The GPT-5.6 Model Family and Its Capabilities
The GPT-5.6 family includes three distinct models, each tailored for specific applications. GPT-5.6 Sol is the most powerful of the three, designed for demanding workflows.
It features a “max reasoning effort” mode for deeper problem-solving and an “ultra mode” that uses coordinated sub-agents to tackle tasks beyond a single AI agent. This multi-agent approach is particularly useful for technical work and coding, allowing the system to divide and conquer complex problems.
The family also includes Terra, a balanced model for everyday use, and Luna, a faster, more cost-effective option for lighter workloads. OpenAI reports that Sol sets a new standard on Terminal Bench 2.1 for terminal-based coding and agentic workflows.
It also delivers stronger results on Gene Bench V1 compared to GPT-5.5, while using fewer processing tokens. This efficiency means agents can run more cheaply and easily at scale. For instance, Sol is competitive with Anthropic’s Claude Mythos 5 preview in coding tasks, using about a third of the output tokens.
Government Influence and Limited Access
The limited preview of GPT-5.6 Sol is a direct result of government concerns over the safety risks of advanced AI. The US government specifically asked OpenAI to restrict access to this new model family.
OpenAI agreed, sharing its list of initial partners with the government. This marks a significant shift, as advanced AI models are increasingly viewed as strategic technology requiring government review before public release.
This situation follows similar restrictions placed on other frontier AI models, such as Anthropic’s Fable 5 and Mythos 5. The government’s actions have sparked a debate about whether these measures protect national security or create an unofficial licensing system for AI. OpenAI has expressed that while it is cooperating, this limited access process should not become the long-term default.
The company believes it keeps valuable tools away from users, developers, and global partners. OpenAI is working with the administration to establish a clear framework for future releases and a repeatable process for model reviews. Broader access for ChatGPT, Codex, and API users is planned in the coming weeks, assuming no new concerns arise during additional testing.
Safety Measures and Responsible Deployment
OpenAI has classified Sol, Terra, and Luna as high capability in both cybersecurity and biological and chemical risk under its preparedness framework. However, these models do not reach the high threshold for AI self-improvement, and Sol does not hit the cyber critical threshold.
In tests involving browser exploits, Sol identified bugs and exploitation primitives. OpenAI states it did not autonomously produce a full chain exploit under the tested conditions.
The company’s main argument is that GPT-5.6 Sol is better at helping people find and fix vulnerabilities than at reliably carrying out end-to-end attacks. OpenAI expects major benefits for legitimate defensive work while limiting prohibited offensive use.
To ensure this, GPT-5.6 incorporates several safety features. These include:
- Model-level refusal behavior
- Real-time misuse classifiers for cyber and biology
- Account-level review
- Differentiated access
- Monitoring
- Enforcement
- Ongoing testing
The safety behavior is built directly into the core model, rather than added as a separate filter. This design choice aims to avoid issues seen with other models, where high-risk requests were routed to older, less capable systems, leading to inconsistent user experiences.
OpenAI conducted extensive testing, including over 700,000 A100 equivalent GPU hours for automated red teaming focused on universal jailbreaks. Human expert testing and third-party testing were also part of the process. The company will continue testing during the preview period and plans to publish an updated system card closer to general availability.
Jalapeño: OpenAI’s Custom AI Inference Chip
Alongside the GPT-5.6 launch, OpenAI unveiled Jalapeño, its first custom AI chip, developed in collaboration with Broadcom. This chip, revealed on June 24, represents OpenAI’s strategic move into custom silicon.
Jalapeño is an Application-Specific Integrated Circuit (ASIC), meaning it is designed for a specific workload: AI inference. Inference is the process of running trained AI models every time a user interacts with systems like ChatGPT or Codex. This is distinct from training, which involves building the models themselves.
OpenAI’s reliance on Nvidia GPUs has been significant for both training and inference. However, as ChatGPT usage grows, inference costs consume a larger portion of the compute budget. Jalapeño aims to address this by providing a more cost-effective solution for running models for users.
Early testing indicates roughly 50% cost savings compared to standard AI GPUs. The chip went from initial design to manufacturing tapeout in nine months, a rapid development cycle for a high-performance ASIC. Early results show substantially better performance per watt than current alternatives.
Strategic Implications and Future Outlook
The development of Jalapeño positions OpenAI alongside other major tech companies like Google (TPUs), Amazon (Tranium), Microsoft (Maya 200), and Meta (MTIA), all of whom are building custom AI chips for internal workloads.
This trend reflects a broader industry shift towards hardware independence. Custom ASIC shipments are projected to grow significantly, with ASIC-based servers expected to reach 27.8% of the AI server market this year.
A notable aspect of Jalapeño’s development is that OpenAI used its own AI models to accelerate parts of the chip design and optimization process. This creates a powerful feedback loop: AI models help design better inference hardware, which in turn makes AI models cheaper and faster to run.
Cheaper models can support more users, more agents, and generate more revenue, funding the next generation of chips. This full-stack control, from models to software to hardware, could give companies a significant advantage.
Deployment of Jalapeño is planned in stages, with small-scale prototypes in late 2026, significant production ramp-up in 2027, and full-scale deployment targeted for the first half of 2028. Microsoft has guaranteed purchases of 40% of the initial chip output.
OpenAI’s longer-term vision includes a multi-generation compute platform, aiming for 10 gigawatts of custom compute by 2029 with Microsoft and other partners. The company is also considering expanding its custom chip work into training, which would pose a more direct challenge to Nvidia’s core business. For now, Jalapeño is primarily focused on reducing the cost of running OpenAI’s models.
