OpenAI Leads Agent Orchestration Platforms Two to One, and Google Wins the Layer Underneath
New survey data from VentureBeat Intelligence puts numbers on a question many businesses are asking right now: when you hand work to an AI agent, whose software is running the show? Among companies that use an agent orchestration platform, 43 percent said they use OpenAI's, 22 percent said they use Anthropic's, and 21 percent said they use Google's.
That is the headline. The more useful story sits underneath it, because the same research found that Google leads where the work actually runs, that Anthropic appears on about as many shopping lists as OpenAI, and that changing the model inside a live application costs far more than most teams budget for.
What an agent orchestration platform actually is
An agent orchestration platform is the software layer that hands tasks to AI agents, calls the tools they depend on, and keeps every agent inside its permissions. If the model is the brain, the agent orchestration platform is the part that decides what an agent may touch, in what order, and what happens when a tool call fails. Our walkthrough of building production grade agentic systems covers what that looks like in practice.
The survey named the products it was measuring: the OpenAI Agents SDK and Responses API, the Claude Agent SDK and Managed Agents from Anthropic, and Google's Enterprise Agent Platform. Those three families cover most of the agent orchestration market at the enterprise end, though the survey also recorded mentions of Salesforce Agentforce, Microsoft Copilot Studio, Amazon Bedrock, and open source frameworks such as LangChain.
The headline numbers
VentureBeat fielded five surveys in September 2026. Three of them are relevant here, and each drew a different group of respondents from the same pool, which matters when you compare them.
| Question | Who led | Numbers |
|---|---|---|
| Agent orchestration platforms used | OpenAI | 43 percent OpenAI, 22 percent Anthropic, 21 percent Google |
| Primary agent orchestration platform | OpenAI | 25 percent OpenAI, 18 percent Google, 16 percent Anthropic |
| Cloud used for production AI | 51 percent Google Cloud, 36 percent Azure, 23 percent AWS | |
| Models used for production AI | OpenAI | 41 percent OpenAI, 39 percent Gemini, 15 percent Anthropic |
| Main provider in production | 32 percent Google Cloud or Gemini, 22 percent OpenAI, 6 percent Anthropic |
The pattern is consistent across the first four rows. OpenAI is the most common choice for the agent orchestration layer itself, while Google is the most common choice for the infrastructure that the agents run on. Counting either a cloud or a model as a vote for the vendor behind it, 68 percent of respondents use Google Cloud or Gemini against 62 percent who use Microsoft Azure or OpenAI.
One caveat sits inside those cloud numbers. A company that runs OpenAI's models through Azure may have answered “Microsoft”, and a company running Claude through Amazon or Google may have named the cloud instead of the model vendor. The survey authors flag this themselves.
Why this contradicts the numbers you may have seen elsewhere
If the finding surprises you, you have probably seen the Ramp data. Ramp tracks which companies its business customers pay, and it reported that Anthropic took the lead in business adoption during 2026, with the September update putting Anthropic at 43.8 percent of businesses against OpenAI at 39.8 percent.
Both sets of numbers can be right, because they measure different things in different populations:
- Ramp counts whether a business pays a vendor for subscriptions or tokens, across the businesses that use Ramp's cards. Each business counts once, so a small company weighs the same as a large one, and small companies are far more numerous.
- VentureBeat asked people at organizations with 100 or more employees, plus consultants who serve them, which agent orchestration platform they use and which providers run their production AI.
A single company can pay Anthropic for Claude subscriptions and for API access to write code, while running its agents on OpenAI's platform. Paying a vendor is not the same as choosing that vendor's agent orchestration platform as the place where work happens, and the two surveys are not in conflict once you separate those questions.
The revenue picture is messier still
The same week produced a reminder that vendor revenue figures depend on what gets counted. OpenAI told investors its annualized revenue for September was approaching 50 billion dollars, roughly 20 billion below the 70 billion figure that circulated widely in September, because the larger number counted sales made through cloud partners in the way Anthropic counts its own. Reporting on the lower figure moved technology stocks on 8 October.
