OpenAI's GPT-6 Astra Drives Enterprise Spend Ahead of Anthropic

GPT-6 Astra is proving to be an absolute hit for OpenAI, judging by the growing quantum of enterprise dollars that the AI lab is now attracting versus its nemesis Anthropic, which now finds itself on the proverbial backfoot for the first time in around two and a half years.

OpenAI's GPT-6 Astra with its novel looped transformer architecture is now proving to be an absolute hit with enterprise customers

According to the data from OpenRouter, users spent more dollars on OpenAI models than those from Anthropic last week, essentially flipping a script that had been running for the past 2.5 years.

This view is also echoed by the data from Ramp, which shows OpenAI's GPT-6 Astra taking a 13 percent share of enterprise dollars last week versus just 8 percent for Anthropic's Fable-class models.

Of course, Anthropic still leads when it comes to the overall adoption rate. But, if the recent trends hold, OpenAI can stage a comeback on the basis of Astra's strength.

As we noted recently, OpenAI has trained the GPT-6 Astra on a 100,000-GPU cluster, with the training run likely to have cost the AI lab anywhere between $500 million and $1 billion. For the benefit of those who might not be aware, instead of requiring developers to furnish dedicated APIs for every application an AI agent needs to use, GPT-6 Astra is able to navigate any given software in a manner that is strikingly similar to how humans do so: by spawning agents that then work across browsers, spreadsheets, websites and desktop applications, producing finished documents and presentations, and implementing multistep workflows rather than merely telling a user how to complete them.

Basically, GPT-6 Astra's core reasoning engine utilizes a native multi-agent structure to solve complex problems. When faced with a complex task, the model's main orchestration agent can formulate a theory and deploy distinct sub-agents to test variations and validate results in parallel. This native delegation makes it highly resilient against "doom loops" - where a model gets stuck in repetitive error cycles - allowing it to autonomously debug its own code and adjust its strategy on the fly.

What's more, according to SemiAnalysis, it is quite likely that OpenAI's GPT-6 Astra uses a looped transformer architecture. To recap, An AI model basically consists of a series of transformer blocks or layers that are made up of two key elements:

1. The Attention Layer looks at how words in a sentence connect with each other. For instance, if you give the model a prompt that says Paris is the capital of France, the attention layer will link the word 'capital' to 'France' and identify 'Paris' as the answer. Even so, this layer does not know what is meant by France or Paris, but would save these notes in the form of KV cache. As context increases, so does KV cache.
2. The Feed-Forward Network (FFN) holds the sum knowledge of a model in the form of weights, and it is this layer that uses deep mathematical formulas to unearth the meaning behind each word. For instance, it would look at France and activate its country encyclopedia, and so on.

In a normal model, an input token moves through layer 1, then layer 2, all the way up to the last layer, with each layer possessing its own unique weights. In a looped transformer architecture, however, the token is fed through a given transformer block multiple times, where each pass through the same weights allows the layer to refine its output. This allows the model's responses to become much more accurate without growing the quantum of weights or parameters. On the flip side, looped transformers significantly increase the size of the KV cache. It is this unique architecture, however, that likely allows OpenAI's GPT-6 Astra to show sizable performance gains.

Finally, do note that Sam Altman has just teased a major model release this week, with next week expected to see multiple launches akin to OpenAI's DevDay 2025, when the AI lab had announced seven major product launches and platform updates.

Follow Wccftech on Google to get more of our news coverage in your feeds.