OpenAI Gives Select Partners Access to GPT-5.6 Sol After Coordinating With Washington

OpenAI has announced a limited preview of its new GPT-5.6 model family, made up of three variants: the flagship Sol, the balanced Terra, and the fast, cheaper Luna. Rather than rolling out to all ChatGPT users at once, the models were first made available to a narrow group of trusted partners through the API and Codex, with broader access expected in the coming weeks.

Three models, three budgets

Sol is OpenAI's most powerful model to date, priced at $5 per million input tokens and $30 per million output tokens. Terra is meant to offer performance close to GPT-5.5 at half the price, $2.5 per million input tokens and $15 per million output tokens. Luna, the cheapest of the three, costs $1 per million input tokens and $6 per million output tokens, which OpenAI describes as its lowest price point ever.

The release also introduces two operating modes: max, built for extended reasoning over hard problems, and ultra, which coordinates multiple subagents at once on multi-step tasks. Sol set a new record on Terminal-Bench 2.1, which tests command-line work, and on GeneBench v1, a genomic analysis benchmark, while using fewer tokens than GPT-5.5.

The government saw it before users did

The most unusual part of the launch is how it was rolled out. Under the Trump administration's guidelines for AI cybersecurity, companies are required to present their most powerful models to the US government roughly 30 days before public release. OpenAI briefed the government on Sol before the preview began, and the list of partners taking part in the test was shared with government officials.

OpenAI stresses that this arrangement is temporary and that the model is ultimately meant to reach standard public availability without such intermediation. The company describes Sol as its most capable model yet on cybersecurity, better at finding and fixing vulnerabilities than at independently carrying out full attacks, and says the model does not cross its internal risk threshold defined as Cyber Critical.

Weeks of testing

Ahead of the launch, OpenAI put in more than 700,000 A100-equivalent compute hours testing the model's resistance to misuse. The safeguards operate on several layers at once: training the model to refuse prohibited actions, real-time classifiers, manual account verification, and access levels that vary by query type. Browser-based tests uncovered bugs in code but did not result in working, complete exploits.

This level of caution reflects growing political pressure around models capable of independently finding and exploiting security vulnerabilities. OpenAI openly acknowledges that Sol pushes the frontier on long-running vulnerability research tasks, making it both a useful tool for security teams and a potential risk if it falls into the wrong hands.

Speed as a product of its own

Alongside the preview, OpenAI confirmed that Sol will run on specialized Cerebras hardware starting in July, reaching speeds of up to 750 tokens per second there. That is roughly ten times faster than typical GPU clusters, which usually deliver between 40 and 120 tokens per second for frontier-class models. In practice, that means tasks like generating a 4,000-token pull request could take seconds instead of minutes, which matters a lot for coding agents working iteratively.

The partnership with Cerebras rests on a multibillion-dollar contract under which OpenAI has secured more than $20 billion in compute capacity and 750 megawatts of inference infrastructure spread across several years. Running Sol on this hardware is also meant to serve as a public proof point for Cerebras technology ahead of the company's planned IPO.

What it means for businesses in Poland

For Polish development teams and companies using the OpenAI API, price and availability will matter more than the launch itself for now. Terra, offering GPT-5.5-level performance at half the cost, could quickly become the default choice for everyday business tasks, while Sol will remain reserved for a narrow set of partners subject to extra access controls. It's also worth watching how quickly OpenAI opens up Sol access in Codex, since that's where teams working with coding agents will feel the difference first.

The mechanism of previewing a model to the US government before release also hints at where regulation of the most powerful AI models may be heading globally, including in the European Union, where work is underway in parallel to implement the AI Act's rules for high-risk systems. A model that runs faster and cheaper will only reach the market after passing through such a filter, a pattern that could become the norm outside the United States as well.

Sources: Previewing GPT-5.6 Sol (openai.com), OpenAI launches a limited preview of GPT-5.6 (engadget.com), OpenAI GPT-5.6 Sol Launched: Benchmarks, Cerebras Speed, and Dark Secrets Revealed (techgenyz.com), OpenAI to run GPT-5.6 Sol on Cerebras at 750 tokens per second in July (aesopacademy.org).