Held At The Door
OpenAI is expanding its offerings throughout the stack, but Washington decides who gets access.
Last week produced the two most consequential product announcements of OpenAI’s year, separated by four days. On Tuesday, the company and Broadcom unveiled a custom inference chip. On Friday, OpenAI released its most capable model family. The two products span the full AI stack from silicon to frontier model, and neither has reached the public.
Government preview
GPT-5.6 arrives as three models under a new naming convention. Sol, the flagship, introduces “max” and “ultra” reasoning modes. The ultra mode deploys subagents to parallelize complex work, and Sol scored 91.9% on Terminal-Bench 2.1, a coding benchmark that tests command-line workflows requiring planning and tool coordination. OpenAI claims that Sol approaches Claude Mythos Preview on ExploitBench, a cybersecurity evaluation, while consuming roughly a third of the output tokens. The system card classified all three models at the “High” risk level for both cyber and biological capabilities, the first time an entire model family has received that designation. Terra offers GPT-5.5-level performance at $2.50 per million input tokens, half of GPT-5.5’s price, and Luna targets high-volume workloads at $1 per million input tokens. The generation number identifies the model vintage, while the tier names persist across updates and can advance independently.
Approximately twenty organizations can use GPT-5.6 during the preview period, each approved by the U.S. government. That arrangement follows the executive order that President Trump signed on June 2, which directs federal agencies to collaborate on benchmarking frontier AI models before wide release and requires developers to provide the government with up to thirty days of access to their most capable systems. GPT-5.6 launched explicitly under that framework, and it became the second frontier release this month to face government-imposed access restrictions, after the Commerce Department suspended Anthropic’s Fable 5 and Mythos 5 on June 12.
OpenAI pushed back in its announcement, arguing that government-gated access withholds frontier tools from the developers, enterprises, and cyber defenders who need them most. The company characterized its compliance as a short-term step and indicated that it expects to expand access to additional organizations this week, with general availability across ChatGPT, Codex, and the API in the coming weeks. A classified process for evaluating frontier models’ cyber capabilities, also mandated by the executive order, must take shape by August 1. How that process treats the next frontier release will determine whether ad hoc government gates harden into a standard feature of deployment.
Owning the stack
Two days before the model launch, OpenAI and Broadcom unveiled Jalapeño, an inference chip that OpenAI designed from scratch and brought from concept to manufacturing tape-out in nine months. Broadcom described that timeline as possibly the fastest ASIC development cycle in high-performance semiconductors. OpenAI’s own models accelerated parts of the chip’s design and optimization. Richard Ho, who leads OpenAI’s hardware program, said that the team optimized the architecture around the kernels, memory movement, and serving patterns specific to frontier LLM inference. Separately, OpenAI announced that Sol will launch on Cerebras hardware in July at up to 750 tokens per second, a speed that current GPU-based serving cannot match.