Maybe due to the disappointing outcomes, Altman had previously written that GPT-4.5 would be the final of OpenAI’s conventional AI fashions, with GPT-5 deliberate to be a dynamic mixture of “non-reasoning” LLMs and simulated reasoning fashions like o3.

A stratospheric worth and a tech dead-end

And about that worth—it is a doozy. GPT-4.5 prices $75 per million enter tokens and $150 per million output tokens by means of the API, in comparison with GPT-4o’s $2.50 per million enter tokens and $10 per million output tokens. (Tokens are chunks of knowledge utilized by AI fashions for processing). For builders utilizing OpenAI fashions, this pricing makes GPT-4.5 impractical for a lot of purposes the place GPT-4o already performs adequately.

In contrast, OpenAI’s flagship reasoning mannequin, o1 professional, prices $15 per million enter tokens and $60 per million output tokens—considerably lower than GPT-4.5 regardless of providing specialised simulated reasoning capabilities. Much more putting, the o3-mini mannequin prices simply $1.10 per million enter tokens and $4.40 per million output tokens, making it cheaper than even GPT-4o whereas offering a lot stronger efficiency on particular duties.

OpenAI has probably recognized about diminishing returns in coaching LLMs for a while. Consequently, the corporate spent most of final yr engaged on simulated reasoning fashions like o1 and o3, which use a distinct inference-time (runtime) method to enhancing efficiency as an alternative of throwing ever-larger quantities of coaching knowledge at GPT-style AI fashions.

Whereas this looks like dangerous information for OpenAI within the quick time period, competitors is prospering within the AI market. Anthropic’s Claude 3.7 Sonnet has demonstrated vastly higher efficiency than GPT-4.5, with a reportedly extra environment friendly structure. It is value noting that Claude 3.7 Sonnet is probably going a system of AI fashions working collectively behind the scenes, though Anthropic has not offered particulars about its structure.

For now, plainly GPT-4.5 would be the final of its sort—a technological dead-end for an unsupervised studying method that has paved the best way for brand spanking new architectures in AI fashions, akin to o3’s inference-time reasoning and even perhaps one thing extra novel, like diffusion-based models. Solely time will inform how issues find yourself.

GPT-4.5 is now accessible to ChatGPT Professional subscribers, with rollout to Plus and Workforce subscribers deliberate for subsequent week, adopted by Enterprise and Training prospects the week after. Builders can entry it by means of OpenAI’s varied APIs on paid tiers, although the corporate is unsure about its long-term availability.