The survey did not ask about revenue, and it did not ask about Anthropic's planned investor day, which Bloomberg reports is scheduled for its San Francisco headquarters this week after the company said in June that it had confidentially filed for an initial public offering. The point is worth keeping in view when a comparison of agent orchestration vendors turns into a comparison of business momentum, because those are separate measurements wearing similar clothes.
Anthropic is on about as many shopping lists as OpenAI
The most forward looking part of the survey asked which agent orchestration platforms companies are considering adding or replacing within the next 12 months, and allowed several answers. Of the 61 companies that named at least one, 33 percent named Anthropic's, 30 percent named OpenAI's, and 25 percent named Google's.
Two details matter when reading that result. First, 60 percent of the companies considering Anthropic's platform do not use it today, which makes those companies a genuine growth pool rather than a satisfied base. Second, the group is small, at 20 companies, so the shortlist finding is directionally interesting and statistically fragile at the same time.
Changing the model inside a live app is the expensive part
The finding with the most immediate practical value concerns model switching, which the survey defines as replacing the AI model inside a production application while keeping the application around it. The replacement can be a move to another vendor or to a newer model from the same one.
- 62 percent of companies had replaced a model inside a production app in the past year.
- Among those, 59 percent saw performance drop and either reworked the application or went back to the original model.
- Only 19 percent said the swap took almost no engineering work, while 75 percent spent a few days of engineering or more, or stopped before finishing.
Angie Brown, executive vice president and chief information officer at Home Depot, explained the stickiness in an interview for VentureBeat's Beyond the Pilot podcast: “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.” The same interview shows how mixed a large company's stack can be, with Home Depot running Google's Agent Development Kit alongside OpenAI models on Azure and Claude for code, a pattern that mirrors how developers are wiring Claude and Codex into their workflows.
One methodological note on that 62 percent figure, which the survey itself supplies: consultants reported swaps at a much higher rate than people working in house, at 81 percent against 51 percent. Taking those two groups together pushes the overall number up.
Retrieval is consolidating into the platforms you already buy
A third survey covered agentic context layers, meaning the systems that feed documents and facts to agents, the mechanism behind retrieval augmented generation. Among companies running retrieval in production, 58 percent use OpenAI's retrieval and 50 percent use Google's Vertex AI Search. The best known standalone tools trail far behind, with Pinecone and Elasticsearch each in the 12 to 14 percent range, though 40 percent of companies use at least one standalone tool.
For anyone choosing components, this is the same story as the platform numbers told from a different angle. The agent orchestration platform vendors are absorbing functions that used to be separate purchases, and retrieval is now one of them.
What this means if you buy or run AI tools
Five practical conclusions follow from the data.
Choose the agent orchestration platform that matches the rest of your stack. The survey's own numbers show that infrastructure habit and platform choice move together. Whether that platform is the OpenAI Agents SDK, the Claude Agent SDK, or a cloud vendor's equivalent, the shortest path is the one that matches the models and cloud you already run. A company already deep in Google Cloud has less reason to bolt on a different vendor's agent orchestration layer, and the reverse holds too.
Treat a model swap as a project, not a configuration change. With 59 percent of swappers needing rework or a rollback, the honest planning assumption is days of engineering and a retest of prompts, tools, and evaluations. Teams that assume a drop in replacement are the ones who end up reverting.
Do not read adoption as a verdict on quality. The vendor that wins on agent orchestration platform usage is not automatically the vendor winning on subscriptions, revenue, or model quality. The survey measures one layer of a five layer stack, and the layers disagree.
Watch the shortlists, not just the install base. An incumbent with a large installed base and a weak shortlist position can lose ground quickly. The reverse is also true, which is why the 33 percent figure for Anthropic matters even though the company trails on current use.
Expect bundling to continue. Retrieval, context management, and tool calling all started as separate purchases and are being folded into the platforms. Budget accordingly, because the standalone line items are the ones most likely to disappear from your renewal list.
A checklist for choosing an agent orchestration platform
Survey data is good at spotting a trend and useless on its own for making a decision. These six questions turn the findings into something you can act on, whether you are buying your first agent orchestration platform or replacing one.
