The OpenAI o3 model could be more expensive to operate than initially estimated

OpenAI's o3 model may cost $30,000 per ARC-AGI task, far more expensive than expected.

: OpenAI's o3 'reasoning' AI model was initially thought to incur a cost of $3,000 for a single ARC-AGI problem. Revised estimates now put the cost at $30,000 per task, significantly higher than first anticipated. o3 High, with its substantial computing demands, appears less efficient and requires multiple attempts to achieve optimal scores compared to other models. Discussions about enterprise-focused pricing strategies, including rates of $20,000 per month, fuel ongoing debates regarding the model's value.

OpenAI's o3 'reasoning' AI model is under scrutiny following revised estimates showing its operational costs might surpass initial expectations. Unveiled in December with fanfare and tested using ARC-AGI, a benchmark for assessing advanced AI, the model's financial feasibility is now questioned. Originally, a single ARC-AGI problem with the o3 configuration was thought to cost around $3,000. However, the Arc Prize Foundation, the body maintaining ARC-AGI, has updated this figure, believing the actual cost could soar to approximately $30,000 per task, indicating a troubling disparity.

Such significant cost differences are noteworthy, reflecting broader concerns about the expenses associated with cutting-edge AI models in their nascent stages. The Arc Prize Foundation has yet to assign a price to o3, instead using OpenAI’s o1-pro model as a comparable reference. Mike Knoop, a co-founder of the Foundation, expresses this approach is based on the immense compute demands witnessed during testing, which seem to align more closely with o1-pro’s cost metrics.

OpenAI's ventures into potentially premium options for enterprise users have fueled speculation regarding their pricing strategies. Previous reports in March suggested that OpenAI might impose fees up to $20,000 monthly for bespoke AI 'agents,' aimed at specific tasks, such as software development, driving intrigue about commercial deployment strategies.

Despite the price tags, some argue the expense of these AI models may still be less than what businesses would spend on human personnel for similar tasks. Nonetheless, Toby Ord, an AI researcher, highlights concerns about the model's inefficiency, noting the high number of attempts required by o3 high to complete tasks within the ARC-AGI benchmark.

The o3 high configuration requiring 172 times more computing resources than the o3 low counterpart to tackle ARC-AGI emphasizes the differential in efficiency and potential applicability across use cases. As industry stakeholders await formal announcements regarding official pricing, the discussions around AI model costs serve as a cautionary tale for potential clients considering incorporating AI at a grand scale.

Sources: TechCrunch, The Information

AItechnologySoftware