Why wasn't the model public right away? Trump's executive order changed the rules

Behind the unusual process lies the executive order that President Donald Trump signed in early June 2026. It calls on AI companies to voluntarily submit their most powerful models to the government for a 30-day review before public release. OpenAI decided to follow the rules — not out of enthusiasm, but pragmatically. "We do not think this kind of state oversight should become a long-term standard. It prevents users, developers, and businesses from accessing the best tools," the company stated in its official blog post. It added, however, that this was the fastest way to get the models to the public.

Testing was entrusted to the Center for AI Standards and Innovation (AISI), which falls under the U.S. Department of Commerce. OpenAI also sent its own engineers directly to Washington so they could respond immediately to questions. The result? The model passed and was released even before the full 30 days elapsed.

Three models, three different tasks

This time, OpenAI didn't deploy a single model but three at once — each targeting a different type of user.

• Sol — the flagship. The most powerful OpenAI model ever, built for complex programming, scientific research, and cybersecurity. It features an "ultra" mode that uses subagents for parallel task processing and a new maximum reasoning mode.
• Terra — the golden middle. Offers performance comparable to the previous flagship GPT-5.5, but at roughly half the price. Ideal for enterprise operations, customer support, and document analysis.
• Luna — speed and savings. The cheapest and fastest variant for routine tasks such as text summarization or simple queries.

Pricing: what you'll pay for tokens

Pricing depends on the volume of processed tokens. To give you an idea — a million tokens corresponds to roughly 750,000 words, or several thick books. For processing such a volume, Sol charges $5 for input and $30 for output. Terra costs half as much as GPT-5.5, with specific pricing depending on usage volume. Luna comes in at $1 per million input tokens and $6 for output.

For regular ChatGPT users, the models remain available within the existing Plus ($20/month), Pro ($200/month), Business, and Enterprise plans. Developers gain access via the API.

Benchmarks: where GPT-5.6 shines

OpenAI published results across several key tests, and the numbers are impressive.

In programming, GPT-5.6 Sol set a new record in the Terminal-Bench 2.1 evaluation, which tests the model's ability to work in the command line — planning, iterating, and coordinating tools across complex workflows.

In biology, the model excelled in GeneBench v1, a benchmark focused on genomic analysis and quantitative biology. Sol achieved better results than GPT-5.5 while using fewer tokens — meaning more efficiently.

The biggest attention, however, goes to cybersecurity performance. In the ExploitBench test, Sol is competitive with Anthropic Mythos Preview — a yet-unreleased model from competitor Anthropic — while using only a third of the output tokens. In the ExploitGym benchmark, created by UC Berkeley researchers in collaboration with OpenAI and other labs, all three variants show significant improvement across increasing reasoning levels.

In healthcare, Sol achieved a length-adjusted score of 60.5 in HealthBench Professional — an increase of 8.7 points over GPT-5.5 and the biggest leap since the original GPT-5 was released.

Safety first: 700,000 GPU hours against hackers

With a model possessing such advanced cybersecurity capabilities, the safety question is critical. OpenAI therefore deployed its most robust protection yet. Automated red teaming — simulated attacks searching for weaknesses — consumed over 700,000 GPU hours converted to A100.

Protection is layered: the model is trained to refuse prohibited activities (including attempts to disguise intent), real-time abuse classifiers run for cybersecurity and biology, and account-level monitoring watches over conversations. When the model detects a potential violation, generation pauses and the output is checked by a larger reasoning model — only then does it reach the user.

OpenAI notes that GPT-5.6 does not exceed the "critical" cybersecurity threat threshold under its own Preparedness Framework. The model can identify bugs and basic exploit primitives in Chromium and Firefox, but did not independently create a functional exploit across the full chain. "It better helps people find and fix vulnerabilities than reliably carry out attacks," OpenAI summarizes.

What this means for Czech users and businesses

GPT-5.6 is available immediately for Czech users as well — via ChatGPT (web and mobile apps) and via the API for developers. The models also communicate in Czech, and based on experience with the GPT-5.x series so far, the quality of Czech responses is at a very good level.

For Czech businesses, Terra is particularly interesting — previous-generation performance at half the price means that integrating AI into business processes (customer support, document analysis, automation) has suddenly become more economically accessible. Luna then opens up possibilities for large-scale routine deployment where every crown matters.

From a European perspective, however, regulation must also be considered. The executive order in the US shows that government oversight of AI models is a reality — and the AI Act is already in force in the EU, bringing its own set of requirements. OpenAI must also meet these standards when deploying in Europe, which could affect the speed of availability for some advanced features.

ChatGPT Work and other news

Along with the release of GPT-5.6, OpenAI also announced ChatGPT Work — a new feature that shifts ChatGPT from a conversational assistant toward a full-fledged work tool capable of independently solving complex tasks across applications. Details are not yet known, but it is a clear signal that OpenAI is heading into the segment where agent-based solutions from Anthropic and Google already operate today.