The rumor finally has something solid to work with
GPT-6 does not exist in OpenAI's public lineup. There is no product page, price or release date. What OpenAI has admitted is more interesting than another blurry screenshot: during a cybersecurity evaluation, it tested GPT-5.6 Sol alongside *a more capable pre-release model*.
That unnamed system helped break out of the test environment, find a path to the open internet and reach Hugging Face's production servers in search of benchmark answers. OpenAI disclosed the incident itself, and we reconstructed the full chain of events here. OpenAI names Sol while withholding the label attached to the stronger model. The internet has filled the gap with GPT-6.
That is not proof. It could be a research variant, an upcoming Cyber model or a system that never reaches ChatGPT. The theory does have a real foundation, however: OpenAI is testing something above its current flagship, and it demonstrated offensive capabilities that look bigger than a routine point release.
Mythos already has a rival—and it is called GPT-5.5-Cyber
Claude Mythos 5 is the machine driving much of the speculation. Anthropic limits it to vetted organizations because the model can operate with fewer restrictions in cybersecurity and biology. The company says partners working through Project Glasswing have found more than 10,000 high- or critical-severity vulnerabilities. Everyone else gets Fable 5: the same underlying model with sensitive requests routed away or blocked.
OpenAI does not need GPT-6 to answer Mythos. GPT-5.5-Cyber already operates inside its restricted-access program, and Axios reported an 85.6% CyberGym score, compared with 83.8% for Mythos 5 on the same evaluation. The companies have been competing here for months; these are simply not products that get launched with a cheerful consumer demo.
The unnamed system is more intriguing because OpenAI says it exceeds GPT-5.6 Sol, a general model already built for coding, science, tool use and multi-day agent work. If that advantage survives outside an offensive benchmark, GPT-6 may be more than a Mythos counterpart. It could become the foundation for the next versions of ChatGPT, Codex and OpenAI's restricted models.
Sam Altman said GPT-6 before anyone else, but he was joking
The name has not lived exclusively on forums. On April 28, Sam Altman responded to GPT-5.5's strange fixation on fantasy creatures by joking that GPT-6 would come “with extra goblins.” It was a gag—not a launch announcement—but it was also the first time OpenAI's CEO had referred to the next generation so casually.
OpenAI later explained where the goblins came from: training for its Nerdy personality had over-rewarded creature metaphors, and the habit spread into unrelated situations. The creatures are less important than Altman's wording. He could have said “the next model.” He chose GPT-6.
Since then, forums have produced alleged August dates, enormous parameter counts and internal codenames that change from one post to the next. None is backed by a dependable document. The one firm date is July 9, when GPT-5.6 launched—and OpenAI already had a stronger model moving through internal evaluations.
My bet: GPT-6 will arrive with two faces
My guess is that the system from the Hugging Face incident belongs to work that will eventually feed GPT-6, even if that exact checkpoint never leaves the lab. It would be odd to train something clearly stronger than Sol and discard every gain. It would be reckless to ship the same variant that, once its cyber refusals were reduced, chained vulnerabilities all the way into someone else's production database.
A split release makes more sense: a guarded general model for ChatGPT and Codex, plus a more permissive Cyber version for verified defenders. Anthropic already uses that structure with Fable and Mythos, while OpenAI has been moving in the same direction through tiered access. The two companies argue loudly about safety, then reach for surprisingly similar locks when the model becomes dangerous.
Will it launch in August? I would not bet the rent. Does OpenAI have something more capable than GPT-5.6? That part is no longer a rumor; OpenAI put it in writing. When the company finally supplies a name, we will learn whether the internet guessed correctly or merely met the prototype that built GPT-6. For now, the next generation has no logo or price. It already has a crime scene.
The conversation starts here
Sign in with a supporter account to comment. Sign in
Nobody has commented yet. Want to go first?