OpenAI Releases GPT-5.6 After Washington-Forced Delay

OpenAI officially released the GPT-5.6 model family on July 9, ending weeks of uncertainty triggered by a launch delay requested by the US administration. The company confirmed that before the public rollout, the model was given to vetted government partners for additional safety testing.

The decision to hold back the launch was made a few weeks earlier, when the Donald Trump administration asked OpenAI for extra time to assess potential national security risks. Rather than shipping to all users at once, the model was first made available exclusively to vetted partners, with test results shared with government institutions.

Why the government stepped in

The reason for the caution was that OpenAI describes GPT-5.6 Sol as the company's strongest model yet in cybersecurity. The model can perform threat modeling, analyze and patch code for vulnerabilities, search for security flaws, and support blue teaming, meaning simulating attacks on one's own infrastructure to catch weak points before criminals do.

A tool of this class in the hands of bad actors could theoretically speed up the discovery of flaws in critical systems, which is why, before the model reached wide distribution, US agencies wanted to check its capabilities themselves. In the end, the launch proceeded without functional restrictions, and the model is available in ChatGPT, Codex, and through the OpenAI API.

54% more token-efficient on coding tasks than the previous generation - Sam Altman, OpenAI

Three variants, three budgets

The split into Sol, Terra and Luna lets companies match cost to task. Sol, the most expensive and most powerful variant, costs $5 per million input tokens and $30 per million output tokens. Terra, built for everyday office and coding work, is noticeably cheaper at $2.5 and $15. Luna, the budget option for simpler tasks and large-scale use, costs $1 and $6.

This pricing structure reflects the strategy OpenAI has used across several model generations: instead of a single universal product, the company offers a ladder of variants so enterprise customers can route simple queries to the cheaper model while sending complex analytical or coding tasks to the more powerful Sol.

What it means for businesses and developers

For security teams and IT departments, the GPT-5.6 launch signals that language models are increasingly moving into territory once reserved for specialized analytical tools. The ability to automatically detect vulnerabilities in code and simulate attacks could speed up the work of security teams, but it also raises questions about who should have access to such powerful offensive tools, and under what conditions.

For developers, the claimed improvement in token efficiency on coding tasks is the key detail, since it directly translates into the cost of using the model in tools like Codex or in code-editor integrations. At rates of a few dollars per million tokens, a 54 percent reduction in token usage makes a real difference in the bills of companies running the model at scale.

A precedent for future launches

The mere fact that a government administration asked a technology company to hold back the launch of an AI model pending a national security review is a precedent worth noting regardless of the outcome. Until now, this kind of pressure mainly concerned chip exports or foreign access to models, not the timing of a product's public launch in the United States itself.

It shows that as models gain real offensive capabilities in cybersecurity, they may fall under the same scrutiny historically applied to dual-use technologies. Future launches of similarly capable models, from both OpenAI and its competitors, may now go through a comparable vetting process before public release.

Sources: OpenAI prezentuje GPT-5.6. Nowa rodzina modeli AI stawia na kodowanie i cyberbezpieczeństwo (itwiz.pl), OpenAI w końcu udostępnia GPT-5.6. Znamy datę premiery nowych modeli (android.com.pl)