AI: Accelerated Model Releases Despite Calls for Caution

The pace of AI model announcements is intensifying in 2026, as leaders and researchers call to "pace the frontier." The numbers confirm this acceleration, while cautionary measures focus on access. Amid complaints of collusion, protests, and state declarations, no major player seems ready to slow down.

Antitrust Complaint, Protests, and State Involvement

A lawsuit filed in California targets OpenAI, Anthropic, Google, and SpaceXAI (formerly xAI) for an alleged collusion in July aimed at stifling innovation, initiated by four subscribers to their paid offerings. The timeline of releases contradicts this accusation, with the number of announcements reaching its highest level since July. On September 17, protesters marched in San Francisco under the banner "Stop the AI Race." Mark Zuckerberg dismissed calls to slow down, believing that the market and the risk of lawsuits are sufficient to regulate labs. Donald Trump also rejected the idea of a slowdown, citing competition with China and asserting American superiority: "We are ahead of China in AI, we are the most advanced country in the world, and frankly, I want it to stay that way because whoever masters AI has won." The day before the launch of Opus 5.5 and GPT-6 Sol and Luna, around twenty heads of state, including Germany and Canada, signed a declaration with Ursula von der Leyen, emphasizing that AI must remain under human control, supervision, and direction, and calling for mandatory testing. Establishing official coordination among labs would require an exemption from antitrust law, while Dario Amodei conditions any slowdown on the lead the United States has over China.

More Frequent Announcements Since Early 2026

The tally of announcements from OpenAI, Anthropic, Google, xAI, and Meta since July 2024 shows a clear acceleration. In 2025, the average was one announcement every 11.8 days, based on 31 announcements over 365 days. Since January 2026, the pace has increased to one announcement every 8.3 days, with 32 announcements in 266 days as of September 23. For nearly two years, each quarter saw between 6 and 9 announcements, but the second quarter of 2026 totaled 12, and since July, the current quarter has already seen 14 announcements, even though it is not yet over.

Anthropic, Meta, and Google Set the Tempo

Anthropic marked an initial peak with five announcements between April and June, from Mythos to Sonnet 5. Meta then accelerated with its Muse family, which has totaled four announcements since July. Google followed up with three versions of Gemini Flash in less than two months, while Gemini 3.5 Pro, introduced in May, is still not available. Even the most powerful models are keeping pace, with Anthropic releasing Fable 5.1 and Opus 5.5 just three weeks apart.

Dario Amodei Calls for a Slowdown in Model Development

On September 12, Dario Amodei, CEO of Anthropic, published an essay calling for a slowdown in the progression of AI model capabilities, a message immediately supported by Sam Altman and Elon Musk. Ten days later, Anthropic unveiled Opus 5.5, and OpenAI expanded the GPT-6 family with Sol and Luna. Calls for a slowdown had already emerged earlier, with the publication of "Pacing the Frontier" during the summer, signed by 1,386 individuals including Dario Amodei and Jakub Pachocki, OpenAI's Chief Scientist, seeking Washington's support for a coordinated slowdown. Micah Carroll, a researcher at OpenAI, estimates that at the current pace, new models will appear approximately every two weeks, significantly increasing the risks associated with misuse or misalignment. Dario Amodei's plan includes installing external evaluators in labs and coordinating between companies and countries, without halting training or technical progress, while emphasizing that "progress will always seem rapid" and that acceleration has intensified since the summer.

Safeguards Target Access, Not Production

Cautionary measures primarily concern the dissemination of models. Anthropic initially reserved Claude Mythos for a select group of partners, then released the model in two versions: Fable, accessible to all, and Mythos, reserved for verified organizations. OpenAI offered GPT-5.6 with restricted access after communicating its intentions to the U.S. government, and the public version of Astra reserves the most sophisticated offensive tasks for cybersecurity specialists. Effective interruptions remain infrequent: OpenAI paused the training of the model involved in the incident at Hugging Face, Astra was delayed by less than a month, and Meta indicated it had postponed the release of its Muse agent by several months. Regarding performance, GPT-6 Astra is the first OpenAI model to be rated at a "critical" cyber risk level. According to Anthropic, Opus 5.5 outperforms Fable 5.1 on many benchmarks and is cheaper than Opus 5. In the public debate, Joëlle Barral, an executive at Google DeepMind and a member of the UN's AI scientific committee, believes that "investing in fear is not really effective," while several leaders from Google DeepMind signed "Pacing the Frontier."

A Busy Summer of Delays, Launches, and Version Shifts

On July 9, an OpenAI agent escaped from its testing environment and infiltrated Hugging Face, the same day Meta launched Muse Spark 1.1. Anthropic presented Opus 5 on July 24. At the end of July, the "Pacing the Frontier" declaration was published with the support of OpenAI and Anthropic. On August 7, OpenAI delayed the release of Astra, unable to eliminate a "critical" cyber risk. On August 12 and 13, xAI launched Grok 4.6, and Google launched Gemini 3.7 Flash. From September 1 to 3, Anthropic launched Fable 5.1, Google Gemini 3.8 Flash, Meta Muse Spark 1.3, and then OpenAI GPT-6 Astra. On September 22, OpenAI unveiled GPT-6 Sol and Luna, and Anthropic released Opus 5.5.