benchrWho needs to migrate
GPT-4o launched May 13, 2024 and carried more real-world traffic than any model of its generation. It made voice conversations normal, made vision input cheap, and for a year and a half it was the default answer to "which OpenAI model should I use." When OpenAI tried swapping ChatGPT users off it in 2025, the backlash forced a partial walk-back. Models don't usually get petitions.
The API doesn't run on sentiment, though. The April 2026 deprecation notice put the launch snapshot on a clock, and after October 23 the requests just error out.
Read your model string carefully
The retirement names gpt-4o-2024-05-13 — the original launch snapshot, which billed $5 input / $15 output per million tokens. Later snapshots billed differently, and bare aliases like gpt-4o resolve to whatever OpenAI currently points them at. Two practical consequences:
First, if you pinned the May 2024 snapshot for output consistency, you've been paying 2024 prices for two years. Whatever you migrate to, you'll likely spend less per token. Second, don't assume an alias saves you. Audit which snapshot your traffic resolves to in the usage dashboard, because "it still works today" tells you nothing about October 24.
Replacement options and listed prices
| Option | Input | Output | Pick it when |
|---|---|---|---|
| GPT-5.5 (OpenAI's official pick) | $5.00 | $30.00 | Agentic work, computer use, 1M context: the loads GPT-4o was never good at anyway |
| GPT-5 (benchr's cost pick) | $1.25 | $10.00 | Chat, drafting, summarization, everyday vision: the actual GPT-4o workload |
| GPT-5.6 Terra (benchr's current-generation pick) | $2.00 | $12.00 | You want OpenAI's current general tier and a 1.05M window; costs more than GPT-5 on both sides |
| GPT-5.6 Luna | $0.20 | $1.20 | High-volume classification, routing, extraction where 4o was overkill; cheaper than GPT-5 Mini on both sides, with a 1.05M window |
| GPT-5 Mini | $0.25 | $2.00 | You already run it and its 400K window is enough |
The official migration path points up-market, which is how provider migration paths usually point. Price the workload instead. Most GPT-4o traffic is conversational and short-form; GPT-5 handles it at a quarter of 4o's original input price and two-thirds of its output price, and the GPT-5 pricing page shows the monthly math at real volumes. GPT-5.6 Terra, at $2 input / $12 output per 1M tokens for GPT-5.6 Terra, costs more than GPT-5 but brings OpenAI's current generation and a 1.05M-token window; the GPT-5.6 pricing page covers the family. If you've outgrown OpenAI entirely, the rankings put Gemini 3.8 Flash and DeepSeek V4-Pro in the same conversation.
ChatGPT and API retirement dates differ
This page covers the API. ChatGPT's model picker follows its own schedule. OpenAI manages consumer model transitions through release notes, and the 4o-era nostalgia debate there is separate from your production code. If your exposure is a Custom GPT or an Assistants-based integration, note that the Assistants API itself sunsets August 26, 2026, which is a tighter deadline than the model's.