Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

GPT-5.5-Cyber preview ends February 1, 2027

Written November 9, 2026

OpenAI's deprecations page (Nov 5) removes gpt-5.5-cyber-preview on Feb 1, 2027 — and names no replacement. What to move, and the unnamed successor.

What ends

The OpenAI API deprecations page carries a November 5 entry for GPT-5.5-Cyber: the gpt-5.5-cyber-preview model is deprecated and will be removed from the API on February 1, 2027. After that date the model is no longer accessible — the page uses "sunset" and "shut down" interchangeably to mean exactly that. Fine-tuned versions of the model follow the same shutdown schedule as the underlying base model, so a fine-tune does not buy extra time.

The entry names no replacement model. Its recommended-replacement column reads, in full, "The most capable cyber model available to you." There is no model name to switch to, no migration guide linked from the entry, and no date by which a successor will be named. If your code pins the preview model by name, the page is telling you to choose the replacement yourself.

Why the date is not short notice

Preview models are the one class OpenAI says may be retired with much shorter notice — as little as two weeks, according to the same page — because they are not recommended for business-critical production workloads unless you can migrate on short notice. This one gets about three months from announcement to shutdown, which is longer than the minimum, but it is still a preview retirement: the runway is a courtesy, not a guarantee of a successor.

The cyber line has been retiring stepwise. gpt-5.4-cyber was deprecated on September 11 and removed from the API on October 1, 2026, with the same unnamed replacement wording. GPT-5.5-Cyber is the next step in the same sequence, not the first.

How to check what you are running

The deprecations page notes that you can monitor model shutdowns programmatically: call the List models endpoint (GET /v1/models) and check each model's shutdown_date field, which holds the shutdown date as a YYYY-MM-DD string when one has been announced, or null if none has. That field is the machine-readable version of the same table — use it to find every integration that still names the preview model, rather than trusting a search of your own code alone.

Verdict

Move now and test the replacement yourself — OpenAI names no successor, so there is nothing to wait for. Pick the cyber model your account can already use, run your security workload against it, and compare the results before February 1, while the preview still answers.

Steps

  1. Find every call to the preview model. Search code, configuration, scheduled jobs, and scripts for gpt-5.5-cyber-preview, and check GET /v1/models for the shutdown_date field.
  2. List the fine-tunes too. Fine-tuned versions follow the same February 1 schedule as the base model. A fine-tune needs a new base, not just a new name.
  3. Choose the replacement from what your account can use. There is no named successor on the page. Test the most capable cyber model available to you against your actual workload — do not assume the newest model is the right one.
  4. Switch before February 1, not on it. The model stops being accessible on the shutdown date. A switch made that week has no room to find a regression.

What can go wrong

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: OpenAI's training ban lands November 20 · Agent change deadlines: retirements and shutdowns · GPT-6 Cyber: OpenAI's restricted security model