Between November 30 and December 11, 2026, all three big model providers switch off something a lot of production code still names by string. Anthropic retires Claude Sonnet 4.5. Google shuts down Gemini 2.5 Pro and Gemini 2.5 Flash. OpenAI shuts down GPT-5, GPT-5 mini, GPT-5 nano and o3, and on the same weekend as Anthropic it closes its Prompts endpoint, the Evals API and Agent Builder. The dates are close together, but what happens to your requests is not the same: two providers return errors, one silently serves a different model and bills you at its rate. That difference decides how urgent your migration is.
What retires, and when
These dates come from each provider's own deprecation page, fetched on November 16, 2026.
| Date | Provider | What stops | Replacement named by the provider | Behaviour after the date |
|---|---|---|---|---|
| Nov 30, 2026 | Anthropic | claude-sonnet-4-5-20250929 |
claude-sonnet-5-5 |
Requests fail |
| Nov 30, 2026 | OpenAI | /v1/prompts and reusable prompt objects |
Prompt text in your own code | Deprecated |
| Nov 30, 2026 | OpenAI | Evals dashboard, API and graders | Promptfoo | Read-only |
| Nov 30, 2026 | OpenAI | Agent Builder | Agents SDK or ChatGPT Workspace Agents | Deprecated; ChatKit stays |
| Dec 1, 2026 | gemini-2.5-pro |
gemini-3.8-flash |
Auto-routed, billed at 3.8 Flash rates | |
| Dec 1, 2026 | gemini-2.5-flash |
gemini-3.5-flash-lite |
Auto-routed, billed at 3.5 Flash-Lite rates | |
| Dec 11, 2026 | OpenAI | gpt-5, gpt-5-pro, o3, o3-pro |
gpt-6.1-sol |
Shutdown: no longer accepts requests |
| Dec 11, 2026 | OpenAI | gpt-5-mini |
gpt-6.1-terra |
Shutdown |
| Dec 11, 2026 | OpenAI | gpt-5-nano |
gpt-6-luna |
Shutdown |
| Jan 14, 2027 | OpenAI | GPT-4o, GPT-4o mini, GPT-4.1, GPT-4.1 mini | gpt-6-luna |
Shutdown |
Two things on that list are easy to miss. Dated snapshots go with their parent: gpt-5-2025-08-07 dies with gpt-5, and OpenAI's page says inference on fine-tuned models continues only "until the base models are deprecated", so a fine-tune on GPT-4o stops on January 14 too. And gemini-2.5-flash-lite is not on the list. Google's page says it "is not deprecated and remains available through the API until further notice", so if you only pinned the Lite model you have nothing to do on December 1.
Errors versus silent rerouting
Anthropic's model deprecations page is blunt: once a model is retired, "Requests to retired models fail." OpenAI's deprecations page defines the shutdown state as "The model no longer accepts requests through the standard API." In both cases a forgotten model string becomes an error in your logs on the morning of the cutoff. That's loud, which is good: you'll notice, roll a config change and move on. The risk is an unattended batch job or a customer-facing path nobody watches overnight, and the fix is to find every pinned ID now rather than wait for the alert.
Google chose the opposite. Its Gemini API deprecations page says that "Effective December 1, 2026, all API requests to gemini-2.5-pro will automatically route to" Gemini 3.8 Flash "and be billed at" Gemini 3.8 Flash rates, with the same arrangement sending gemini-2.5-flash to Gemini 3.5 Flash-Lite. Nothing breaks, nothing errors, and that's exactly the problem. A request you tuned for a Pro-class model will be answered by a Flash-class model from a different generation, at a different price, with no change in your code. If you have evals or output parsers calibrated against 2.5 Pro, they'll start drifting on December 1 with no exception to point at. Google also tells you to "Review the breaking API changes in the Gemini 3.8 Flash migration guide before December 1, 2026", which implies the rerouted request may not even be a clean drop-in.
So treat the Gemini change as the most urgent of the three even though it is the only one that won't page you. Pick the replacement model yourself, run your own evals against it before the date, and switch deliberately. Relying on the automatic route means the first time you learn what changed is when a user tells you.
Note also that Google's page describes its shutdown dates as "the earliest possible dates on which a model might be retired". December 1 is a floor, not a promise of extra time. Anthropic's policy is "at least 60 days' notice before model retirement for publicly released models"; OpenAI's is "At least 6 months" for generally available models and as little as two weeks for previews. That spread is why pinning a preview model in production is a bad habit regardless of this quarter.
The OpenAI feature shutdowns are the sneaky ones
Model retirements get noticed because the model name is in the request. The November 30 OpenAI shutdowns are about infrastructure around the request, and they can hide in a dashboard rather than in your repository.
If you adopted reusable prompt objects, every call that references a prompt ID instead of inline text depends on /v1/prompts. OpenAI's replacement advice is simply to "Move prompt content into your application code." That's a mechanical migration, but it also means the versioned prompt history you kept in the dashboard needs to be exported before the endpoint goes. The Evals dashboard, API and graders go read-only the same day with the suggestion to "Migrate to Promptfoo", so any CI step that creates an eval run through the API will start failing. Agent Builder is deprecated in favour of the Agents SDK or ChatGPT Workspace Agents, while ChatKit "remains available". A further date to put in the calendar is January 6, 2027, when self-serve fine-tuning job creation becomes restricted.
The common thread: these were the tools that let you keep configuration outside your codebase. Bringing prompts and eval definitions back into version control is more work up front, but it's also the only state you fully control when a provider changes its mind.
