OpenAI is retiring gpt-5.4-cyber on 1 October 2026: here's what you need to do
gpt-5.4-cyber is deprecated with a hard removal on 1 October 2026. The replacement is gpt-5.6-cyber, but migration requires Daybreak Red approval.
OpenAI has formally deprecated gpt-5.4-cyber, with a hard removal date of 1 October 2026. From that date, any API call referencing the old model ID will return an error. The documented replacement is gpt-5.6-cyber, though as you’ll see below, switching is not as simple as updating a string in your config.
What is being removed and when
The gpt-5.4-cyber model ID is deprecated with immediate effect and will be completely inaccessible from 1 October 2026. OpenAI’s standard policy for generally available models is to provide at least six months’ notice before a hard shutdown, so the timeline here is consistent with how the company has handled previous deprecations. Affected developers should have received an email notification from OpenAI, and the change is documented on the official deprecations page.
It is worth noting that 1 October 2026 sits inside a particularly busy deprecation window. The Assistants API was discontinued on 26 August 2026, and a broader legacy model sunset follows on 23 October 2026. If you maintain multiple integrations built on OpenAI’s API, this period warrants a thorough audit of everything in production.
What replaces it
The documented replacement is gpt-5.6-cyber, a purpose-trained cybersecurity model built on GPT-5.6 Sol. It is available through Daybreak Red, the higher of the two tiers in OpenAI’s Daybreak programme.
The capability jump is significant. In OpenAI’s internal benchmarking, gpt-5.6-cyber completes 95.0% of advanced cybersecurity task requests. For context, gpt-5.6-cyber’s predecessor gpt-5.5-cyber completed 57.3% of those same requests, and GPT-5.6 Sol without any specialist access managed just 1.5%. Under OpenAI’s Preparedness Framework, gpt-5.6-cyber has been assessed as reaching the High cybersecurity capability threshold, though not Critical.
One clarification worth spelling out: that 95% figure is a refusal rate metric, not a measure of accuracy. It tells you how often the model responds rather than refuses. Completions still need to be evaluated for correctness in your specific use case.
The Daybreak Red access requirement is the important part
This is where the migration becomes more involved than a simple model ID swap.
gpt-5.6-cyber is only available through Daybreak Red, the programme tier designed for approved defenders conducting advanced, authorised vulnerability research, exploit validation, and security testing. Access requires a separate application and vetting process. Existing access under earlier schemes, including Trusted Access for Cyber or gpt-5.5-cyber access, does not automatically carry over to Daybreak Red.
Daybreak itself has two tiers. Daybreak Blue gives approved users access to GPT-5.6 Sol with system-level cybersecurity guardrails removed, supporting defensive work. Daybreak Red goes further, providing models that are purpose-trained for cybersecurity tasks and trained to reduce refusals for higher-risk, dual-use work. If your workflows sit at the sharper end of authorised security research, it is Daybreak Red you need to apply for.
The programme is designed for vetted organisations and individuals. OpenAI’s partners in the Daybreak ecosystem include Cisco, CrowdStrike, Cloudflare, and Palo Alto Networks, and access requests may be handled through an organisation-level intake or an individual verification flow depending on your situation.
What you should do before 1 October 2026
Start the Daybreak Red application now. Vetting takes time, and if you wait until September you are accepting real risk of a gap in access. Visit the Daybreak overview in the OpenAI Help Centre to understand the requirements and begin the process.
Audit your AI dependency register. Identify every place in your codebase or infrastructure where gpt-5.4-cyber is referenced as a model ID. This includes environment variables, configuration files, prompt management tools, and any third-party integrations that proxy OpenAI calls.
Test before you migrate. gpt-5.6-cyber is a more capable model, but capability changes can affect output behaviour in ways that matter for your use case. Test quality, cost, and latency before promoting the new model ID to production. Token pricing for gpt-5.6-cyber applies a 2x input and 1.5x output multiplier for prompts exceeding 272K input tokens, and cache writes are billed at 1.25x the standard uncached input rate.
Add graceful error handling. Build your integration so that a model error returns a meaningful signal rather than a silent failure. This is good practice regardless of deprecations, but particularly relevant given the density of changes expected through October 2026.
Explore dedicated capacity if you need continuity. In some cases, OpenAI will provision dedicated access to a model beyond its standard shutdown date. If that is relevant to your situation, contact OpenAI’s sales team before the removal date to discuss options.
The broader picture
gpt-5.6-cyber builds on gpt-5.5-cyber, which OpenAI released in June 2026. The Daybreak programme itself was launched in direct competition with Anthropic’s Project Glasswing, and the original programme ran three model tiers built on GPT-5.5. The consolidation to two tiers (Blue and Red) reflects how OpenAI has refined the access model as the programme has matured.
The core principle behind Daybreak has not changed: OpenAI wants its most capable cybersecurity models in the hands of legitimate defenders, not misused for offensive purposes outside authorised testing. The vetting process exists to make that distinction, and it is why migration here requires more than a config change.
If you are using gpt-5.4-cyber in production, the deadline is firm and the access process takes time. The place to start is the Daybreak application, and the time to start is now.