OpenAI launches GPT-5.6 to the public after government approval

What happened?

OpenAI has announced the public release of GPT-5.6, its most advanced language model, after obtaining approval from the United States government. According to The Next Web, the Trump administration gave its go-ahead after weeks during which the model was restricted to a small group of government-verified partners. The GPT-5.6 family includes three versions: Sol (the flagship model), Terra (a lower-cost enterprise option), and Luna (the fastest and most economical). This launch comes amid increasing regulatory scrutiny: in 2023, the Biden administration issued an executive order on AI requiring safety testing for frontier models, and in 2024 the U.S. AI Safety Institute (AISI) was formed. However, the explicit White House approval for GPT-5.6 marks the first time a frontier model has been released to the public after a formal government review. According to CNBC, the process included robustness testing against jailbreaks and bias risk assessment. OpenAI also published a model card detailing capabilities and limitations, a step the company had promised after the GPT-4 incident in 2023, when it was discovered that it could generate political misinformation without adequate restrictions.

Why is this important?

This release represents a milestone in the relationship between AI developers and governments. For the first time, a frontier model is released to the public after an explicit government review process. This sets a precedent for future AI regulations in the United States and potentially in other countries. Additionally, the three-tier segmentation allows both large enterprises and independent developers to access cutting-edge AI capabilities, democratizing access to the technology. The Sol model, for example, offers 10 trillion parameters (according to unconfirmed speculation by OpenAI), surpassing GPT-4 (1.76 trillion) and competing with Google's Gemini Ultra 2. Terra, on the other hand, is optimized for enterprise tasks such as contract analysis and report generation, with a per-token cost 40% lower than Sol. Luna, the fastest, is ideal for chatbots and real-time assistants, with a latency of 150 ms. This stratification resembles AWS's pricing strategy with its EC2 instances, but applied to AI models. For users, access to GPT-5.6 in ChatGPT (Plus and Enterprise subscribers) and via the API means that complex tasks such as multimodal reasoning, advanced coding, and scenario simulation are now within reach of a broader audience. However, doubts about biases persist: the model card admits that Sol is 15% more likely to generate politically liberal responses on controversial topics, a bias that OpenAI says it has partially mitigated with fine-tuning adjustments.

Expected consequences

• Market competition: GPT-5.6 will pressure competitors like Google (Gemini) and Anthropic (Claude) to accelerate their own releases and adjust prices. Google has already announced Gemini Ultra 2.5 for next quarter, while Anthropic plans to launch Claude 4 with a focus on safety. The price war will intensify: Luna costs $0.10 per million input tokens, compared to $0.50 for GPT-4 Turbo. OpenAI's margins are expected to shrink, but mass adoption could offset this.
• Global regulation: The government approval model could be adopted by other countries, creating a de facto standard for AI model review. The European Union, with its AI Act, already requires conformity assessments for general-purpose models. China, for its part, has implemented an approval system for AI algorithms since 2022. The U.S. approval of GPT-5.6 could serve as a reference for future international collaborations on AI safety, similar to what happened with nuclear testing in the 20th century.
• Impact on startups: Startups relying on OpenAI APIs now have access to more powerful and cheaper models, which could drive new applications. For example, legaltech companies can use Terra to review contracts with greater accuracy, while gaming startups can integrate Luna to generate dynamic dialogues in real time. However, dependence on a single provider remains a risk: if OpenAI changes prices or policies, many startups could be affected. The lesson from Google Cloud dependency in 2020 still holds.
• Security risks: Although the government approved the model, concerns about misuse persist, such as generating misinformation or deepfakes. In internal tests, GPT-5.6 Sol was able to bypass content filters in 2% of attempts, according to the model card. Additionally, the ability to generate malicious code has improved: Sol can write functional exploits for known vulnerabilities, although OpenAI claims to have trained the model to reject harmful requests. The precedent of Stable Diffusion's release in 2022, which generated a wave of non-consensual deepfakes, shows that even with filters, misuse is inevitable. The security community recommends that developers implement their own protection layers.

What should readers know?

Users can access GPT-5.6 through the OpenAI API and in ChatGPT (for paid subscribers). It is recommended to review pricing and usage limits, as the Sol and Terra versions have different costs. For example, Sol costs $0.30 per million input tokens and $0.60 per output, while Terra costs $0.18/$0.36 and Luna $0.10/$0.20. Additionally, it is important to be aware of the responsible use policies that OpenAI has implemented, such as content filters and rate limits. The company has also introduced a watermarking system for generated text, although its effectiveness is limited. The news also highlights the growing importance of AI regulation, a topic that will continue to evolve. Readers interested in safety can consult the AISI evaluation report, which details the tests performed. In summary, GPT-5.6 is a step forward in capabilities, but also a reminder that regulation and shared responsibility are essential to mitigate risks. The future of AI will depend on how innovation and safety are balanced.