OpenAI Kills Four Legacy Models Today — Hours Before DevDay 2026

Four OpenAI API models stop answering calls on September 28, 2026. The shutdown lands one day before OpenAI DevDay in San Francisco — and four days after the Videos API and every Sora 2 alias already went dark.

If your stack still pins gpt-3.5-turbo-instruct, gpt-3.5-turbo-1106, babbage-002, or davinci-002, today is not a soft warning. Per OpenAI’s deprecations page, those four IDs are removed from the API. The named replacement for all of them is gpt-5.6-terra.

The quieter risk is confusion. The base gpt-3.5-turbo alias is not on today’s list. It stays until the larger October 23 wave. Mixing those dates is how production tickets get filed at 2 a.m.

What Dies on September 28

OpenAI announced this set on September 26, 2025, under “Legacy GPT model snapshots.” The shutdown date has been public for a year. Here is the table straight from the docs:

Model ID	Shutdown	Official replacement
gpt-3.5-turbo-instruct	Sept 28, 2026	gpt-5.6-terra
gpt-3.5-turbo-1106	Sept 28, 2026	gpt-5.6-terra
babbage-002	Sept 28, 2026	gpt-5.6-terra
davinci-002	Sept 28, 2026	gpt-5.6-terra

Two of these matter more than the others in real apps. gpt-3.5-turbo-1106 is a Chat Completions snapshot teams pinned to avoid alias drift. gpt-3.5-turbo-instruct is the Completions-endpoint holdout — the one that still takes a plain prompt string instead of a messages array. Moving off instruct is not a model-ID swap. You change the request shape.

What Already Died: Sora 2 and the Videos API

On September 24, OpenAI removed the Videos API and every listed Sora 2 model — sora-2, sora-2-pro, and their dated snapshots. The deprecations page lists no replacement. That is a harder break than today’s GPT-3.5 retirements: there is no OpenAI video-generation model to point at.

If you still have a feature flag that calls the Videos API, it is already returning errors. Treat that as a product decision, not a migration ticket.

What Comes Next on the Deprecation Calendar

Today is not the only cutoff this week:

Date	What shuts down	Replacement
Sept 24, 2026	Videos API; all sora-2 aliases	None listed
Sept 28, 2026	Four legacy GPT IDs above	gpt-5.6-terra
Oct 1, 2026	gpt-5.4-cyber	gpt-5.6-cyber
Oct 23, 2026	gpt-3.5-turbo-0125, gpt-4-turbo, several o-series snapshots, and more	Varies — OpenAI lists gpt-5.6-terra or gpt-5.6-sol depending on the model

The October 1 Cyber swap is gated. Per OpenAI’s GPT-5.6 Cyber model page, the model needs Daybreak program approval. List price there is $12.50 input / $75 output per million tokens — a defender-focused tier, not a default chat model.

The October 23 wave is the one that finally retires the main gpt-3.5-turbo-0125 snapshot and a long list of GPT-4 / o-series pins. If you are cleaning house today, audit those strings in the same pass.

Official Replacement vs. What You Should Actually Call

OpenAI’s docs name gpt-5.6-terra as the substitute for today’s four models. From the Terra model page, that means:

• $2.00 per million input tokens
• $0.20 cached input
• $12.00 per million output tokens
• 1,050,000 context; 128,000 max output
• Knowledge cutoff: February 16, 2026

That is the compliant answer for a migration ticket. It is not automatically the cheapest answer.

OpenAI’s current API pricing page lists GPT-6 Luna at $0.10 / $0.50 and GPT-6 Sol at $2 / $10 for short context — the rates we covered when Sol and Luna cut prices in half. For high-volume classification and summarization that used to sit on GPT-3.5, Luna is usually the better default than Terra. For coding agents, Sol is the closer peer to what teams expected from mid-tier GPT-5.6.

One practical rule: use Terra when you need the officially named substitute with minimal retesting. Use Luna or Sol when you are willing to re-eval prompts against the GPT-6 family and cut the bill.

The Instruct Migration Trap

gpt-3.5-turbo-instruct lived on the legacy Completions endpoint. Current GPT-5.6 and GPT-6 models expect Chat Completions or the Responses API with a messages (or input) structure. A string-replace on the model field will fail even if Terra is available.

