OpenAI gates GPT‑5.4
OpenAI released a cybersecurity‑tuned model, GPT‑5.4‑Cyber, but is restricting access to vetted security vendors, researchers and trusted programmes rather than offering it through the general API or ChatGPT flows. The rollout is tiered and paired with product packaging changes—OpenAI also documents fallback routing (GPT‑5.4 mini as a fallback and Auto routing options) that make entitlement and runtime routing part of the product surface. ( )
OpenAI has released GPT‑5.4‑Cyber, a cybersecurity model it is keeping behind a vetting gate instead of opening to all ChatGPT or application programming interface users. (openai.com) The company said on April 14 it is expanding its Trusted Access for Cyber program to “thousands” of verified individual defenders and “hundreds” of teams that protect critical software. Reuters reported the model is tuned for defensive cybersecurity work and arrived one week after Anthropic announced its own restricted cyber model, Mythos, on April 7. (openai.com, reuters.com) OpenAI said eligible users include security vendors, researchers, bug bounty hunters, maintainers of major open source projects, and nonprofit or public-interest groups working on cyber defense. Axios reported the company is using a tiered access system rather than a single public launch, with different permissions tied to trust level and use case. (openai.com, axios.com) In plain terms, OpenAI is treating access itself as part of the product. The model is not just a new engine; it is a new permission layer that decides who can use the more cyber-permissive version and under what controls. (axios.com, openai.com) That same logic now shows up in OpenAI’s broader product packaging. OpenAI’s model release notes say GPT‑5.4 mini is used in ChatGPT as a rate-limit fallback for GPT‑5.4 Thinking, and Enterprise customers can choose Auto routing to default to GPT‑5.4 mini. (help.openai.com) Those release notes also say GPT‑5.4 mini will not appear as a selectable model in the picker for most users, even though it may run behind the scenes after limits are hit. A separate ChatGPT help page says older ChatGPT models including GPT‑5 Instant and GPT‑5 Thinking were retired on February 13, 2026, while application programming interface access remained unchanged. (help.openai.com, help.openai.com) OpenAI introduced the base GPT‑5.4 model on March 5 as its flagship for professional work, with coding, tool use, computer use, and a one million token context window. GPT‑5.4‑Cyber is a specialized branch of that model, aimed at helping defenders find and fix software flaws faster without making the capability broadly available on day one. (openai.com, openai.com) Anthropic took a similar restricted path with Mythos Preview under a program called Project Glasswing, according to Reuters. OpenAI’s move shows the leading model labs converging on the same rollout pattern: release stronger cyber tools first to screened organizations instead of the general public. (reuters.com) For customers, that means the question is no longer only which model they bought. It is also which model they are entitled to, when the system silently falls back, and whether the most capable version is available only through a trust program instead of the standard product menu. (help.openai.com, axios.com, openai.com)
