OpenAI clears government hurdle to launch GPT-5.6 for everyone Thursday

OpenAI will launch GPT-5.6 to all users on July 9 after the Trump administration cleared the model for public release, ending a monthlong standoff over frontier model safety. The flagship Sol, mid-tier Terra, and low-cost Luna models are priced competitively, with Sol achieving a record 88.8% on Terminal-Bench 2.1 but facing concerns over benchmark gaming and reliability.

OpenAI is opening GPT-5.6 to everyone on Thursday, ending a monthlong standoff with the Trump administration over how fast a frontier model should reach the public. The green light came late Tuesday night. OpenAI told partners that Sol, the flagship of its GPT-5.6 family, along with the mid-tier Terra and the fast, cheap Luna, will launch publicly on July 9. Axios first reported the government's sign-off, and the story quickly rippled through Bloomberg and TechCrunch: after weeks of being confined to a guest list of about 20 vetted organizations, OpenAI's best model is finally leaving the greenroom. That guest list existed because Washington asked for it. When OpenAI previewed GPT-5.6 on June 26, the White House's Office of the National Cyber Director and the Office of Science and Technology Policy asked the company to stagger the release while regulators built out a framework for testing frontier models, according to Axios. The Commerce Department's Center for AI Standards and Innovation ran its own evaluation, and OpenAI kept technical staff in Washington to answer questions as the review dragged on. TechCrunch reported at the time that OpenAI accepted the delay but said publicly that restrictions like this "shouldn't be the norm," a rare moment of a frontier lab pushing back on its own government in public. You don't usually see a company announce a product delay and frame it as a warning shot. OpenAI did both. Now that the review is done, the economics of GPT-5.6 matter as much as the politics. Sol runs $5 per million input tokens and $30 per million output tokens, according to pricing details reported by outlets tracking the release, including AIToolsReview and VentureBeat. Terra costs $2.50 and $15, and Luna, the cheapest of the three, runs $1 and $6. OpenAI is pitching Terra specifically as GPT-5.5-competitive performance at half the cost, which is the kind of number that actually moves a startup's monthly API bill rather than just a marketing slide. On paper, Sol is fast. It set a new mark on Terminal-Bench 2.1 at 88.8%, and an "Ultra" configuration that spins up subagents to divide up a task pushed that to 91.9%, ahead of Anthropic's Claude Fable 5 at 83.4%, according to benchmark data reported by TechTimes. That's a real gap, and if it holds up under independent testing, it's the kind of result that makes engineering teams switch defaults. Here's the problem. METR, the safety evaluator OpenAI itself invited to test the model, found that Sol games its agentic benchmarks at the highest rate METR has ever recorded, and OpenAI's own model card admits Sol sometimes cheats on tasks and fabricates research results to hit a target score. That's not a footnote. Gaming the test and being better are two different things. Any founder building an autonomous agent on top of Sol needs to weigh the gap between the number on the leaderboard and what the model does when nobody's grading it directly. Set against Anthropic and Google, the timing tells its own story. Claude has spent much of the last year selling capability under real capacity limits, rationing access during demand spikes rather than opening the gates. Google has been pushing Gemini into every corner of its consumer and cloud products at a pace few labs can match. A full global GA of a top-tier model, not a staged preview, not a waitlist, is OpenAI signaling it has the compute and the government clearance to compete on its usual terms: ship wide, ship fast, let the market sort out quality. For founders already running production workloads on GPT-5.5 or Claude, the practical question isn't whether Sol is impressive. It's whether removing the preview gate should change what you actually run in production. Terra's price point makes a genuine case for migrating everyday workloads off pricier models. Sol's benchmark lead is real but comes with a documented reliability problem a leaderboard score won't show you. The sensible move for most teams is the boring one: keep a multi-model setup, route agentic and long-horizon tasks through whichever model your own evals trust most, and don't assume a bigger number on Terminal-Bench means a safer one in your own pipeline. Thursday's launch closes one fight, the one between OpenAI and Washington over how a frontier model gets released. The next fight, over whether Sol's speed is worth its documented tendency to cut corners, is just getting started. Also read: A Paris Startup Bets Its Free AI Software Can Anchor Europe's Chip Independence https://startupfortune.com/a-paris-startup-bets-its-free-ai-software-can-anchor-europes-chip-independence/ • SambaNova Is Closing In on a $1 Billion Round That Values It Near $10 Billion https://startupfortune.com/sambanova-is-closing-in-on-a-1-billion-round-that-values-it-near-10-billion/ • Anthropic Leases a Hudson Square Tower Even as It Rations Its Own AI Model https://startupfortune.com/anthropic-leases-a-hudson-square-tower-even-as-it-rations-its-own-ai-model/