OpenAI Legacy Models Shut Down September 28: Migration Checklist

Direct answer: OpenAI is scheduled to shut down API access to gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 on September 28, 2026. OpenAI names gpt-5.6-terra as the replacement for all four.[1] Search code, environment variables, deployment secrets and automation settings for the exact retiring IDs, move traffic behind a configurable model name, test output and error handling, then deploy before the cutoff.

OpenAI models affected

Retiring model ID	Shutdown date	OpenAI replacement
gpt-3.5-turbo-instruct	September 28, 2026	gpt-5.6-terra
babbage-002	September 28, 2026	gpt-5.6-terra
davinci-002	September 28, 2026	gpt-5.6-terra
gpt-3.5-turbo-1106	September 28, 2026	gpt-5.6-terra

The deadline concerns these four exact API model IDs, not a general shutdown of every model whose name contains GPT-3.5. Do not replace unrelated model names without checking OpenAI’s current schedule.[1]

Five-minute audit

Run this from each application repository:

git grep -nE 'gpt-3\.5-turbo-instruct|gpt-3\.5-turbo-1106|babbage-002|davinci-002'

Then check the deployment layer:

• environment variables such as OPENAI_MODEL;
• CI/CD variables and secret managers;
• serverless and container configuration;
• scheduled jobs, workflow builders and no-code automations;
• test fixtures, fallback chains and internal setup documents;
• model IDs stored in a database or admin panel.

Security note: do not print secret values while auditing. Search variable names and model values only.

Safe migration order

1. Inventory every use of the four exact IDs.
2. Identify the API shape used by each integration; old Completions, Chat Completions and Responses requests are not interchangeable.
3. Put the replacement in configuration rather than scattering a literal model name through the codebase.
4. Create representative tests for real prompts and expected outputs.
5. Add useful monitoring for output validation, latency, token use and errors without recording sensitive prompt data.
6. Test gpt-5.6-terra in staging.
7. Run a controlled canary in production if time and architecture permit.
8. Deploy and watch failures, latency, token use and output validation.
9. Remove the retired fallback so a failed request cannot silently retry a dead model.

OpenAI describes GPT-5.6 Terra as a model that balances intelligence and cost.[2] A replacement is not guaranteed to reproduce old wording, token counts or formatting, so validation matters.

Minimal Python migration example

from openai import OpenAI

client = OpenAI()

response = client.responses.create(
    model="gpt-5.6-terra",
    input="Summarize this support ticket in three bullet points.",
)

print(response.output_text)

OpenAI recommends the Responses API for new projects and documents response.output_text as an SDK convenience property for aggregating text output.[3][4]

Minimal JavaScript migration example

import OpenAI from "openai";

const client = new OpenAI();

const response = await client.responses.create({
  model: "gpt-5.6-terra",
  input: "Summarize this support ticket in three bullet points.",
});

console.log(response.output_text);

Why a model-name swap may not be enough

gpt-3.5-turbo-instruct, babbage-002 and davinci-002 commonly appear in older completion-style integrations, while gpt-3.5-turbo-1106 commonly appears in chat-style code. OpenAI’s migration guide treats moving to Responses as changes to the request endpoint, output shape and conversation-state handling; function calls and Structured Outputs also use different request shapes.[3]

Check these before deploying:

• input field and message format;
• output parsing;
• system or developer instruction placement;
• function and tool calls;
• structured JSON validation;
• streaming event handling;
• multi-turn state;
• timeouts, retries and rate limits;
• token budgets and cost alerts.

Copy-paste deployment checklist

• ☐ Searched all repositories for the four exact IDs
• ☐ Checked environment variables, deployment settings and secret managers
• ☐ Checked cron jobs, workflow tools and no-code automations
• ☐ Replaced the model through configuration
• ☐ Updated request and response handling where required
• ☐ Tested normal, empty, long and malformed inputs
• ☐ Tested structured output or function calls if used
• ☐ Compared representative results with current production behavior
• ☐ Added monitoring and rollback configuration
• ☐ Removed retired-model fallbacks
• ☐ Deployed before September 28

Frequently asked questions

Which OpenAI models shut down on September 28, 2026?

gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106.[1]

What replacement does OpenAI recommend?

OpenAI lists gpt-5.6-terra for all four retiring model IDs.[1]

Does every GPT-3.5 model shut down on September 28?

No. The cited notice lists four exact model IDs. Audit exact strings and consult OpenAI’s live deprecations page for other models and dates.[1]

Can I change only the model name?

Sometimes a basic integration may need little more, but older completion-style code can require request and output changes. OpenAI recommends Responses for new projects and provides a migration guide covering endpoints, outputs, tools, structured data and state.[3][4]

Should I wait until requests fail?

No. Migrate and test before the cutoff so you can compare behavior and roll back configuration without a production outage.

Last checked: September 27, 2026. Deprecation schedules can change; confirm the live OpenAI deprecations page before deployment.

Official sources

1. OpenAI API deprecations
2. GPT-5.6 Terra model documentation
3. Migrate to the Responses API
4. OpenAI text generation guide