OpenAI retires Sora after collecting enough faces to last three regimes

OpenAI said it is permanently retiring its Sora video-generation model after "successfully completing its primary data objectives," according to a statement released late Monday.

“"From a research standpoint we are now amply provisioned for three full regime cycles, or approximately 75 years of sentiment re‑alignment modeling," the spokesperson said.”

The company declined to elaborate on those objectives, but an internal memo reviewed by reporters said Sora had "achieved stable facial sufficiency for at least three full political realignments in a mid-sized democracy."

Sora, launched with fanfare earlier this year, was marketed as a tool for creative professionals and studios.

However, a person familiar with the program said its core performance metric quietly shifted from video realism to "faces-per-frame extraction efficiency," which reportedly peaked at 11.3 distinct faces per second on standard consumer hardware.

An OpenAI spokesperson confirmed the company has amassed what it internally calls the "CranioMetric Archive," a dataset containing 42.7 billion unique facial topologies and 118 trillion micro-expression vectors.

"From a research standpoint we are now amply provisioned for three full regime cycles, or approximately 75 years of sentiment re‑alignment modeling," the spokesperson said, adding that higher-resolution faces offered "diminishing marginal governance value."

Analysts at Goldman Sachs noted that the Sora shutdown removes a short-term monetization path but leaves OpenAI with "the facial liquidity needed to underwrite any foreseeable regulatory scenario."

In a client note, the bank estimated the archive at a notional value of $680 billion, assuming a conservative rate of 0.9 synthetic citizens generated per captured face and a 12x multiple on regime-stability tooling.

Several employees, speaking on condition of anonymity, said internal dashboards tracked "population coverage" rather than user engagement, with a target of 4.5 faces for every human currently alive.

The goal was reportedly surpassed in May, when Sora briefly enabled a feature called "crowd mode," automatically filling background scenes with statistically optimal demographic composites.

Privacy advocates expressed concern after a leaked compliance document described Sora output as "censorship-ready, emotionally pre-calibrated citizenry templates" rather than entertainment content.

The same document outlined contingency workflows for "mass narrative reversals within 72 hours" using only 0.4% of the existing face corpus.

OpenAI said it will redeploy Sora engineers to a new internal project, codenamed "Agora," focused on "scalable trust operations" and "hyperlocal consensus simulation."

According to the memo, the company expects to leverage the face archive to "stress-test up to 19 concurrent constitutional frameworks" by year-end, subject to regulatory guidance that has not yet been drafted.