✗ This model is not supported when using X-OpenAI-Internal-Codex-Responses-Lite.

One internal header, two error strings, and searchers land on either. Desktop users see This model is not supported when using X-OpenAI-Internal-Codex-Responses-Lite. after an auto-update; CLI users pointing at Azure Foundry get the sibling X-OpenAI-Internal-Codex-Responses-Lite only supports function tools, custom tools, and client-executed tool search as an invalid_request_error. The root is the same Responses-Lite header, so this page covers both.

Symptoms

• The Desktop app errors on new conversations with gpt-5.5-high or gpt-5.5-xhigh, typically when config.toml carries model_provider = "openai_http". Reports on #30224 span builds 26.616.41845 through 26.623.42026.
• CLI 0.144.0 against Azure Foundry fails every turn on gpt-5.6-sol with the tools-parameter invalid_request_error (code unsupported_value), while gpt-5.5 keeps working on the same setup, per #31870.

Cause

The GPT-5.6 entries in Codex’s bundled model catalog ship with use_responses_lite: true, which makes the client send the Responses-Lite header plus a multi-agent “collaboration” tool. OpenAI’s own endpoint accepts that; Azure’s Responses endpoint rejects the tool. On Desktop, the header collides with custom-provider routing — the openai_http provider, or extensions like CC Switch that rewrite conversation handling, per #30224. No OpenAI maintainer had commented on either thread as of July 14, so everything below is community-reported.

Fix

1. Desktop: remove or comment out model_provider = "openai_http" (and any custom-provider block you’re not actively using) in config.toml. Users report this clears the error immediately, at the cost of conversation history tied to that provider.
2. Skip the model_cache.json edit. Setting "use_responses_lite": false there looks like the obvious cure, but reporters confirm the app rewrites it to true on startup.
3. Running CC Switch? One reporter says disabling its unified conversation history fixed it on its own.
4. CLI with Azure or a proxy: run codex debug models > ~/.codex/model-catalog.json, set "use_responses_lite": false and "multi_agent_version": null on the gpt-5.6 entries, then add model_catalog_json = "/absolute/path/model-catalog.json" at the root of config.toml (see the config reference). Multiple users on #31870 confirm this works — though one notes prompt caching stays broken on Azure gpt-5.6 regardless, an upstream limitation.
5. Or roll back: one Azure user reports rolling back to CLI 0.142.5 temporarily bypasses the error.

Affected versions

Desktop builds 26.616.41845 through 26.623.42026 appear in #30224; #31870 is filed against CLI 0.144.0, with 0.142.5 reported clean. Both issues remain open, with activity through July 14, 2026.

Facts on this page verified July 15, 2026.