OpenAI has reportedly made guest access to ChatGPT much cheaper to operate, according to details attributed to a person familiar with internal discussions and reported by The Information.
The reported change focuses on inference costs: the expense of running existing AI models when users ask questions and receive responses. For visitors who do not have an account, OpenAI engineers reportedly said those costs had been cut by more than half.
What OpenAI reportedly changed
The optimization was applied to ChatGPT for visitors who do not have an account. That is an important boundary around the report, because it means the improvement was not described as a broad change across every version of ChatGPT.
According to the report, the number of Nvidia GPUs required to serve those guest users fell to just a few hundred. The source does not state how many GPUs were needed before, and it does not describe the techniques OpenAI used to achieve the reduction.
That leaves the central fact clear but narrow: OpenAI reportedly made one limited part of ChatGPT less expensive to serve. The scale of the previous resource use, the exact engineering path, and the applicability to other parts of the product were not disclosed.
Why inference costs matter
Inference is the recurring cost of using an already built AI model. Every answer generated by a system like ChatGPT requires compute, and that compute is served by hardware such as Nvidia GPUs.
Lower inference costs can matter because they change the operating pressure on an AI service. If a system can handle the same class of requests with fewer GPUs, the freed-up resources can be used in several ways.
- They could support scaling services to more users.
- They could help deliver faster responses.
- They could be directed toward better models.
- They could improve margins.
The report does not say which of those outcomes OpenAI will prioritize. It only states that the reported optimization reduced costs for guest ChatGPT users and lowered the GPU requirement for that group.
The limits of the reported gain
The biggest limitation is that guest users can access only a very limited set of ChatGPT features. Because of that, it is not clear whether the same gains would appear in the full product.
That distinction matters. A limited guest experience may involve fewer product paths than the full ChatGPT service. The source does not describe whether the optimization depends on that narrower feature set, so it would be premature to treat the result as proof that all ChatGPT response costs can fall by the same amount.
The report also does not identify the technical method behind the savings. Without that detail, there is no basis for saying whether the change came from model behavior, serving infrastructure, request handling, or another optimization route.
What can be said from the source is simpler: OpenAI reportedly reduced the cost of serving one defined user segment by more than half, and that segment now requires just a few hundred Nvidia GPUs.
How this fits the wider AI efficiency push
OpenAI is not the only organization tied to recent inference efficiency news. Deepseek also just dropped a new open-source method that can speed up inference requests by 60 to 85 percent.
That comparison points to a broader pressure in AI: as demand for model responses grows, companies and researchers are looking for ways to get more output from available hardware. Faster or cheaper inference can create more room inside existing compute limits.
Still, efficiency gains do not automatically mean demand for chips falls. The source notes that data center buildouts are moving slowly, which means gains like these will probably give labs more breathing room rather than reduce chip demand.
In practical terms, that means optimization may ease short-term constraints while services continue to grow. If fewer GPUs are needed for one workload, those resources can be shifted elsewhere instead of sitting unused.
What remains unknown
The report leaves several important questions unanswered. It does not say how many Nvidia GPUs guest ChatGPT previously required. It also does not explain what OpenAI changed under the hood.
Most importantly, it does not show whether the same kind of reduction can be applied to the full ChatGPT product. Guest access is narrower, and the source explicitly leaves open whether the improvement would carry over beyond that limited use case.
For now, the reported result is still significant. Cutting inference costs by more than half for any live user segment suggests that operational efficiency remains a major lever for AI companies. But the scope matters: this is a reported gain for visitors without an account, not a confirmed across-the-board cost reduction for all of ChatGPT.