1. Where does your data already live? The survey shows infrastructure habit and platform choice moving together, with 51 percent of respondents running production AI on Google Cloud and Google's position strengthening once its cloud and its models are counted together. Starting from the agent orchestration platform that matches your existing cloud and warehouse avoids building a bridge that someone then has to maintain.
2. How much of your own code does the SDK assume? The OpenAI Agents SDK, the Claude Agent SDK, and Google's Enterprise Agent Platform each expect a particular shape of application around them. Ask for a worked example that resembles your actual use case rather than a hello world sample, and count the files you would have to touch.
3. Is retrieval already bundled? With 58 percent of respondents using OpenAI's retrieval and 50 percent using Google's Vertex AI Search, the context layer is steadily being absorbed into the agent orchestration platform. Check what your shortlist already includes before renewing a standalone vector store for another year.
4. What does leaving cost? The model switching data is the closest available proxy, and it is not encouraging: 59 percent of companies that swapped a model in a live application had to rework it or revert. Keep prompts, tools, and evaluations in a form the vendor does not own, so that a future move is a project rather than a rewrite.
5. Who will actually run it? Consultants reported model swaps at 81 percent against 51 percent for in house teams, which suggests the two groups work at different speeds and carry different tolerance for churn. Match the agent orchestration platform to your own capacity rather than to a case study from a company with a much larger platform team.
6. Will the vendor still be on next year's shortlist? Anthropic appears on 33 percent of shortlists against OpenAI's 30 percent while trailing well behind on current use. An install base tells you what already happened. A shortlist tells you what buyers are about to do.
Answer those six questions and the survey stops being trivia. The vendor leading today's usage numbers is not automatically the right agent orchestration platform for your stack, and the switching data shows why guessing is the expensive option.
How much should you trust these numbers
The methodology deserves attention, because the survey is transparent about its limits and you should be too when you repeat the figures:
- The three surveys cited here had 109, 102, and 113 qualified respondents. 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 people who opt into the surveys, drawn from VentureBeat readers and a research panel. These are not the same people month to month.
- The results describe these specific respondents and may not hold for all companies.
- The claim that OpenAI leads in agent orchestration platform use did pass a statistical test at 95 percent confidence, including among in house respondents alone.
- The shortlist comparison rests on just 20 companies that named Anthropic's platform, which the survey authors describe as small.
One more honest note: these numbers come from a single publisher's opt in panel, and I have not seen the underlying data. Treat them as a strong signal about a specific population of enterprise respondents, not as a census of the market.
FAQ
Is OpenAI really twice as popular as Anthropic for AI agents?
On this survey, among companies that use an agent orchestration platform, 43 percent use OpenAI's against 22 percent for Anthropic's. That is close to double, and the lead passed a statistical test. It says nothing about subscriptions, model quality, or revenue, where other data sources tell a different story.
Why do Ramp and VentureBeat disagree?
They measure different things. Ramp counts businesses that pay a vendor, weighting a small company the same as a large one. VentureBeat asks larger organizations which agent orchestration platform they use and which providers run their production AI. A company can pay Anthropic and still run its agents on OpenAI's platform.
Should a small business care about agent orchestration at all?
Only when a task genuinely needs several steps and tools, the kind of work browser agents already attempt. The practical takeaway from the model switching data applies to any size of business: whatever platform you choose, assume that changing it later costs engineering time, so pick the one that fits the stack you already run.
The short version
Among companies using an agent orchestration platform, OpenAI leads Anthropic roughly two to one, while Google leads on the cloud and infrastructure underneath. Anthropic is on about as many shortlists as OpenAI, which makes the next year more competitive than the current install base suggests. The most actionable finding for anyone building with these tools is that swapping a model inside a live application usually means rework, with 59 percent of companies that tried it needing changes or a rollback.
Sources
- Axios on OpenAI's annualized revenue: https://www.axios.com/2026/10/08/openai-50-billion-arr-anthropic-revenue-accounting
- Ramp AI Index on business AI adoption: https://ramp.com/data/ai-index-august-2026