OpenAI updated the AI model powering Operator from the previous custom version of GPT-4o to a model based on o3, one of the latest in OpenAI’s o series of “reasoning” models. The o3 Operator was fine-tuned with additional safety data for computer use and included safety datasets designed to teach the model decision boundaries.
OpenAI upgraded the Operator in ChatGPT with a new Computer-Using Agent (CUA) model based on a version of OpenAI o3. With the new model, the Operator became more persistent and more accurate when interacting with the browser, improving the overall task success rate. It also delivers better-structured responses that are more clear and thorough.
According to OpenAI, the new CUA model showed stronger performance relative to the industry, achieving SOTA on OSWorld and WebArena. It also showed stronger relative performance to the previous version, both in established benchmarks and human preference evaluations.
OpenAI replaces the GPT‑4o-based model with a version based on o3
OpenAI hints at a big upgrade for ChatGPT Operator Agent pic.twitter.com/iGPQp9butD
— SabatAge (@sabatage) May 22, 2025
OpenAI replaced the existing GPT‑4o-based model for Operator with a version based on OpenAI o3, although the API version will remain based on 4o. The AI firm also claimed that the o3 Operator uses the same multi-layered safety approach used for the 4o version.
However, compared with other models in the o3 family, the o3 Operator was fine-tuned with additional safety data for computer use, including safety datasets designed to teach the model decision boundaries on confirmations and refusals.
OpenAI released a technical report showing the o3 Operator’s performance on specific safety evaluations. Compared to the GPT-4o Operator model, the o3 Operator was less likely to refuse to perform “illicit” activities and search for sensitive personal data and less susceptible to a form of AI attack known as “prompt injection.”
“o3 Operator uses the same multi-layered approach to safety that we used for the 4o version of Operator…Although the o3 Operator inherits o3’s coding capabilities, it does not have native access to a coding environment or Terminal.”
–OpenAI
The AI firm also disclosed that the new o3-based model went through standard safety evaluations, and Operator continued to be available as a research preview to ChatGPT Pro users globally. However, this upgraded model was only available in Operator in ChatGPT.
Knoop suspects running OpenAI’s o3 model might be costlier than expected
Last week, the Arc Prize Foundation, which maintains and administers ARC-AGI, updated its approximate computing costs for o3. The organization originally estimated that the best-performing configuration of o3 it tested, o3 high, cost around $3K to solve a single ARC-AGI problem. However, the Foundation now believes that the cost could be 10x higher than previously estimated, possibly around $30K per task.
Also, while OpenAI has yet to price o3 or release it fully even, one of the co-founders of the Arc Prize Foundation, Mike Knoop, believes the o1-pro model pricing is a reasonable proxy and a closer comparison of the true cost of o3. He, however, added that o3 would continue to be labeled as a preview on the leaderboard to reflect the uncertainty until official pricing was announced.
According to the Arc Prize Foundation, a high price for o3 high would not be out of the question, given the amount of computing resources the model reportedly uses. o3 high used 172x more computing than o3 low, the lowest-computing configuration of o3, to tackle ARC-AGI.
Rumors have been flying since early March about the pricey plans OpenAI was considering introducing for enterprise customers. The information reported that the company may charge up to $20K per month for specialized AI “agents,” like software developer agents.
However, while some argued that even OpenAI’s priciest models would cost well under what a typical human contractor or staffer would command, AI researcher Toby Ord pointed out that the models may not be as efficient. For instance, o3 high needed 1,024 attempts at each task in ARC-AGI to achieve its best score.
KEY Difference Wire: the secret tool crypto projects use to get guaranteed media coverage