How to Escape OpenAI, Anthropic, and Their Expensive APIs

Access to advanced AI has become a challenge for enterprises and organisations. 

As powerful models become accessible with just a few clicks and are heavily promoted in software engineering, companies are increasingly debating how much AI access to allow their employees.

Some have taken a strict approach. After exceeding its annual AI budget in just one month, Uber introduced a $1,500 monthly per-employee limit to contain token costs. Amazon and Meta, too, have reportedly begun steering employees towards cheaper AI models and imposing tighter controls on usage. 

This creates a paradox. Enterprise adoption is expected to become the primary revenue driver for model providers, yet enterprises are simultaneously restricting AI usage. 

Founder and CEO of Palo Alto Networks, Nikesh Arora, argued on X that if consumer AI largely involves giving away free AI while burning capital, enterprise customers will ultimately fund frontier AI.

He suggested that labs need to ‘forward price’ tokens by reducing costs to encourage experimentation and workflow reimagination. “CIOs are busy restricting AI use and working on making the use more efficient,” he wrote.

Today, much of the enterprise AI ecosystem is focused on reducing AI costs by redesigning workflows. Enterprises are increasingly adopting powerful open source models, aided by Chinese AI companies that offer frontier-level performance at a fraction of the cost, or building routing architectures that send tasks to the most appropriate model.

However, the comparison between open source and proprietary models based solely on token prices is misplaced.

“Comparing proprietary APIs with open source models purely on token prices misses the real economic tradeoff,” says Leslie Joseph, Principal Analyst at Forrester Research, in an interaction with AIM.

The total cost of ownership includes infrastructure, engineering, governance, operations and ongoing model management.

“Open source becomes attractive at sufficient scale or where data sovereignty is critical. For many organisations, managed proprietary models remain the lower-cost option once operational complexity is considered,” adds Joseph.

Even companies that have successfully migrated to open source models acknowledge this trade-off.

Cloudflare recently highlighted why this operational complexity matters. While announcing support for Moonshot AI’s Kimi K2.5 model, it said that serving frontier open-source models at scale requires substantial engineering beyond downloading model weights, including extensive optimisation of the inference stack. 

Although the effort reduced inference costs for some internal workloads by 77%, Cloudflare noted that these optimisations require specialist infrastructure expertise, which platforms such as Workers AI package into a managed service.

Recently, Florent Crivello, CEO and co-founder of Lindy, an AI platform for building customisable no-code agents, wrote on X that the company had shifted all of its traffic to DeepSeek V4, replacing Anthropic’s models.

According to Crivello, the migration is saving the company millions of dollars while improving performance, although it required significant engineering effort. “I think these things are stickier than they sound—it’s not easy to switch models,” he tells AIM. 

He highlighted the trust US frontier AI models have earned and the brand impact shown in the saying, “no one ever got fired for picking Anthropic.”

Crivello added that for many companies, the cost of switching to Chinese AI models is not worth the effort.

The New Playbook for Enterprise AI

Another approach enterprises are adopting is improving AI discipline across the organisation through model routing and usage optimisation.

Recently, Brian Armstrong, CEO of Coinbase, revealed on X that the company made open-weight models such as GLM 5.2 and Kimi 2.7 the default via its internal LLM gateway, while still allowing engineers to choose any model they deemed appropriate.

Armstrong said the company had “cut our AI spend nearly in half, while our token usage continues to grow.”

This was also possible because 91% of employees never reached their AI usage limits, allowing Coinbase to replace tighter restrictions with cheaper default models.

The company also implemented automated model routing. Its internal AI preprocesses prompts and directs requests to the most cost-effective model based on the task, pricing and cache availability.

Employees are also encouraged to keep context windows lean by starting new sessions for different tasks, limiting file context, disconnecting unused tools and avoiding unnecessary token consumption.

Similarly, in an interview with VentureBeat, Farhan Thawar, head of engineering at Shopify, revealed that the company has built an internal LLM proxy that sits between employees and AI providers. 

