not automatically make it the most economical option. Summarising a routine document, classifying an incoming request and extracting a date from an invoice rarely require the same computing power as analysing a complex contract or writing production software. Using the most advanced model for every task is comparable to assigning the company’s most expensive specialist to every administrative query.
Dependence Extends Beyond The Contract
Vendor dependence is sometimes described as a procurement issue: a company negotiates with one provider and may later move to another. With generative AI, the relationship is more complicated.
Prompts are refined for the behaviour of a particular model. Applications are built around its technical interface. Employees become familiar with its strengths and limitations. Security procedures, evaluation methods and approval workflows are designed for the provider’s system. Even the tone and structure of AI-generated material may gradually reflect how that model responds.
Changing provider can therefore require more than replacing an API key. Prompts may need to be rewritten, outputs retested and integrations rebuilt. A competing model may interpret instructions differently, support different file formats or require another approach to tool use. Processes that appeared portable can prove surprisingly specific to one provider.
The risk increases when AI is connected to company systems and permitted to take action. An assistant that merely drafts an email is relatively easy to replace. An agent that retrieves client information, updates records, prepares reports and triggers internal workflows is tied to permissions, data sources and business rules. Migrating such a system becomes an operational project.
This is one reason companies are beginning to separate their AI applications from the underlying model. Rather than allowing each department to build directly on one provider, they are introducing an intermediate layer that can route tasks to different models. The approach requires more discipline at the beginning, but it reduces the cost of changing providers later.
The Best Model Depends On The Work
OpenAI and Anthropic are frequently compared as though a company must select a single winner. In practice, their models may be suitable for different types of work, while neither is necessarily the right option for every task.
One model may perform better with complex reasoning, another with long documents or programming. A smaller commercial model may be sufficient for routine internal processes. An open model running within the company’s own infrastructure may be preferable when data sensitivity matters more than marginal differences in output quality.
This is turning model selection into a portfolio decision. Companies can reserve premium models for tasks where their capabilities materially improve the result, while directing repetitive, high-volume work to less expensive alternatives. Sensitive data can remain within controlled infrastructure, and teams can switch providers when performance or prices change.
The model-router concept remains unfamiliar outside technology departments, but its logic is straightforward. The employee or application submits a request without choosing a provider. The system classifies the task and sends it to the model that offers the best combination of quality, speed, cost and data protection.
A short translation may be processed by a compact model. A legal comparison may go to a stronger reasoning system. Confidential information may be handled locally. If one provider is unavailable, another can take over. The company manages the service it needs rather than organising its work around the product range of a single AI company.
Open Models Are Becoming Credible Alternatives
The position of the leading American providers is also being challenged by the rapid development of open and lower-cost models. Some can be downloaded, adapted and operated on a company’s own servers or through a chosen hosting provider. Others offer commercial access at prices designed to attract organisations concerned about the cost of frontier systems.
Open models do not remove complexity. A company may need specialist expertise to deploy, secure and maintain them. Hardware costs can be substantial, and the model still requires evaluation before it is trusted with business-critical tasks. For a smaller organisation, using a managed commercial service may remain simpler and cheaper.
Their significance lies in the alternative they create. Companies no longer have to assume that every AI workload should be sent to one of two American providers. An open model may be adequate for classification, search, extraction, translation or internal knowledge retrieval. It can also be adapted to a company’s terminology and operated under stricter control.
Chinese AI developers have added further pressure by demonstrating that capable models can be offered at markedly lower prices. Not every organisation will be comfortable using them, particularly where regulatory requirements, data transfers or geopolitical exposure are involved. Their influence nevertheless extends beyond direct adoption: lower-cost competition makes it harder for established providers to treat high prices as inevitable.
Human Work Sometimes Remains The Economical Choice
The rush to automate has encouraged companies to compare the cost of AI with the salary of an employee. That comparison can be misleading.
An automated process may require model usage, software integration, monitoring, quality controls and exception handling. When errors carry financial, legal or reputational consequences, a person must often review the output. If the task occurs infrequently or varies significantly from one case to another, the cost of building and maintaining the AI workflow may exceed the saving.
This does not mean that automation has failed. It means the company must consider the full process rather than the apparent cost of generating an answer. AI may still improve the employee’s productivity without taking over the entire task. A model can prepare the first analysis, identify missing information or suggest a response, while the employee remains responsible for the decision.
The most economical division of work is not always full automation. In some cases, it is a carefully designed combination of machine speed and human judgement.
What Companies Should Do Before Expanding Their AI Use
The first step is to map where AI is already being used. Many organisations underestimate the number of tools, subscriptions and informal workflows that have appeared across departments. Without that overview, they cannot calculate costs, assess data exposure or identify dependence on a particular provider.
Each use case should then be evaluated separately. The relevant questions are practical: how important is the task, what level of accuracy is required, what data is involved, how frequently does it run and what happens when the model is wrong? The answers determine whether the company needs a premium model, a smaller alternative, local processing or direct human involvement.
New AI applications should be designed for portability wherever possible. Prompts, data connections and business rules should not be buried inside one provider’s platform. Companies should maintain independent tests that allow them to compare models against the same tasks and expected results. Contracts should also address data use, service availability, price changes and the practicalities of moving to another supplier.
None of this requires companies to abandon OpenAI or Anthropic. Dependence becomes dangerous when it is accidental, poorly understood and difficult to reverse. A deliberate relationship with a leading provider can be entirely rational, provided the company knows which processes rely on it, what they cost and what alternatives exist.
The next phase of corporate AI will be less concerned with access to the most celebrated model and more concerned with control. Companies that retain the ability to compare, combine and replace models will be in a stronger position than those that allow one provider to define the architecture of their AI strategy.