Checklist that actually works:

• Search the repo for the four model IDs and for /v1/completions
• Convert prompt templates into system/user messages
• Re-check max tokens, stop sequences, and logprobs assumptions
• Run a golden-set eval before flipping production

Teams that pinned -1106 “just in case” have less runway than they think. That snapshot shares today’s date, not the October 23 date that applies to gpt-3.5-turbo-0125.

DevDay Is Tomorrow — What Is Confirmed

OpenAI’s event site confirms the logistics, not the product roadmap:

• When: Tuesday, September 29, 2026
• Where: Fort Mason, San Francisco
• Keynote: 10:00 a.m. Pacific, featuring Sam Altman
• Livestream: opening keynote free; other sessions recorded and posted later
• In-person: applications closed; invited registration was $650
• Exchanges: later events listed for Bengaluru, Tokyo, Seoul, Paris, Berlin, London, São Paulo, and Mexico City

That is all OpenAI has put on the record for the day. Secondary outlets have floated a cybersecurity model preview — often called “GPT-6 Cyber” — around DevDay. OpenAI has not confirmed that name, a model card, or a preview slot. Treat it as unconfirmed until it appears on the changelog or the keynote stream.

The timing still matters for builders. OpenAI has spent September shipping Sol and Luna, pausing tool-use on its most capable research models after a DNS sandbox breakout, and clearing legacy models off the API. DevDay is the moment those threads either get explained — or get another round of ship emojis without a migration guide.

What It Means for Indian Developers

Legacy GPT-3.5 pins are common in Indian SaaS and agency codebases: cheap defaults from 2023–2024 that never got a second look. Today’s shutdown is the forced cleanup. Terra at $2 / $12 is fine for a stopgap. Luna at $0.10 / $0.50 is the price that matches how most Indian products actually use LLMs — support bots, document extraction, lead scoring, and light coding helpers.

Pair the API swap with the India-priced tooling stack we have already mapped: Cursor Start at ₹649/month and the AI IDE range from ₹96 to ₹958. Token price and IDE price are the same budget line.

Bengaluru is also on OpenAI’s DevDay Exchange list later this year. The San Francisco keynote stream at 10:00 a.m. PT is 10:30 p.m. IST on September 29 — watchable without a $650 ticket.

Confirmed vs. Unconfirmed

Claim	Status
Four model IDs shut down Sept 28, 2026	Confirmed on OpenAI’s deprecations page
Official replacement is gpt-5.6-terra	Confirmed by OpenAI
Videos API / Sora 2 removed Sept 24 with no replacement	Confirmed by OpenAI
DevDay Sept 29 at Fort Mason; Altman keynote 10 a.m. PT livestream	Confirmed on OpenAI’s DevDay site
Terra $2 / $12; Luna $0.10 / $0.50; Sol $2 / $10	Confirmed on OpenAI model/pricing pages
GPT-6 Cyber preview at DevDay	Unconfirmed — secondary reporting only; no OpenAI announcement
Base gpt-3.5-turbo dies today	False — main snapshot runs to Oct 23, 2026

Frequently Asked Questions

Which OpenAI models shut down on September 28, 2026?

gpt-3.5-turbo-instruct, gpt-3.5-turbo-1106, babbage-002, and davinci-002, according to OpenAI’s deprecations documentation. The recommended replacement for each is gpt-5.6-terra.

Is gpt-3.5-turbo itself shutting down today?

No. The alias and the gpt-3.5-turbo-0125 snapshot are listed for October 23, 2026 — not September 28. Do not conflate them with gpt-3.5-turbo-1106, which does die today.

What should I migrate to instead of GPT-3.5?

OpenAI’s named substitute is gpt-5.6-terra at $2 / $12 per million tokens. For many high-volume tasks, gpt-6-luna at $0.10 / $0.50 is the more economical GPT-6 option; for coding agents, gpt-6-sol at $2 / $10 is the stronger peer.

When is OpenAI DevDay 2026 and how can I watch?

Tuesday, September 29, 2026, at Fort Mason in San Francisco. The opening keynote at 10:00 a.m. Pacific (10:30 p.m. IST) is livestreamed free via OpenAI’s DevDay site. In-person applications are closed.

Is OpenAI launching GPT-6 Cyber at DevDay?

Not confirmed. Fortune/Reuters-style secondary reports have said a cybersecurity model could be previewed around DevDay. OpenAI has not published a model card, name confirmation, or agenda item for it as of this writing.