If your production configs or CI scripts are still pointed at older OpenAI model snapshots like gpt-5.2-codex, they are scheduled to fail starting July 23, 2026. This isn't breaking news—the schedule was announced back on April 22—but the deadline is now close enough to force a review of your dependencies. This is more than a version bump; the replacement models have different performance and pricing profiles, requiring active migration, not just a find-and-replace.
what's actually happening
On July 23, OpenAI will shut down access to a batch of 13 older model snapshots. While this wave includes various preview and specialized research models, the most significant impact for many builders is the retirement of five specific Codex-lineage models. These are the workhorses that powered many early AI coding assistants and agents. The list includes every Codex variant before GPT-5.3.
This is part of a standard platform hygiene process to simplify the available models and focus resources on newer, more capable ones. According to OpenAI's own policy, specialized variants like the Codex series are given at least three months' notice before shutdown, and this wave complies with that.
why this requires more than find-and-replace
The critical takeaway is that the migration path isn't straightforward. You can't just swap the old model ID for the new one and expect identical behavior or cost. Most of the retiring Codex and chat snapshots are being mapped to gpt-5.5. However, other models, like the deep-research variants, map to gpt-5.5-pro, which comes with a different price tier.
Furthermore, the replacement models may have different performance characteristics or architectural details. For instance, gpt-5.5 introduces a long-context surcharge, where prompts over a certain token count are billed at a higher rate. A workload that previously relied on stuffing large codebases into a Codex model's context could see a significant change in its monthly bill. The only way to know the impact is to test.
Here is a simple example of what an API call update might look like. The change itself is trivial, but it implies a non-trivial validation workload.
import openai
# Deprecated call, will fail after July 23, 2026
# response_old = openai.chat.completions.create(
# model="gpt-5.2-codex",
# messages=[
# {"role": "system", "content": "You are a helpful coding assistant."},
# {"role": "user", "content": "Write a python function to calculate a factorial."}
# ]
# )
# Updated call to a supported model
# Note: Performance, latency, and cost may differ. Must be re-evaluated.
response_new = openai.chat.completions.create(
model="gpt-5.5", # Or another suitable replacement
messages=[
{"role": "system", "content": "You are a helpful coding assistant."},
{"role": "user", "content": "Write a python function to calculate a factorial."}
]
)
print(response_new.choices[0].message.content)
your audit checklist before july 23
This isn't a fire drill, but it is an action item. Before the deadline, you should run a thorough audit.
- Grep your codebase: Search your entire application, including notebooks and utility scripts, for the retiring model IDs. The full list is on OpenAI's deprecations page.
- Check your infrastructure: Look for pinned model versions in your CI/CD pipelines, Terraform scripts, or any other infrastructure-as-code definitions.
- Review agent backends: If you run autonomous agents, double-check the model configurations they use for different tasks.
- Re-run your evaluations: Once you've pointed a service at a new model, run your existing test suites and evals to check for regressions in quality, format, or latency.
- Update your cost models: Analyze the pricing for the replacement models and adjust your budget forecasts accordingly, keeping an eye on factors like long-context surcharges.
This is a forcing function to treat model IDs as the managed dependencies they are. The process is routine, but ignoring it will lead to broken systems. More waves of deprecations are already scheduled for October and December, so building this audit into your regular process is the only sustainable path.
Top comments (0)