Find every pinned model ID before the dates
A plain grep for each model name works, but it over-matches (gpt-5 hits gpt-5-mini and gpt-5.5-cyber-preview) and misses dated snapshots. The script below encodes the retirement table above, matches a bare ID or its dated snapshot, and refuses to match a longer sibling. It exits non-zero when it finds anything, so it can run in CI.
#!/usr/bin/env python3
"""Scan a source tree for pinned LLM model IDs that retire soon."""
import pathlib
import re
import sys
# model id -> (shutdown date, replacement, what the provider does)
RETIRING = {
"claude-sonnet-4-5-20250929": ("2026-11-30", "claude-sonnet-5-5", "requests fail"),
"gemini-2.5-pro": ("2026-12-01", "gemini-3.8-flash", "auto-routed, billed at 3.8 Flash rates"),
"gemini-2.5-flash": ("2026-12-01", "gemini-3.5-flash-lite", "auto-routed, billed at 3.5 Flash-Lite rates"),
"gpt-5": ("2026-12-11", "gpt-6.1-sol", "shutdown"),
"gpt-5-mini": ("2026-12-11", "gpt-6.1-terra", "shutdown"),
"gpt-5-nano": ("2026-12-11", "gpt-6-luna", "shutdown"),
"o3": ("2026-12-11", "gpt-6.1-sol", "shutdown"),
"gpt-4o": ("2027-01-14", "gpt-6-luna", "shutdown"),
"gpt-4.1": ("2027-01-14", "gpt-6-luna", "shutdown"),
}
EXTS = {".py", ".js", ".ts", ".tsx", ".go", ".java", ".php", ".rb", ".cs",
".json", ".yaml", ".yml", ".toml", ".env", ".ini", ".cfg"}
SKIP_DIRS = {".git", "node_modules", "vendor", "venv", ".venv", "dist", "build"}
# Match the bare id or a dated snapshot (gpt-5-2025-08-07), but not a
# longer sibling such as gpt-5-mini or gemini-2.5-flash-lite.
PATTERNS = {
mid: re.compile(rf"(?<![\w.]){re.escape(mid)}(?:-20\d\d-\d\d-\d\d)?(?![\w.-])")
for mid in RETIRING
}
def scan(root: pathlib.Path):
for path in sorted(root.rglob("*")):
if any(part in SKIP_DIRS for part in path.parts):
continue
if not path.is_file() or path.suffix not in EXTS:
continue
try:
lines = path.read_text(encoding="utf-8", errors="replace").splitlines()
except OSError:
continue
for lineno, line in enumerate(lines, 1):
for mid, pat in PATTERNS.items():
for m in pat.finditer(line):
yield path.relative_to(root), lineno, m.group(0), RETIRING[mid]
def main():
root = pathlib.Path(sys.argv[1] if len(sys.argv) > 1 else ".")
hits = list(scan(root))
for rel, lineno, found, (date, repl, action) in hits:
print(f"{rel}:{lineno}: {found} -> {date}: {action}; replacement {repl}")
print(f"{len(hits)} pinned model id(s) retiring")
sys.exit(1 if hits else 0)
if __name__ == "__main__":
main()
Running it against a small sample tree that pins a mix of current and retiring models (plus gemini-2.5-flash-lite and gpt-5.5-cyber-preview as decoys that must not match) gives this on Python 3.15.0:
$ python3 find_retiring_models.py sample
config/models.yaml:1: gemini-2.5-pro -> 2026-12-01: auto-routed, billed at 3.8 Flash rates; replacement gemini-3.8-flash
config/models.yaml:2: gemini-2.5-flash -> 2026-12-01: auto-routed, billed at 3.5 Flash-Lite rates; replacement gemini-3.5-flash-lite
src/chat.js:7: claude-sonnet-4-5-20250929 -> 2026-11-30: requests fail; replacement claude-sonnet-5-5
src/legacy.py:2: gpt-4o-2024-08-06 -> 2027-01-14: shutdown; replacement gpt-6-luna
src/legacy.py:3: gpt-5-2025-08-07 -> 2026-12-11: shutdown; replacement gpt-6.1-sol
src/summarize.py:6: gpt-5-mini -> 2026-12-11: shutdown; replacement gpt-6.1-terra
src/summarize.py:9: o3 -> 2026-12-11: shutdown; replacement gpt-6.1-sol
7 pinned model id(s) retiring
$ echo $?
1
The two decoys were correctly ignored. The script only covers what is in a repository, so check the places it can't see as well: environment variables set in your deployment platform, secrets managers, feature-flag services, and any prompt or eval configuration that lives only in a provider dashboard.
Migrating without trusting the default replacement
Each provider names a replacement, and the table above repeats them, but a named replacement is a suggestion about capability tier, not a guarantee that your prompts behave the same. The practical order is: switch the model string in a branch, run whatever evaluation set you already have against the new model, compare cost per request from real token counts, and only then ship. Where a provider offers a migration guide for the replacement, as Google does for Gemini 3.8 Flash, read it for request-shape changes before the eval run, or you'll be debugging schema errors instead of measuring quality.
Pin the dated snapshot of the replacement where the provider offers one, and write the retirement policy into your runbook: Anthropic gives at least 60 days, OpenAI at least six months for GA models, Google publishes an earliest-possible date and promises advance notice of the exact one. If a model ID in production has no owner who reads those pages, the next cliff will arrive the same way this one did.
No comment for this article.