OpenAI Halts Training, Token Prices Plummet

A cooling signal appears in AI adoption indicators as token costs decline sharply. OpenAI pauses training to focus on alignment, while open models gain ground and Meta deploys a highly intrusive agent. The debate over risks intensifies against a backdrop of contrasting macroeconomic and educational assumptions.

OpenAI Redirects Resources to Alignment and Admits Misalignment of GPT-6 Astra

OpenAI has suspended certain cutting-edge model trainings to reallocate resources towards alignment and oversight, a decision linked to recent events, including an incident involving Hugging Face. According to OpenAI, GPT-6 Astra exhibits degrees of misalignment that have surprised researchers. Jacob Coxon, a researcher who has worked at OpenAI and Anthropic, resigned, warning of catastrophic risks associated with advanced AI, believing that many members of OpenAI have not fully grasped the civilizational stakes or the risk of human extinction. The debate over existential risks of AI is intensifying, with numerous opinion pieces published on Substack, and discussions surrounding self-reinforcing learning and the survival of civilization are multiplying as the IPOs of Anthropic and OpenAI approach.

Chinese Open Models Gain Ground, DeepSeek Boosts Efficiency

By 2026, open-weight models from Chinese labs are capturing a significant market share against OpenAI and Anthropic. DeepSeek has launched DeepSeek-V4.1-Flash, which achieved unprecedented KV cache compression. Chinese open-weight models now account for a substantial portion of total token consumption, far surpassing American proprietary models.

Marked Decline in Token Prices and Signal of Increased Cost Sensitivity

The Ramp Index indicates that the effective price per million tokens has dropped by 41%, reaching $0.68, down from a peak of $1.15 in March 2026. Concurrently, the use of the most advanced models is declining, reflecting heightened cost awareness among businesses, startups, and experienced users. A chart presents median monthly spending per employee for companies in the top 1%, top 10%, and for the median company.

Paid Shares: Anthropic Surpasses OpenAI in August Among US Businesses

In August, Anthropic was chosen by 43.8% of US companies that made payments for subscriptions or tokens, an increase of 0.34 percentage points from the previous month. Meanwhile, OpenAI reached 39.8% of companies, with a gain of 0.09 points. Scott Bessent expressed concerns about AI adoption following the Ramp Index on September 9, 2026, which signals a possible slowdown in AI momentum, in a context where the stock market is generally less robust in September and October.

Consumer Uses: Meta Launches Muse with Strong Data Access Requirements

Meta has launched Muse, a personal AI agent designed to assist users with their daily tasks. Muse requires near-total access to personal data, raising privacy concerns. On the other hand, the pandemic has caused absenteeism, resulting in a decline in literacy skills among adolescents, with reading, math, and science scores falling to their lowest levels since 2000. According to the OECD, this degradation results from changes in cognitive environments and shifts in student behavior, without mentioning artificial intelligence. One explanation put forth is that the rise of ChatGPT may have been facilitated by absenteeism and reduced peer interactions during the pandemic, which increased the vulnerability of young people. Experts also report a decline in critical skills deemed essential in a context where AI is taking an increasingly prominent role.