Your pinned GPT-5 snapshot expires December 11

Pinning a model snapshot is the correct engineering instinct. You freeze gpt-5-2025-08-07 so your prompt behaves the same in December as it did in August, and you stop chasing silent quality drift. What most teams miss is that the pin has an expiry date. OpenAI removes six pinned snapshots on December 11, 2026, and that clock started on June 11.

What actually happened

Per OpenAI's deprecations page, announced June 11, 2026, these snapshots go away on December 11, 2026:

Removed	Replacement
gpt-5-2025-08-07	gpt-5.6-sol
gpt-5-mini-2025-08-07	gpt-5.6-terra
gpt-5-nano-2025-08-07	gpt-5.6-luna
gpt-5-pro-2025-10-06	gpt-5.6-sol with reasoning.mode: pro
o3-2025-04-16	gpt-5.6-sol
o3-pro-2025-06-10	gpt-5.6-sol with reasoning.mode: pro

Six months of notice. That is a reasonable window and considerably more than some vendors give — but it is a window, not an exemption.

Also on the page and worth a calendar entry: January 20, 2027 ends the legacy audio, realtime, and transcription snapshots (gpt-realtime, gpt-audio, gpt-4o-audio, gpt-4o-realtime and friends), with gpt-realtime-2.1 and gpt-audio-1.5 as replacements. If you built a voice intake line or a call-transcription pipeline in the last eighteen months, that one is aimed at you.

Why pinned model deprecation matters for your business

The failure mode here is specific and we have watched it happen: a team pins a snapshot, ships, validates once, and then treats the model as a settled dependency — because that is exactly what pinning was supposed to buy. Six months later the pin evaporates, traffic silently lands on a newer model with different behavior, and the first signal is a customer complaint about an output that used to be right.

gpt-5.6-sol is not a drop-in for gpt-5-2025-08-07. It is a better model. That is precisely the problem: better models break prompts that were tuned against the old one's quirks. Formatting shifts. Refusal boundaries move. Reasoning-mode parameters change what you get back. A prompt with a brittle "respond with only the JSON, no preamble" instruction can pass every eval you wrote and still fail in a way you did not think to test.

Three things to do before December.

Own an inventory of pinned model IDs. Grep your codebase, your env vars, and your third-party tool configs for model strings. Most teams find them in more places than they expected — a background job, a Zapier step, an internal script somebody wrote once.

Build a regression set from real traffic, not from your imagination. Fifty to a hundred actual inputs with known-good outputs. This is a two-hour job that pays for itself the first time a vendor moves your floor, and it is reusable for every future migration.

Run the swap early and measure, not on December 10. Send the regression set through the replacement, diff the outputs, and look at the disagreements by hand. Then flip.

The larger point: any model ID in your code is a dated dependency, the same as a pinned package version with an end-of-life. Treat vendor deprecation pages the way you treat CVE feeds — something a person reads on a schedule, not something you discover from an error rate.

Sources: OpenAI API Deprecations.