BODY: The AI arms race just took another sharp turn. On June 26, 2026, OpenAI announced its GPT-5.6 series, headlined by a top-tier model called GPT-5.6 Sol that—according to the company's own benchmarks—surpasses Anthropic's flagship Claude Mythos 5. There's just one catch: almost nobody can actually use it yet.
The GPT-5.6 lineup follows a celestial naming scheme, with Sol sitting at the top alongside reported siblings named for Terra and Luna. Sol is positioned as OpenAI's most capable reasoning model to date, and the company claims it pushes past Claude Mythos 5 on the metrics that the two labs have spent the past year trading blows over.
But unlike a typical splashy OpenAI launch, GPT-5.6 Sol is not generally available. At the time of reporting, the model is offered only as a limited preview to a small set of organizations, rather than rolling out to the broad ChatGPT and API user base that has come to expect day-one access.
The reason cited is notable: the restricted release reportedly comes at the direction of the U.S. government. That framing places GPT-5.6 Sol squarely in the growing conversation about frontier-model governance, export sensitivity, and who gets to wield the most powerful systems first—a conversation that has shifted from hypothetical to operational over the past year.
The insider take
From Tokyo, the pattern is familiar and a little frustrating. Japanese enterprises and developers have repeatedly found themselves at the back of the line for frontier AI—first for capacity reasons, now increasingly for regulatory ones. A government-gated "limited preview" model signals that access to the absolute cutting edge is becoming a geopolitical privilege, not a market product. For Japan's AI ecosystem, which leans heavily on both OpenAI and Anthropic APIs, the takeaway is clear: the headline benchmark wins matter less than whether you can ever get your hands on the model. Expect renewed local interest in domestic and open-weight alternatives as a hedge.
Originally reported by GIGAZINE (Japanese).