Calling GPT-5.5 via Codex

GPT-5.5 was announced on April 23, but the API was left out. OpenAI explained that more deployment safeguards are needed, and said it would soon bring GPT-5.5 and GPT-5.5 Pro to the API. When using the model, the API is preferred to avoid the hidden system prompt effects of ChatGPT or other agent harnesses.

The recent point of contention is how agent harnesses like OpenClaw and Pi connect to the subscription products of major model providers. OpenClaw had directly added Anthropic subscription integration but was blocked, and after OpenAI hired OpenClaw's creator Peter Steinberger, it said it would allow its own subscription integration in the way Codex CLI uses it. When Jeremy Howard asked whether /backend-api/codex/responses was supported, Romain Huet explained that the intent is to let Codex and ChatGPT subscriptions be used anywhere—apps, terminal, JetBrains, Xcode, OpenCode, Pi, Claude Code—and Steinberger also replied that OpenAI subscriptions are officially supported.

Building on this development, Simon Willison reverse-engineered the openai/codex repository using Claude Code to figure out how auth tokens are stored, and created llm-openai-via-codex, a plugin for LLM. This plugin pulls in an existing Codex subscription to call openai-codex/gpt-5.5.

The usage steps are simple.

• Install the Codex CLI and log in with an OpenAI plan.
• Install LLM with uv tool install llm.
• Add the plugin with llm install llm-openai-via-codex.
• Send a prompt with llm -m openai-codex/gpt-5.5 'Your prompt goes here'.

Existing LLM features are also preserved. Image attachments, llm chat, llm logs, and llm --tool all work as before, and an actual test generated "an SVG of a pelican riding a bicycle." The default output was somewhat broken, but giving -o reasoning_effort xhigh took about 4 minutes instead and produced a much better result, using 9,322 reasoning tokens. Compared to the default run's 39 tokens, xhigh produced a more elaborate SVG with a more CSS-heavy approach.