OpenAI Error: model_not_found — Model Not Found

import openai

try:
    response = openai.chat.completions.create(
        model="gpt-4-turbo-preview-2024",   # ← deprecated alias
        messages=[{"role": "user", "content": "Hello"}],
    )
except openai.NotFoundError as e:
    # e.status_code == 404
    # e.code == 'model_not_found'
    # e.message: "The model `gpt-4-turbo-preview-2024` does not exist or you do not have access to it."
    ...

model_not_found is OpenAI’s catch-all 404 for any model identifier that doesn’t resolve to something your API key can call. The error message is intentionally ambiguous — “does not exist or you do not have access” — because OpenAI doesn’t want to confirm the existence of gated models to unauthorised callers. That ambiguity is the source of most debugging time spent on this error.

The fastest path to clarity is client.models.list(). If the model is in the response, you have access — recheck spelling, casing, and which endpoint you’re calling. If it isn’t, check OpenAI’s deprecation page (the model may have been retired) and your account’s usage tier (the model may be gated). Combine those two checks and you’ll resolve 90%+ of model_not_found errors in under a minute.

Why this happens

• Typo or wrong casing in model name. Model IDs are case-sensitive and exact. `GPT-4o` won't match `gpt-4o`. `gpt4o` (no hyphen) won't match. `gpt-4-turbo` is now an alias to a specific snapshot, but `gpt-4-turbo-preview` is deprecated. A single character off and the API returns 404.
• Deprecated or retired model. OpenAI deprecates older models on a rolling schedule. `text-davinci-003`, `gpt-3.5-turbo-0301`, `gpt-4-32k`, and the original `gpt-4-vision-preview` have all been retired. Calls to retired models return `model_not_found`. The deprecation page lists current dates.
• Tier-gated model and your org isn't qualified. Newer flagship models often gate by usage tier. `o1` historically required tier 5 ($1k+ paid). Some experimental models are limited to tier 1+ paid customers. Free-tier orgs can't see `gpt-4o` at all in some periods. The error is the same `model_not_found` even though the issue is access, not existence.
• Wrong API endpoint for the model type. Calling `/v1/chat/completions` with an embedding model (`text-embedding-3-large`) returns `model_not_found` even though the model exists — it just doesn't serve chat. Embedding models go to `/v1/embeddings`, image models to `/v1/images/generations`, audio to `/v1/audio/*`.
• Fine-tuned model from another org or project. Fine-tuned model IDs (`ft:gpt-4o-mini-2024-07-18:org-abc:custom-name:abc123`) are scoped to the org/project that created them. Calling them from a different org returns `model_not_found`. Project-scoped fine-tunes are also invisible to other projects in the same org.

How to fix it

Fixes are ordered by likelihood. Start with the first one that matches your context.

from openai import OpenAI
client = OpenAI()

models = client.models.list()
chat_models = [m.id for m in models.data if m.id.startswith(('gpt-', 'o1', 'o3'))]
print('Available chat models:', sorted(chat_models))

# Defensive caller:
def pick_model(preferred, fallback="gpt-4o-mini"):
    available = {m.id for m in client.models.list().data}
    return preferred if preferred in available else fallback

# WRONG: chat completion with an embedding model
# client.chat.completions.create(model="text-embedding-3-large", ...)

# RIGHT: embeddings endpoint
from openai import OpenAI
client = OpenAI()
emb = client.embeddings.create(
    model="text-embedding-3-large",
    input="Hello world",
)

Detection and monitoring in production

Treat `model_not_found` as a critical config error, not a transient one. Alert on first occurrence per deploy — it usually means a deploy shipped with an outdated model name or a tier change broke access. Log `error.code` and the requested model name together so you can spot pattern (typo on one model vs deprecation across the board).