The system purchases tokens in bulk, automatically routes traffic across providers and monitors token consumption, using circuit breakers rather than hard usage limits.

In an interaction with AIM, Nikhil Gandhi, senior vice president of engineering at Yahoo, said the company created a central AI council early in its adoption journey after realising that allowing every team to build independently would lead to fragmentation. 

The council evaluates models, develops reusable AI agents and shares best practices across engineering teams.

“None of the models are designed to work out of the box. You have to give them the right context,” he says. 

He added that the centralised approach has delivered “a lot more bang for the buck” by preventing teams from duplicating work.

Even so, many enterprises are unlikely to have the time, budget or engineering resources to build their own routing systems and AI infrastructure. For many, managed AI APIs remain the more practical option.

Gandhi adds that for enterprises deeply integrated with a particular AI stack, the engineering overhead of switching can outweigh the potential savings, making it an ROI decision rather than a purely technical one.

Not Using LLMs for Everything

“The cost of the underlying infrastructure has risen sharply in a short span of time,” Sujatha Iyer, head of AI security at Zoho, tells AIM.

“As storing enormous volumes of data, processing it, training models and running inference at scale all carry a significant cost, eventually someone has to pay for it.”

She adds that consumer AI apps often pay for this with user data, whereas enterprise AI bears the cost through token spending and rising AI budgets.

She also argues that the larger question is whether every AI requirement needs an LLM, whether open or closed. 

“Enterprises can also work with purpose-built machine learning models wherever they can,” she says. Iyer believes that adopting machine learning models and right-sized models where appropriate can often be more effective in enterprise settings.

She also notes that context is the real differentiator in enterprise workflows. “Even a simple query may draw from multiple systems, permission layers and business rules before it gives a meaningful answer.”

According to Iyer, ease of access to context layers must therefore begin early in the software development lifecycle.

 “The data access and permission layer has to be designed from day one.”

FDEs Will Train ‘Enterprise Waymos of the World’

The trend also explains the growing focus on Forward Deployed Engineers, or FDEs, across leading AI labs. 

These are engineers who work closely with enterprise companies on behalf of AI model makers to help integrate LLMs efficiently into their processes. 

Arora says these engineers will train the “enterprise Waymos of the world” by teaching generic AI models how each business operates, gradually turning them into autonomous enterprise systems capable of running complex workflows.

OpenAI has built a dedicated Forward Deployed Engineering organisation and has recently expanded the effort through its Deployment Company, while Anthropic has been scaling its Applied AI team and deploying engineers directly to strategic enterprise customers.

However, not everyone believes this FDE-heavy model is permanent.

Joseph argues that “Today's reliance on Forward Deployed Engineers reflects the immaturity of enterprise AI platforms more than a fundamental limitation of LLMs.”

He believes today’s enterprise AI tooling, including orchestration frameworks, governance systems, and evaluation platforms, still cannot automate much of the integration work required to deploy AI within businesses. As those platforms mature, much of that engineering will become built into the software itself.

FDEs will still be needed, he says, but their role will increasingly shift from integrating and deploying AI systems to adapting them to each organisation’s business processes.

Iyer shares a similar view. While acknowledging the value of FDEs, she also says, “Over time, more reliability should move into the platform itself, but enterprises will still need people who can translate business context into the right AI solution.”

Besides, Joseph also says, “High token prices are creating budget pressure, but they are not the main reason enterprise AI adoption is slowing. The market is moving from experimentation to economic discipline.”

He argues that CIOs are no longer funding AI because of its novelty, but because it can deliver measurable business outcomes. 

“Token costs matter, but they are only one part of the equation alongside governance, orchestration, data access, reliability, and integration.”

Joseph believes that enterprises focusing only on reducing inference costs, or indulging in ‘tokenmaxxing’, are often optimising the wrong variable. “The bigger challenge is designing AI systems that consistently generate business value.”