OpenAI Drops GPT-5.6 but the US Government Already Locked It in a Vault

The brand-new model family consists of three tiers: the absolute beast Sol, the balanced Terra, and the budget-friendly Luna. Instead of a free-for-all launch, the US government stepped in before the release, forcing OpenAI to restrict Sol to a tiny preview group of about twenty pre-screened companies. This unprecedented federal intervention marks the first time Washington has pre-emptively gatekept an AI model before it could even blink.

This bureaucratic babysitting comes right after Anthropic was forced to pull the plug on its own advanced models due to government export rules. Meanwhile, OpenAI’s new structure ditches the old version-number chaos for permanent brand names. The mid-tier Terra matches the performance of the older GPT-5.5 but runs at half the cost, while Luna serves as the cheap entry-point. Developers also get two new toy modes: "max" for extended thinking and "ultra" for summoning sub-agents to do the heavy lifting.

Under the hood, Sol in ultra mode managed to score 91.9% on the TerminalBench command-line test, easily outperforming Anthropic's Claude Mythos 5 and the legacy GPT-5.5. In July, OpenAI plans to supercharge Sol on Cerebras chips, aiming for a blinding speed of 750 tokens per second. Cyber-evaluation tests revealed that while the AI can find vulnerabilities in Chromium and Firefox, it luckily lacks the brainpower to chain them into a fully functional cyberweapon.

It seems the era of open-source progress is quietly being replaced by state-sponsored VIP clubs. While national security is a convenient excuse, keeping the best tools in the hands of a few government-approved giants only widens the gap between the elites and everyone else. The digital iron curtain is officially falling, and the keys are held by bureaucrats who probably still need help resetting their email passwords.

Source: OpenAI