OpenAI officially released its latest AI model family, GPT-5.6, worldwide on July 9 (local U.S. time). What makes this launch notable is not just performance. It is, in effect, the first commercial AI model whose safety and security were vetted directly by the U.S. government before release — a case that shows how AI development and government oversight are beginning to intersect. 🤖

Summary (TL;DR)

  • OpenAI released three GPT-5.6 models — Sol, Terra, and Luna — globally on July 9. After unveiling them on June 26 and initially supplying them only to select partners at the U.S. government’s request, the Trump administration lifted broad restrictions and shifted to full access.
  • Before release, the AI Standards and Innovation Center (CAISI) under the U.S. Commerce Department handled security-vulnerability testing, and OpenAI stationed technical experts in Washington, D.C. to respond to government inquiries. It is the first model to undergo government pre-review under President Trump’s cyber executive order.
  • Pricing per 1M tokens is $5 input / $30 output for Sol, $2.50 / $15 for Terra, and $1 / $6 for Luna. Terra delivers performance on par with the previous generation, GPT-5.5, at half the cost.

🏛️ Why Did GPT-5.6 Go Through Government Review First?

The standout element of this launch is “government pre-review.” On June 26, OpenAI unveiled three models: the flagship “Sol,” the balanced everyday-work model “Terra,” and the fast, economical “Luna.” At the U.S. government’s request, however, it initially opened access only to a handful of trusted partners. Then, on July 7, an Axios report revealed that the Trump administration had lifted broad release restrictions, and OpenAI promptly announced via X that a global launch would follow on Thursday (the 9th).

The review was carried out at the AI Standards and Innovation Center (CAISI) under the U.S. Commerce Department. OpenAI went through additional testing and meetings with government officials, and afterward kept technical experts in Washington, D.C. to answer potential questions. According to OpenAI, this is part of the process of shaping President Trump’s cyber-related executive order framework, and the company also stated its position that “this kind of government access process should not become the long-term default.” Anthropic’s “Fable 5” had earlier gone through a similar lifting of restrictions, suggesting that coordination between government and industry around frontier models is settling into a pattern.

🔀 How Do Sol, Terra, and Luna Differ?

GPT-5.6 introduces a new naming system that uses the number for the generation and the name for the capability tier. The number “5.6” denotes the model generation, while Sol, Terra, and Luna each mark a durable tier that varies in intelligence, speed, and cost.

  • Sol: The top flagship model. It aims for state-of-the-art performance in coding, science, and cybersecurity, and adds a new “max” reasoning option for deep, extended reasoning as well as an “ultra” mode that uses subagents to speed up complex work.
  • Terra: A balanced model optimized for everyday work. It maintains performance comparable to the previous generation, GPT-5.5, while cutting cost roughly in half.
  • Luna: The fastest and cheapest model, notable for delivering solid performance even at a low price.

Pricing per 1M tokens is $5 input / $30 output for Sol, $2.50 / $15 for Terra, and $1 / $6 for Luna. OpenAI said it would initially provide access through the API and Codex to a select group of trusted partners, then gradually expand to ChatGPT, Codex, and the API. It added that it would serve Sol at up to 750 tokens per second on Cerebras infrastructure during July.

🛡️ The Key Point: Both Performance and Safety Were Raised

OpenAI highlighted performance gains and stronger safeguards together. On performance, Sol set a new record on Terminal-Bench 2.1, which evaluates command-line workflows, and on GeneBench v1, which covers long-horizon genomics analysis, it produced better results than GPT-5.5 while using fewer tokens. It also showed improved capabilities on the cybersecurity benchmarks ExploitBench and ExploitGym.

At the same time, OpenAI said it applied its most robust safety stack to date. It devoted more than 700,000 A100-equivalent GPU hours to automated red-teaming to find so-called “universal jailbreaks” that work across many prompts and contexts, and combined multiple layers of defense — training the model to refuse, real-time misuse classifiers, and account-level review. That said, in evaluations targeting Chromium and Firefox, Sol identified bugs and the building blocks of exploits but did not autonomously produce a full attack chain under the test conditions, and the company said it did not cross the “Cyber Critical” threshold under its Preparedness Framework. Even so, given the large jump in capability, OpenAI’s policy is to pair a phased release with strengthened safeguards.

📌 Takeaway

The GPT-5.6 launch shows two trends at once. One is that the performance race toward specialized domains such as coding, science, and cybersecurity has intensified; the other is that frontier models are now beginning to pass through the gate of government review. In particular, Terra delivering previous-generation-class performance at half the cost signals that cost efficiency has become as important a competitive axis as raw performance.

Three points are worth watching. First, whether government pre-review remains a one-off or becomes an institutionalized, repeatable process. Second, how large a gap Sol, Terra, and Luna show against rival models in real user experience and benchmarks. Third, whether stronger safeguards end up excessively blocking legitimate security and development work. Since OpenAI itself said it would examine this during the preview period, the coming weeks — as real-world use cases accumulate — will be the point to watch.

※ This article is for informational purposes only and is not investment advice.

Sources