OpenAI revealed that in the past few months, there has been a significant increase in the use of AI coding proxies by their internal research team. Converted at the pricing of API, by mid-August, the median daily token consumption for researchers in the research department had exceeded $600, with the top 10% of high-frequency users spending over $7,000 per day.

Rapid increase in usage intensity within three months

The company stated in its latest blog that researchers' dependence on the AI coding proxy is deepening, and the growth rate is faster than that of other teams within the company. OpenAI mentioned that such tools have significantly changed the way researchers work, including enabling faster code delivery, conducting more experiments, and entrusting higher-level tasks to the proxies.

The data provided by OpenAI also shows that in July, the average daily token expenditure for researchers in the research department was $162, which rose to over $600 by mid-August. The company also noted that these figures are based on API estimated prices and reflect theoretical costs, rather than the actual OpenAI payment amounts.

Technical support requirements have also decreased simultaneously.

As coding proxies take on more troubleshooting tasks, the activity of the manual technical support channels is also declining. OpenAI stated that since January of this year, the daily post volume on the internal technical support channels manned by humans has decreased by more than half.

This indicates that the purpose of the AI coding proxy is no longer limited to generating code; it has also begun to take over tasks such as debugging, problem localization, and handling some technical support issues. For the development team, this means that more repetitive tasks are being handed over to the model for processing.

Automation researchers continue to make progress.

The company also stated that it has achieved the goal of "automated research interns" proposed by CEO Sam Altman last year, which refers to a system that can complete tasks under human guidance. OpenAI mentioned that they are making progress towards creating "fully automated AI researchers" by March 2028.

There has been more recent discussion within the industry regarding the cost associated with the intensive use of AI. Previously, some companies encouraged employees to use AI as much as possible and even established leaderboards to track high-frequency users. However, as the costs of top-tier models have risen, companies have begun to pay more attention to whether the investment matches the output. Both Meta and Amazon have discontinued their internal token leaderboards. Amazon executives have also reminded employees not to use AI just for the sake of using AI.

Additional information:OpenAI previously held a more open attitude towards the high consumption of token. In June of this year, Sam Altman stated that among the company's employees, those with the highest usage could consume up to 100 billion token per month. OpenClaw, the founder who joined OpenAI in February, also shared a screenshot of their monthly token bill in May, which amounted to approximately 1.3 million US dollars.