Enterprises that use OpenAI's agent platform are far more likely to build on it than enterprises that use Anthropic's, according to new VB Pulse data from August. Of the 75 enterprises that have OpenAI's Agents SDK or Responses API somewhere in their agent stack, 52, or 69%, name it their primary orchestration platform. Of the 45 that have Anthropic's Claude Platform and Agent Skills, 17, or 38%, do the same. Enterprises that use both are counted in both groups.
That 38% is the lowest rate of any platform in at least 25 enterprise stacks, and the pattern holds in every company-size and industry group we looked at, though those groups are too small to report as percentages. On primary platform alone, OpenAI leads the August wave: 53 of the 162 enterprises that named a primary platform chose OpenAI, or 33%, followed by Google's Enterprise Agent Platform at 39, or 24%, and Anthropic at 18, or 11%. These counts include one respondent each for OpenAI and Anthropic who named that platform primary without marking it as one they use.
That gap between using a platform and building on it is the measurement missing from the argument now running through the industry about which model provider is winning the enterprise. The figures being cited in that argument, from the corporate spend platform Ramp and the model routing service OpenRouter, measure spending. Neither can tell the difference between a platform a company is trying and one it builds on.
Anthropic leads on one measure: consideration relative to its current user base. Among the 121 enterprises that named platforms they are considering adopting, adding or replacing in the next 12 months, 36 named Anthropic, against the 45 that use it today. That ratio of consideration to current use is the highest of any platform used at 40 or more enterprises in the survey. OpenAI drew 41 considerers against 75 users, Google 30 against 65 and Microsoft 15 against 31.
Methodology
This story uses the August wave of the VB Pulse agentic orchestration tracker, one of five monthly VentureBeat surveys of enterprise AI buyers. The wave drew 221 responses and qualified 169 at organizations with 100 or more employees. Of those 169 respondents, 83 said they are the final decision-maker for AI purchases and 55 said they are a recommender or influencer. The tracker's leading platform has changed from wave to wave as its sample has changed: Microsoft AI Foundry / Copilot Studio led the April through May wave, with OpenAI second; Anthropic led in June; and Microsoft led again in July. The vendor figures in this article are an August snapshot, not a trend. Question bases, sample limits and how to read the vendor figures are at the foot of this article.
What enterprises are actually deciding
The reason the primary-platform question matters more than an installation count is what sits on top of it. When a company runs AI agents that take several steps — for example, reading a file, calling a tool, deciding something and calling another tool — a coordinating layer has to decide what runs when, what each agent may touch and what happens when a step fails. That layer is the agent control plane, and where it lives is the structural decision enterprises are making right now. (Engineers building a single agent more often call its loop a “harness”, and the managed environment that loop runs inside a “runtime.” The control plane is the layer above both, the one that decides which agents exist at all, what each is allowed to touch, and who answers for them. Amazon Web Services, for example, encodes that order in its own interface, where the Bedrock AgentCore control plane is the thing that creates and manages agent runtimes.)
A quarter of enterprises expect a model provider to own it. Asked where they expect the primary control plane for enterprise agents to live by the end of 2026, 44 of 169 respondents, or 26%, said a provider-managed agent service, second only to a hybrid of provider-native and external orchestration, chosen by 55, or 33%.
Inflexibility across models and tools is the main concern enterprises have about letting a model provider hold orchestration control, chosen by 48 of the 169 respondents, or 28%. Vendor lock-in ranks fourth at 16%, or 27 of 169.
What the widely cited numbers can and cannot see
Two data sets have driven most of the September discussion. OpenRouter, a service that routes developer requests to whichever model the developer selects, published weekly figures showing spending on OpenAI models passing spending on Anthropic models in the week of September 7, for the first time since February 2024. Ramp, which builds an index from card and bill-payment transactions across tens of thousands of businesses, published a chart on model-level share of enterprise AI spend, showing OpenAI's Astra overtaking Anthropic's Fable in September.
Both are real measurements of the populations they cover, but neither covers the buyers in this survey. OpenRouter's 2025 State of AI report describes its users as developers and end users, with more than half of usage originating outside the United States. Large enterprises often buy model access on negotiated contracts or through Amazon Bedrock, Google Vertex AI and the Microsoft Azure marketplace rather than through a routing aggregator.
Ramp's own two published cuts of its data point in different directions in September. The model-level spend chart that circulated shows OpenAI's Astra climbing past Anthropic's Fable. Ramp's broader AI Index, published September 9, put Anthropic at 43.8% of United States businesses in August and OpenAI at 39.8%, with Anthropic gaining slightly faster that month. Ramp counts corporate card and bill-payment transactions from businesses on its own platform, which leaves out large enterprises that use other spend-management providers. Its published methodology says the index may understate adoption because it misses free AI tools and employees' personal accounts. This survey has limits of its own: its respondents are self-selected, and this wave skews toward mid-market companies outside tech, as the methodology at the foot of this article explains.
The bottom line
On the evidence of this wave, OpenAI has the stronger primary-platform position among its current users and Anthropic has the fuller pipeline. OpenAI is in 75 of 168 enterprise stacks and is the primary platform at 52 of them. Anthropic is in 45 and is primary at 17, the lowest primary-platform rate of any platform in at least 25 stacks, while drawing more forward consideration relative to its installed base than any competitor.
Sixty percent of this wave, or 101 of 169 enterprises, said they plan to adopt, add or replace an agent orchestration platform within 12 months, and 47 of 169 said they plan to move within the quarter. For Anthropic, the open question is whether its consideration turns into primary-platform positions, or whether enterprises keep installing it alongside a platform they build on instead.
How we measured this
Question bases. Not every question was answered by all 169 respondents. Which platforms an organization uses: 168 respondents, several answers allowed. Which single platform is primary: 162 respondents who named a real platform. Which platforms an organization is considering: 121 respondents, because 48 left that block empty after saying they have no plans to change.
The sample. Self-selected from VentureBeat's audience and research panel, not a probability sample. The figures describe these respondents and generalize to the wider market only as far as that mix allows.
This wave skews mid-market. Twenty-seven of the 169 respondents, or 16%, are at organizations of 10,000 or more employees, against 51% in the July wave. Seventeen, or 10%, are in technology and software, against 53% in July.
Why there aren't any trend lines. Each monthly wave draws a different mix of respondents, and results have moved with that mix, so every figure in this article is an August cross-section rather than a trend.