Twice as many companies use OpenAI's agent orchestration platform as Anthropic's, VentureBeat Intelligence found in its September survey on agent orchestration. OpenAI holds that lead even though Anthropic may lead OpenAI on other important indicators. For example, Ramp, the corporate card and spending company, reports that more businesses now pay for Anthropic than for OpenAI.
An agent orchestration platform is the software that hands tasks to AI agents, calls tools, and keeps agents within their permissions. Among companies using one, 43% said they use OpenAI's, 22% said they use Anthropic's, and 21% said they use Google's. The survey named the products: OpenAI’s Agents SDK and Responses API, Anthropic’s Claude Agent SDK and Managed Agents, and Google’s Enterprise Agent Platform.
We also asked companies to name their primary agent orchestration platform. Of the companies using one, 25% named OpenAI's, 18% named Google's, and 16% named Anthropic's. The rest named platforms such as Salesforce’s Agentforce, Microsoft’s Copilot Studio, Amazon Bedrock, and open-source frameworks such as LangChain.
A second September survey, on AI infrastructure and compute, asked which cloud and model providers companies use to run AI in production. For cloud, 51% of respondents said they use Google Cloud, 36% Microsoft Azure, and 23% Amazon Web Services. For models, 41% said they use OpenAI’s, 39% Google’s Gemini, and 15% Anthropic’s.
When asked which provider is their main used in production, 32% named Google Cloud or Gemini, and 22% named OpenAI. Just 6% named Anthropic. It's important to note that, some companies that run OpenAI's models through Microsoft Azure, or Anthropic's Claude models through Amazon's or Google's clouds, may have named those clouds instead.
To give customers choice, Google sells both a major cloud and its own leading models. Counting either offering, 68% of respondents use Google Cloud or Gemini, against 62% that use Microsoft Azure or OpenAI.
Home Depot is a good example of how a company can run mostly on Google and still use other providers' models. "We're a pretty hefty Google shop, so we've got a lot of Google ADK running around there," Angie Brown, Home Depot's EVP and CIO, said in an interview for VentureBeat's Beyond the Pilot podcast, which will air later this month. ADK is Google's Agent Development Kit software for building AI agents. Brown said the vast majority of Home Depot's data sits in Google's BigQuery data warehouse. Home Depot also uses OpenAI's models through Microsoft Azure to judge the results it gets from Google's models, and it uses Anthropic's Claude to write code and OpenAI's models to review that code, Brown said. "So it's not that we are a Google-only shop," Brown said.
Anthropic, which said in June that it had confidentially filed for an initial public offering, plans to host selected institutional investors at a pre-IPO investor day at its San Francisco headquarters on Wednesday, Bloomberg reports. Ramp, which tracks what U.S. businesses on its platform pay for, reported on September 9 that 43.8% paid Anthropic for subscriptions or tokens, against 39.8% for OpenAI.
However, it’s important to note that Ramp’s data is biased toward the customers it tracks, which tend to be more tech-forward companies than average. Ramp’s figures cover only businesses that use Ramp’s cards and software, and Ramp counts each business once, so a 50-person company counts the same as a 50,000-person one. Because small companies far outnumber large ones, they weigh heavily in Ramp’s figures.
Ramp and VentureBeat measure different things in different groups of companies. Ramp counts how many businesses on its platform pay each company. VentureBeat asked people at organizations with 100 or more employees, and consultants serving them, which agent orchestration platforms they use, and which providers they use to run AI in production. A business can pay for Anthropic’s Claude subscriptions, and for Claude’s API to run coding tools, while running its agents on OpenAI’s platform.
Revenue comparisons between OpenAI and Anthropic also depend on what exactly is counted. OpenAI's annualized revenue is about $50 billion, roughly $20 billion below the $70 billion figure widely reported in September, according to Axios. The $70 billion figure counted OpenAI's sales through cloud partners, the way Anthropic counts its own. Tech stocks dropped on October 8 after the lower figure was reported, according to CNN. Our surveys did not ask about either company's revenue or Anthropic's planned IPO.
Anthropic's platform is on about as many shopping lists as OpenAI's
We asked companies planning to adopt a new, additional, or replacement agent orchestration platform within 12 months which platforms they are considering, and told them they could name several. Of the 61 companies that named at least one, 33% named Anthropic's, 30% named OpenAI's and 25% named Google's.
So, Anthropic's platform shows up on about as many shopping lists as OpenAI's.
Of the companies considering Anthropic's agent orchestration platform, 60% do not currently use it.
Swapping AI models in a live app usually meant extra work
Businesses on Ramp's platform are switching AI model providers more often than at any time in Ramp's data. Ramp reported on October 6 that 8% of businesses switched providers in an average month over the three months through September. Our agent orchestration survey also asked about a related step: Whether companies had replaced the AI model inside a production app in the past year while keeping the app around it. A replacement can mean moving to another provider's model, or to a newer model from the same provider. Of the companies surveyed, 62% had replaced a model inside a production app. Among those companies, 59% saw performance drop and had to rework the app, or went back to the original model.
Only 19% of the companies that swapped a model said the swap took almost no engineering work. For 75% of the companies that swapped, the switch took a few days of engineering work or more, or they stopped before finishing.
Brown described why switching models is costly at Home Depot: "One of the things is the context training that has to be done anytime there's a switch, so there's a stickiness level there that comes in."
Consultants reported swaps at a higher rate than in-house respondents, meaning people who work on AI for their own employer. Of consultants, 81% said their main client had swapped a model in the past year, against 51% of in-house respondents at their own companies. So, consultants pushed the overall 62% above what in-house respondents reported.
OpenAI’s and Google’s retrieval each run at half or more of companies, several times any standalone tool
In a third September survey on agentic context layers, 58% of companies running retrieval in production said they use OpenAI’s retrieval and 50% use Google’s Vertex AI Search. The most-used standalone tools, Pinecone and Elasticsearch, are each used by 12% to 14%, and 40% use at least one standalone tool.
What the data means for technical leaders
Well over half (59%) of companies that swapped a model in a live app reworked the app or went back to the original model. This data suggests that technical leaders planning to swap an AI model inside a live app should budget for rework. This means retesting the prompts, tools, and evaluations built around the old model.
VentureBeat Intelligence will ask the same questions in its October VB Pulse survey, so readers will see whether OpenAI's lead in agent orchestration platform use shows up again in a new group of respondents.
The full research reports can be found here.
About the data: VentureBeat Intelligence fielded five VB Pulse surveys in September 2026. Three are cited in this article: the Agentic Orchestration survey, with 109 qualified respondents; the AI Infrastructure and Compute survey, with 102; and the Agentic Context Layers survey, with 113. Respondents qualified if they work at an organization with 100 or more employees, or consult for one as their main client.
Each monthly survey is a snapshot of a different group of respondents who opt into the surveys. They come from VentureBeat’s readers and a research panel. Because September’s screening questions differed from August’s, the Google comparison counts only because it passed the statistical test on all September respondents and again on in-house respondents alone.
The results describe these specific respondents and may not hold for all companies. OpenAI's lead in agent orchestration platform use passed a statistical test at 95% confidence, including among in-house respondents alone. The group of companies considering Anthropic's platform is small, at 20 companies.