Codex GPT-5.5 model not found: diagnose the 404

If Codex reports that gpt-5.5 does not exist or you do not have access to it, check the combination of model ID, provider and authentication before changing your subscription. A valid ChatGPT login does not establish access through an API key, and a valid provider key does not establish that the provider accepts the model ID in your configuration.

This guide is for a Codex CLI session showing an error such as:

unexpected status 404 Not Found: The model `gpt-5.5` does not exist or you do not have access to it.

The wording identifies the requested model, not the cause. It is not evidence that GPT-5.5 was retired. The steps below are based on the official Codex configuration and authentication documentation checked on September 8, 2026; they do not claim that GPT-5.5 is available to every account or provider.

Start with the route, not a reinstall

Run these inspection commands in the same terminal where Codex fails:

codex --version
codex login status
codex --help

Record the version and authentication method. Also record the directory you launched from, any --model, --profile or -c arguments, and whether an IDE launched the session. Do not share API keys, access tokens or the contents of auth.json.

codex login status describes authentication; it is not a complete report of the endpoint selected by a custom provider. Read it alongside the provider configuration.

OpenAI documents the distinction between ChatGPT subscription access and API-key usage-based access. Keep those routes separate throughout the diagnosis.

Find the setting Codex actually reads

A common troubleshooting mistake is editing the user configuration while the launch command still selects another model. Inspect the relevant configuration files locally and record only the fields you need:

model
model_provider
openai_base_url
model_providers.<provider>.base_url
model_providers.<provider>.env_key
model_providers.<provider>.requires_openai_auth

According to the current configuration basics, precedence runs from highest to lowest:

1. CLI flags and --config overrides.
2. Trusted project configuration, with the file closest to the working directory winning.
3. The profile file selected with --profile.
4. User configuration at ~/.codex/config.toml.
5. System configuration, then built-in defaults.

There is an important restriction: current advanced configuration documentation says project configuration ignores provider-related keys such as model_provider, model_providers and openai_base_url, with a startup warning. Keep provider definitions in user-level configuration. A project can still override other allowed settings, including the model, so inspecting only one file can miss a mismatch.

The current docs describe profile files as ~/.codex/profile-name.config.toml. Older tutorials may use a different profile layout. Check your installed version before copying either format.

Repair the connection you intended to use

If you intended ChatGPT sign-in

Confirm the account and workspace, then select a model actually available in that session. Remove an obsolete model override from the launch command or the configuration layer that supplied it. If the model is not offered, check account availability rather than assuming that manually entering gpt-5.5 grants access.

Only reauthenticate when the active account is wrong or authentication has failed. Reinstalling Codex or deleting credentials is not a necessary first step for a model lookup error.

If you intended an OpenAI API key

Confirm that the request uses the intended OpenAI endpoint and API account/project. Compare the requested ID with that account’s current model availability. Check openai_base_url for an old proxy override before concluding that OpenAI rejected the request.

API access and billing are separate from included ChatGPT plan credits. If another model works with the same key, that is useful evidence about the route, but it does not prove access to GPT-5.5.

If you intended a third-party provider

Treat the URL, model identifier and key as one matched set. A provider may use a namespaced model ID; use its documented value rather than adding or removing a prefix by guesswork.

This is a configuration template, not a working endpoint or an assertion of GPT-5.5 availability. Replace both placeholders with values from your provider. Put these entries in user-level configuration, merging with existing sections rather than duplicating them:

model = "REPLACE_WITH_PROVIDER_MODEL_ID"
model_provider = "diagnostic_provider"

[model_providers.diagnostic_provider]
name = "My provider"
base_url = "https://api.example.com/v1"
env_key = "PROVIDER_API_KEY"
requires_openai_auth = false
wire_api = "responses"

The selected provider must support the Responses API; the current Codex configuration reference lists responses as the only supported wire_api value. A provider offering only Chat Completions cannot use this template. Set PROVIDER_API_KEY through your usual local secret-management method; do not paste the secret into a shared TOML file.

requires_openai_auth defaults to false; it is explicit here to keep this template on the provider-key route. If an existing provider definition sets it to true, Codex uses OpenAI authentication and ignores env_key. Check this setting when a provider key appears to have no effect.

Before editing, save a copy of the affected configuration. Keep unrelated settings. After editing, start a new session from the same directory and run a small request that does not ask Codex to edit files. API usage may be billed. Restore the saved configuration if the change breaks a previously working route.

For a broader setup walkthrough, see custom model providers in Codex. If you use Ofox, check the current model catalog for the exact ID and supported connection before configuring it.

Use the next response to narrow the cause

If you compare a direct API request with Codex, use the same host, key, model and Responses endpoint. A successful Chat Completions request to another host does not isolate a Codex problem. A direct success on the matched route shifts attention toward Codex configuration or request differences; it does not prove which difference caused the error.

A local model metadata ... not found warning is also distinct from an HTTP 404. Capture the final server response rather than assuming those messages have the same cause. For non-Codex SDKs and Azure cases, use the general OpenAI model-not-found guide.

What to include if you still need support

Send a redacted report containing the Codex version, operating system, authentication type, selected model/provider, endpoint host and path, working-directory context, relevant startup warnings and the exact final error. Include the request ID if the provider supplies one. Exclude secrets and private prompts.

That report lets support distinguish account availability from a stale override or provider mismatch. It is a better next step than changing the account, model and endpoint all at once and losing the original evidence.

Sources

• OpenAI Codex authentication
• OpenAI Codex configuration basics
• OpenAI Codex advanced configuration
• OpenAI Codex configuration reference
• OpenAI Codex CLI reference