[KongchangAI]
· 1 min read· 990 words

AI Gateway Index Reveals: New Paying Teams Choose Anthropic Models Nearly 2:1 Over OpenAI

AI Gateway Index Reveals: New Paying Teams Choose Anthropic Models Nearly 2:1 Over OpenAI

AI gateway data shows new paying teams choose Anthropic over OpenAI nearly 2:1.

Based on the AI Gateway Index, the article notes that among new paying teams accessing models through specific gateways, the ratio of Anthropic to OpenAI adoption is nearly 2:1—contrasting with the perceived OpenAI-dominated narrative. The author attributes the shift to Anthropic's safety and compliance positioning for regulated industries like finance and healthcare, Claude's capabilities in long-context and complex reasoning plus friendly pricing, and multi-model gateways lowering switching costs so teams can vote by task performance. The article stresses the data covers only gateway-connected teams, not overall market share; OpenAI retains barriers via existing users, enterprise edition, and developer ecosystem, but continued skew in new traffic will gradually erode its share, explaining its recent moves on enterprise features and pricing. The recommendation for enterprises is to evaluate multiple models based on task types, compliance requirements, and cost structure, and to leverage gateways for testing and hybrid deployment.

Introduction

The latest AI Gateway Index shows that when it comes to enterprise AI model selection, new paying teams are favoring Anthropic's models over OpenAI's by a ratio of nearly 2:1. This data comes from real traffic statistics at the gateway level, offering a unique perspective on enterprise AI adoption trends.

Data at a Glance: The Preference Shift Among New Paying Teams

According to the index, new paying teams adopt Anthropic and OpenAI models at a ratio of nearly 2:1. This means that for every three new paying teams, roughly two choose Anthropic's models while only one chooses OpenAI. This ratio deviates significantly from the widely perceived "OpenAI-dominated" narrative, suggesting a structural shift is underway in the enterprise market.

It should be noted that this data only reflects new paying teams accessing models through specific AI gateways and is not a whole-market statistic. However, as an important hub for enterprises to uniformly access multiple models, gateway data carries signal value.

An AI gateway is middleware software that provides enterprises with unified access to multiple LLM APIs. Common forms include open-source projects (such as LiteLLM and OpenRouter) and commercial managed services. On the request side, it provides a unified interface, authentication, rate limiting, and cost accounting; on the model side, it handles routing, load balancing, and failover. The unique value of its data lies in the fact that gateways naturally record the raw facts of "which team called which model and how many tokens were consumed," and these choices are made after cross-vendor comparison—making them more reflective of true preferences than surveys or leaderboards. The limitation is that the sample only covers teams that adopt gateways, skewing toward medium-to-large organizations with higher technical maturity and multi-model strategies, and cannot be equated with overall market share.

Why Are Enterprises Turning to Anthropic?

Safety and Compliance Advantages

Anthropic has long emphasized AI safety and controllability, and its models perform notably well in reducing harmful outputs and following instructions. For heavily regulated industries such as finance and healthcare, compliance risk is a core consideration in model selection, and Anthropic's positioning fits this need precisely.

AI safety alignment refers to training models so that their behavior conforms to human intent and value norms. Anthropic's "Constitutional AI" uses a set of written principles in place of extensive human annotation to guide model self-correction, and its combination of RLHF (reinforcement learning from human feedback) and RLAIF (reinforcement learning from AI feedback) is the technical foundation of its safety narrative. In the context of enterprise procurement, "safety" also often refers to data not being used for training, regionalized storage, audit logs, and compliance certifications such as SOC 2—these and model alignment capabilities are two different evaluation dimensions and should be verified separately during model selection.

Model Capabilities and Cost-Effectiveness

The Claude series models are competitive on tasks such as long-context understanding and complex reasoning, and their pricing strategy is friendly to teams deploying at scale. New paying teams often conduct comparative testing across multiple models and ultimately choose the solution with the best overall cost-effectiveness.

Long context windows (e.g., hundreds of thousands to millions of tokens) allow models to process an entire contract, codebase, or long meeting transcript in a single request, reducing the information loss caused by chunked retrieval; reasoning models, meanwhile, generate a longer internal chain of thought before answering, significantly improving accuracy on math, coding, and multi-step planning tasks at the cost of increased latency and token consumption. When evaluating cost-effectiveness, enterprises should combine "price per thousand tokens" with "the total number of tokens and retries required to complete a task," using a fixed internal evaluation set for end-to-end cost comparison rather than looking only at the price list.

Multi-Model Gateways Lower Migration Costs

The proliferation of AI gateways allows enterprises to switch between or run multiple models in parallel at low cost. As migration barriers decrease, teams are more inclined to vote based on actual task performance rather than being locked into a single vendor.

Vendor lock-in refers to enterprises being deeply bound to a single model vendor due to data formats, proprietary APIs, fine-tuning artifacts, or prompt engineering, facing high rewriting and regression testing costs when migrating. A unified gateway abstracts an OpenAI-compatible request format, decoupling application-layer code from specific models; combined with prompt templates and evaluation pipelines, teams can distribute the same batch of requests to different models for A/B testing within minutes. It should be noted that migration costs are reduced but not eliminated—differences in how each model handles system prompts, tool-calling formats, and refusal boundaries still require targeted tuning.

Is OpenAI's Lead Shaking?

Despite trailing among new paying teams, OpenAI still has a massive existing user base and ecosystem moat thanks to its first-mover advantage. ChatGPT's enterprise edition, its API ecosystem, and the scale of its developer community will be difficult to fully displace in the short term.

But the directional shift in new traffic is worth watching: if the 2:1 preference among new paying teams persists, it will erode OpenAI's market share over time. This also explains why OpenAI has recently been making frequent moves on enterprise features and pricing strategy.

Implications for Developers and Enterprises

For enterprises currently selecting models, this index suggests: don't blindly follow the default option—conduct multi-model evaluations based on your own task types, compliance requirements, and cost structure. As a unified access layer, AI gateways can be an ideal tool for testing and hybrid deployment.

For model vendors, competition in the enterprise market has shifted from "single-point capability breakthroughs" to a comprehensive contest of "safety, cost, and ecosystem." The voting results of new paying teams may be more real than any evaluation leaderboard.

Conclusion

The AI Gateway Index provides a window into the dynamics of enterprise AI adoption. Anthropic's nearly 2:1 lead among new paying teams is a signal of market diversification, not the final outcome. Future changes in the index are worth continued attention.

#AI Gateway Index#Anthropic vs OpenAI#enterprise AI model selection#AI gateway#multi-model deployment
Share:

Related articles