OpenAI limits GPT-5.6 rollout after government request, says restrictions shouldn’t be the norm

OpenAI is limiting the merchandise of its newest AI models to a “small radical of trusted partners” astatine the behest of the U.S. government, the institution said Friday.

The adjacent procreation GPT-5.6 lineup includes Sol, its flagship model; Terra, a much balanced exemplary for mundane use; and Luna, a faster, little outgo option. Although Sol is the company’s astir almighty mode, the Trump medication has restricted the merchandise of each three. OpenAI said the preview is constricted to partners “whose information has been shared with the government.”

The administration’s request comes arsenic the US authorities puts caller unit connected AI companies to restrict their astir precocious systems. After Anthropic released its astir almighty nationalist exemplary Fable 5, the medication ordered the institution to region entree for immoderate overseas national, prompting Anthropic to instrumentality the exemplary down entirely. 

The incidental has brought up questions of however overmuch powerfulness the authorities should person implicit AI exemplary releases. Dean Ball, a erstwhile White House AI advisor and soon-to-be OpenAI employee, says President Trump’s recent enforcement order — which asks definite AI companies to voluntarily taxable their astir precocious models for authorities reappraisal up to 30 days earlier merchandise — has created a de facto involuntary licensing regime for frontier AI, starring to heavy-handed restrictions. 

The occupation compounds, Ball argues, erstwhile the authorities doesn’t person intelligibly defined information standards, which could pb to endless motorboat delays that mightiness not lone springiness a manus to China successful the AI race, but besides jeopardize the billions of dollars going to AI infrastructure buildouts. 

And portion OpenAI did arsenic the medication asked this clip around, the AI steadfast made it wide it wasn’t blessed with the arrangement.

“We don’t judge this benignant of authorities entree process should go the semipermanent default,” reads a Friday blog post. “It keeps the champion tools from users, developers, enterprises, cyber defenders, and planetary partners who request them.”

OpenAI called the preview a “short-term step” that volition enactment GPT-5.6 connected the way to broader availability successful the coming weeks, arsenic the institution works with the medication to make a caller enforcement bid model connected cybersecurity, arsenic good arsenic a “repeatable process for aboriginal exemplary releases.”

GPT-5.6 Sol specs

OpenAI says GPT-5.6 Sol is its strongest exemplary yet, with improved agentic capabilities successful coding, biology and cybersecurity. Sol introduces a “max” reasoning effort mode and an “ultra” mode that uses coordinated subagents to lick highly analyzable tasks (just the benignant of neat instrumentality that sends your token usage skyrocketing).

GPT-5.6 excels astatine respective benchmarks, says OpenAI, including being somewhat amended astatine coding workflows than Anthropic’s Claude Mythos 5, which the Trump medication besides efficaciously banned this month. OpenAI says GPT-5.6 Sol is besides competitory with Mythos preview, but uses a 3rd of the output tokens. 

To assuage immoderate fears of its almighty models being unsafe, OpenAI says Sol includes its astir robust information stack yet. It is, OpenAI says, heavy hardened against adversarial attacks and intentionally optimized to favour antiaircraft cybersecurity enactment implicit violative exploits. In different words, it’s designed to beryllium hard to jailbreak, portion prioritizing showing users however to support against exploits, alternatively than however to hack into systems. 

OpenAI besides says its information guardrails are built straight into the halfway model’s behavior, alternatively than relying connected a abstracted filter connected apical of it. The steadfast is apt trying to debar the trap that caught Anthropic with Fable 5. In the little moments erstwhile Fable 5 was available, whenever the model’s classifiers detected a high-risk topic— similar cybersecurity, biology, oregon chemistry — it wouldn’t conscionable artifact the prompt; it would way the petition to an older model. The full over-cautious travel and invisible downrouting led to galore mendacious positives and idiosyncratic backlash. 

While the GPT-5.6 models are initially disposable lone to a prime radical of partners, OpenAI plans to marque them much broadly disposable to radical utilizing ChatGPT, Codex, and the API soon. 

GPT-5.6 comes successful 3 sizes with tiered pricing: Sol costs $5 per cardinal input tokens and $30 per cardinal output tokens; Terra costs fractional that; and Luna costs $1 and $6, respectively. OpenAI says it has besides improved punctual caching to marque repeated prompts cheaper and much predictable.