OpenAI Deprecating 6 Codex Models Today (April 14) — Full Migration Guide

April 14, 2026 · 7 min read

Today, April 14, 2026, OpenAI is completing a deprecation cycle that removes six legacy Codex and GPT-3 era models from its API. If your applications, scripts, or integrations call any of the affected model IDs, they will break today. Here is everything you need to migrate cleanly.

Which Models Are Being Deprecated

The following model IDs will return HTTP 404 errors from the OpenAI API after April 14:

Which Models Are NOT Affected

To be clear about what remains available:

How to Find Affected Calls in Your Codebase

Run this command in your project root to surface all deprecated model strings:

grep -r "code-davinci-002\|code-cushman-001\|text-davinci-00\|text-curie-001\|gpt-3.5-turbo-instruct-0914" . --include="*.py" --include="*.js" --include="*.ts" --include="*.json" -l

This outputs every file containing deprecated model IDs. For a quick replacement on most code-focused files:

# Python (replace code-davinci-002 with gpt-5.3-codex)
sed -i 's/code-davinci-002/gpt-5.3-codex/g' your_file.py

# Node.js
sed -i 's/text-davinci-003/gpt-4.1-mini/g' your_file.js

Migration Guide by Use Case

Code Completion (code-davinci-002 / code-cushman-001)

These were the original Codex models, released in 2021. They're substantially outperformed by modern alternatives:

• gpt-5.3-codex — direct successor, significantly better on all code benchmarks (HumanEval: 95% vs 72%)
• gpt-4.1 — better for complex multi-file reasoning and architecture tasks
• o3-mini — best for algorithmic problems and competitive programming

If you're calling the Codex API endpoint directly (rather than the chat completions endpoint), you need to switch to chat completions format. The Codex-style completions endpoint is also being retired with these models.

Text Generation (text-davinci-002/003, text-curie-001)

These are GPT-3 era completion models. The modern equivalents in the chat completions API are faster, cheaper, and better:

• gpt-4.1-mini — replaces text-davinci-002/003 for most tasks at lower cost
• gpt-4.1-nano — replaces text-curie-001 for high-throughput lightweight tasks
• gpt-4.1 — for tasks where text-davinci-003 was genuinely being pushed

The key migration change: text completion models used a prompt-in, completion-out format. Chat models require messages format. Update your API calls to use messages: [{role: "user", content: your_prompt}] instead of passing a raw prompt string.

Instruct Format (gpt-3.5-turbo-instruct-0914)

This snapshot was the last instruct-format (non-chat) GPT-3.5 model. Migrate to gpt-4.1-mini, which offers the chat completions interface with equivalent or better instruction-following at lower latency and cost.

What Happens If You Miss the Deadline

After April 14, API calls to deprecated model IDs return HTTP 404 with error code `model_not_found`. Your application will receive an exception and fail. OpenAI does not auto-redirect deprecated model calls to successor models — the migration is on you.

If you're using an OpenAI client library (Python, Node, etc.), also check that you're on a recent version — older library versions may cache deprecated model lists and not handle the 404 cleanly.

Quick Migration Checklist

• Run grep to find all deprecated model strings in your codebase
• Replace code-davinci-002/code-cushman-001 → gpt-5.3-codex
• Replace text-davinci-002/003 → gpt-4.1-mini
• Replace text-curie-001 → gpt-4.1-nano
• Replace gpt-3.5-turbo-instruct-0914 → gpt-4.1-mini
• Update completion-format calls to chat completions format (messages array)
• Update OpenAI client library to latest version
• Test all affected endpoints before end of day

Frequently Asked Questions