Snowflake Cortex model end of life: what fails from 14 October and what to switch to

The Snowflake Cortex model deprecation covers 16 models that went legacy on August 12, 2026. Two of them, claude-4-sonnet and openai-gpt-4.1, reach end-of-life on October 14, 2026. The other 14 are listed as "No sooner than October 14, 2026." After end-of-life, any call that names the model fails. Here's what to check this week.

What is retiring and when

Snowflake's August model deprecations release note lists 16 rows, and every one has the same legacy date. Fine-tuned variants are separate rows with their own entries, so a llama3-70b call and a fine-tuned llama3-70b call are two different problems.

"No sooner than" means the date isn't fixed yet, and we haven't seen Snowflake say when it will be. As a judgment call, plan for October 14 on all 16. Moving early costs you a test cycle. Moving late costs you a failed job.

Who is affected: legacy versus end-of-life

The release note opens with a distinction that's easy to skim past: "Legacy is not the same as end-of-life." What your account can do today depends on whether it used the model before August 12.

If it did, Snowflake says the account "can continue to use it in every supported application" (including AI_COMPLETE, CORTEX.COMPLETE, the Agents API, Cortex Inference and Snowflake Intelligence) until the model's end-of-life date. If it hadn't, the account "can't start using it. Queries and API calls that name the model fail." So a team that never touched llama3-8b before August is already blocked, and a team that did has until the EOL date.

That grace ends for everyone together. The lifecycle section of Snowflake's regional availability page says that after end-of-life, "the model is no longer available to any account, regardless of prior usage." It also recommends migrating before that date "even if your account can still call the legacy model."

One exemption. Managed functions such as AI_CLASSIFY, AI_FILTER and AI_AGG aren't affected, because Snowflake manages inference of the models underneath them. If you only use those, nothing here touches you.

What fails after the date

Anything with a named model parameter. The release note (Ref: 2388) says that if the parameter isn't updated, "queries or API calls that reference the model fail after its end-of-life date for every account, including accounts that were allowed to use the model during the legacy period."

The risky callers are the ones nobody opens often. A task that runs weekly, a stored procedure behind a dashboard, a Streamlit app or agent configuration with a model string pasted in two years ago. Nothing fails at deploy time. The failure shows up when the job next runs after the date, and possibly with no owner who remembers writing it.

What to switch to

Snowflake doesn't name replacements. The release note tells you to "transition to a replacement model" and stops there, so there is no official claude-4-sonnet-to-X mapping to follow. Everything in this section is our judgment, not Snowflake's guidance.

The candidates come from the models Snowflake's "Choosing a model" section lists as current: claude-opus-5-5 (Public Preview, requires cross-region inference), claude-opus-5, claude-opus-4-8, claude-sonnet-5, claude-sonnet-4-6, claude-haiku-4-5, openai-gpt-5-mini (requires cross-region inference for Azure US accounts), llama3.3-70b (Medium), llama3.1-8b (Small), and gemini-3.1-pro (Public Preview, needs cross-region inference). Snowflake's page also still lists mistral-7b and llama3-70b even though the release note retires them.

Our read, to be tested and not trusted:

• For claude-4-sonnet, the natural place to start is claude-sonnet-4-6 or claude-sonnet-5, since you keep the same model family. Retiring a model on a platform you don't control follows the same pattern as Anthropic's own Sonnet 4.5 retirement, where a model ID swap was the easy part.
• For openai-gpt-4.1, openai-gpt-5-mini is the only OpenAI model on the current list, but it carries a cross-region requirement for some accounts (Azure US), so check that first. OpenAI's own removal dates for gpt-5.1 and TTS show how quickly that vendor moves its lineup.
• For the llama rows, llama3.3-70b (Medium) and llama3.1-8b (Small) are the open-weights Llama models on the current list. Mistral-7b also appears there but is itself retiring, so teams on mistral, mixtral or pixtral rows have no same-family replacement and should expect a family change and a bigger re-test.
• For fine-tuned rows, Snowflake gives no replacement at all. Treat each as its own project, because a fine-tune doesn't carry over to a different base model.

Cross-region availability varies by account region. Check the regional tables before you commit to anything, and where the prose and the release note disagree, trust SHOW CORTEX BASE MODELS over prose.

How to migrate off a retiring Cortex model

1. List the retiring models your account can see

Run this in your account:

SHOW CORTEX BASE MODELS IN SCHEMA SNOWFLAKE.MODELS;

Snowflake's SHOW CORTEX BASE MODELS reference lists each model's lifecycle status, availability, legacy date and end-of-life date, for the models available to your current role. Read the lifecycle_status, legacy_date and eol_date columns, and note that the reference recommends specifying the IN SCHEMA SNOWFLAKE.MODELS clause. You can narrow the output with LIKE or STARTS WITH.

2. Find every call that names one

Search for all 16 names across stored procedures, tasks, SQL and dbt repositories, Streamlit apps and agent configurations. Include fine-tuned model names, which won't match a plain search for the base name. Give each hit an owner, because someone has to say yes to a changed model.

3. Pick a replacement per call

Use the candidates above as a starting list, then narrow by what each call does. Check cross-region requirements for your account first, since that rules models out faster than quality does.

4. Re-test outputs on a fixed sample

Take a fixed set of real inputs per call and compare old and new output side by side. Don't judge on a handful of prompts. Fine-tuned variants need their own plan here, since there's no stated replacement for them.

5. Re-estimate credit spend

A swap can change what a call costs. Snowflake's Cortex cost documentation is the place to compare, and we haven't quoted per-model figures because we haven't verified them. Compare performance per credit, not list price alone.

6. Change the parameter, ship, and verify

Update the named model in each call, deploy, then re-run SHOW CORTEX BASE MODELS and confirm nothing you still reference is on the list. For the 14 "no sooner than" models, set a calendar check for when Snowflake fixes their dates.

What this means for your Snowflake AI roadmap

By October 14, every call that names one of the 16 models needs an owner, a chosen replacement, a quality re-test and a re-estimated credit cost. That's a small list for most accounts, but it's easy to miss a task that only runs monthly. If you can't produce the list from step 2 in a day, you have a visibility problem that matters more than this one deadline.

Two judgment calls. Treat every "no sooner than" row as October 14. And pin to current models, not ones already on a legacy path, then schedule a lifecycle review whenever Snowflake publishes a deprecation note. It has published deprecation notes in April, May, June and August 2026 (the June page lists July 10 dates), so a quarterly check is the floor.

For how these forced moves land on engineering budgets across vendors, see AI model releases and pricing: what changes for engineering teams.

FAQ

Does my account still work with a legacy model until October 14?

Only if it used that model before August 12, 2026. In that case it can keep using it in every supported application until the end-of-life date. After that date the model is unavailable to every account.

Do AI_CLASSIFY, AI_FILTER and AI_AGG break?

No. Snowflake says managed functions aren't impacted, because it manages and optimizes inference of the underlying models.

What happens if I never used the model before August 12?

Your account can't start using it, and queries and API calls that name the model fail now, not on the end-of-life date.

Which models have a fixed end-of-life date?

Two: claude-4-sonnet and openai-gpt-4.1, both on October 14, 2026. The other 14 are "No sooner than October 14, 2026."

How do I see end-of-life dates in my account?

Run SHOW CORTEX BASE MODELS IN SCHEMA SNOWFLAKE.MODELS and read the lifecycle_status, legacy_date and eol_date columns.

Does Snowflake name a replacement model?

No. The release note says to transition to a replacement model but doesn't say which. Any mapping is a judgment you have to test.