News · 2026-09-29

OpenAI retires four legacy GPT-3 and GPT-3.5-era API model IDs

OpenAI shut down four legacy GPT-3/GPT-3.5-era API model IDs on September 28 and recommends gpt-5.6-terra as the replacement. The news is operational rather than glamorous: any application still calling these IDs needs to migrate and re-test behavior, especially if it relies on the older completions interface.

Key facts

• The retired IDs are gpt-3.5-turbo-instruct, babbage-002, davinci-002, and gpt-3.5-turbo-1106.
• The shutdown date listed by OpenAI is September 28, 2026.
• OpenAI recommends gpt-5.6-terra for each retired ID.
• The primary record is OpenAI’s deprecations documentation.

This corrects a misleading shorthand that “GPT-3 discontinued today.” babbage-002 and davinci-002 are legacy completion-era models; gpt-3.5-turbo-instruct belongs to the older instruct/completions path; and gpt-3.5-turbo-1106 is a chat-model snapshot. OpenAI had already retired the original ada, babbage, curie, and davinci base-family models in January 2024, as its historical notice records.

Think of this as retiring four old electrical connectors, not turning off electricity. A service may still work after swapping the connector, but its assumptions about message format, instruction following, tool use, output style, latency, token counting, and safety behavior can all change. That is why the recommended replacement is a starting point for a migration, not proof of drop-in equivalence.

OpenAI’s current model catalog and changelog describe a very different platform from the one these models entered: reasoning controls, Responses API workflows, asynchronous tools, mid-turn steering, and managed agent infrastructure. Legacy code can fail not only because an endpoint disappears, but because a newer model produces a different decision boundary on the same prompt. Test fixtures should include bad inputs, structured-output parsing, refusal paths, and the actual business action downstream of each completion.

There is no major controversy in the fact of the shutdown; the controversy is in the headline. Calling it the demise of all GPT-3 models erases a two-year history and can make teams think a migration is unexpected. The accurate claim is smaller and more useful: OpenAI retired four named legacy IDs, mainly affecting older API surface area.

A practical migration sequence is inventory first, pin test cases, compare outputs and failure modes, then roll out behind a fallback and observable metrics. The goal is not to recreate an old model forever. It is to identify whether the product depended on accidental old behavior. This is particularly important when a completion triggers an automated action, because a higher-capability replacement can be more assertive as well as more capable.

The caveat is that the deprecations page supplies the status and recommended replacement, not a guarantee that every workload will improve. Developers should treat “recommended” as OpenAI’s supported direction and verify their own cost, quality, latency, and safety constraints before removing a fallback.

Primary source, verified: read the paper →

Ground Truth. (2026, September 29). OpenAI retires four legacy GPT-3 and GPT-3.5-era API model IDs. Ground Truth. https://groundtruth.day/news/openai-retires-four-legacy-gpt-era-api-model-ids.html

@misc{groundtruth:openai-retires-four-legacy-gpt-era-api-model-ids,
  title  = {OpenAI retires four legacy GPT-3 and GPT-3.5-era API model IDs},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/openai-retires-four-legacy-gpt-era-api-model-ids.html}
}

Topics: openai · api · deprecation · developers