The Last of GPT-3: OpenAI's Four Legacy Models Go Dark the Day Before DevDay

On Monday, September 28, 2026, four model identifiers that had been answering API requests for years quietly stopped responding. gpt-3.5-turbo-instruct, gpt-3.5-turbo-1106, babbage-002, and davinci-002 hit their formal shutdown date on OpenAI’s deprecation calendar, and with them the last completion-style endpoints of the GPT-3 lineage went dark. Any application still calling those IDs now receives an error instead of a completion.

The timing is the story. Tomorrow — September 29 — OpenAI hosts its annual DevDay conference at Fort Mason in San Francisco. The company that spent 2026 shipping GPT-6 Astra, an agent runtime in beta, a per-minute-priced voice model, and a three-tier GPT-5.6 family has deliberately cleared the oldest code off the runway before landing its next generation of developer tooling. The old API surfaces die the day before the new ones take the stage.

What exactly shut down

OpenAI’s deprecation table lists the removal with surgical precision. Four exact model IDs left the API on September 28:

• gpt-3.5-turbo-instruct — the last widely-used instruct-tuned completion model, a direct descendant of the model that powered ChatGPT at launch in November 2022
• gpt-3.5-turbo-1106 — a frozen November 2023 snapshot of GPT-3.5 Turbo, kept alive for reproducibility long past its contemporaries
• babbage-002 and davinci-002 — the replacement generations of the original GPT-3 base models, named after Charles Babbage and Ada Lovelace’s contemporaries, serving raw text completion since the API’s earliest days

All four are mapped to a single successor in OpenAI’s migration table: gpt-5.6-terra, the mid-tier of the GPT-5.6 family launched July 9, 2026. The routing is deliberate — legacy base models and instruct variants land on Terra, not the flagship Sol, because Terra occupies the price-performance slot those workloads actually need.

This follows the September 24 shutdown of the entire Sora 2 video stack — the Videos API and both Sora models — which means OpenAI has now retired both its oldest text models and its first-generation video models within a single week.

Six years, in four identifiers

The GPT-3 era began in June 2020, when OpenAI opened private beta access to an API serving a 175-billion-parameter model that astonished developers by writing prose, code, and translation from raw text completion. Babbage and Davinci were the names on that first pricing page. When GPT-3.5 arrived in 2022, the -turbo variants made the API cheap enough to power consumer products, and ChatGPT’s launch that November turned the model family into a household name.

The instruct suffix marks a philosophical waypoint: the period when OpenAI taught completion models to follow instructions rather than merely continue text. The -1106 snapshot represents an earlier era’s approach to versioning — freeze a model, promise it forever, and let enterprise contracts depend on exact reproducibility. Both conventions are now historical artifacts. Modern API design ships snapshot-dated models with explicit deprecation windows, and the industry has largely converged on chat-structured and agentic interfaces rather than raw completion.

When these four identifiers stop answering, no living OpenAI model will accept a bare completion-style prompt through the legacy endpoint. The API surface that built the company’s developer ecosystem will exist only in documentation, tutorials, and an enormous body of blog posts whose curl examples no longer work.

The DevDay timing is not a coincidence

OpenAI’s deprecation calendar has become a messaging instrument. The August 2026 purge cleared the GPT-4o/GPT-4.1 middle generation. The September 28 date closes the GPT-3 lineage. Tomorrow’s DevDay — the first since GPT-6 Astra’s September 3 release — is expected to push hard on the agentic future: the agent runtime that entered beta this month, the rumored always-on “o” agent, and deeper developer tooling around GPT-6’s computer-use capabilities.

Clearing legacy models the day before a developer conference serves three purposes. It forces every laggard integration to modernize before the event, so the audience arriving at Fort Mason is building on current-generation APIs. It simplifies OpenAI’s own serving infrastructure — maintaining 2023-era completion paths has real cost. And it makes the narrative clean: the platform that tomorrow’s demos run on contains nothing older than the GPT-5 generation.

What developers should actually do

The practical guidance is straightforward. Audit dependency lists and infrastructure code for the four shutdown IDs — automated dependency scanners routinely miss hardcoded model strings in environment variables and legacy service configurations. Applications calling gpt-3.5-turbo-instruct for completion-style workloads should migrate to gpt-5.6-terra using the Responses API, which handles structured tasks natively. Note that completion-style prompting does not transfer directly: a raw continuation prompt will underperform on a chat-trained model, so the migration is behavioral, not just an identifier swap.

Fine-tuned derivatives follow on October 23, 2026: ft-babbage-002, ft-davinci-002, ft-gpt-3.5-turbo, ft-gpt-4, and ft-gpt-4.1-nano-2025-04-14 all shut down that day, redirecting to Terra and the budget-tier Luna. Organizations with fine-tuned GPT-3.5 checkpoints still in production have three weeks to retrain on current base models or accept the breakage.

Pricing context matters for the migration math. Terra entered the market at $2.50 per million input tokens and $15 per million output tokens; OpenAI cut Terra’s price by 20% on July 30, bringing it to $2.00/$12.00. That is dramatically cheaper per token than GPT-3.5-turbo’s original 2023 pricing while being a categorically more capable model — one of the rare migrations where the modern replacement costs less in both absolute and capability-adjusted terms.

Why this matters beyond OpenAI

The shutdown is a small event with a large symbolic footprint. GPT-3 was the first model to demonstrate that a general-purpose language model could be a product. Every AI application in production today descends from the API conventions those four identifiers established: the tokens-per-pricing-unit model, the model-as-a-service abstraction, the completion request itself.

Their removal marks the completion of a full platform generation cycle — roughly six years from experimental beta to deprecated relic. That cycle time is compressing. GPT-4o, once the most-deployed model in the industry with a reported 37.6% share of cloud deployments, was retired in June 2026 barely two years after launch. Model lifespans on the frontier are now measured in months, and the deprecation calendar has become as strategically important to track as the release calendar.

For the developers heading to Fort Mason tomorrow, the message is implicit in the calendar: build for the model generation that exists today, because the identifiers you’re calling now are already on someone’s shutdown list.