GPT-5.5 accessed via Codex backdoor API with new LLM plugin
The article describes how GPT-5.5 — currently unavailable via the official API — can be accessed through OpenAI's semi-official Codex subscription endpoint using a new plugin called `llm-openai-via-codex`, built by reverse-engineering the open-source Codex CLI.
Score breakdown
Developers can access GPT-5.5 today — before its official API launch — by installing the `llm-openai-via-codex` plugin and routing prompts through their existing Codex subscription, with OpenAI's explicit blessing.
- 01GPT-5.5 is available in OpenAI Codex and rolling out to paid ChatGPT subscribers, but its API release is delayed for safety and security reasons.
- 02A new plugin, `llm-openai-via-codex`, was built by using Claude Code to reverse-engineer the open-source `openai/codex` repo and extract its authentication mechanism.
- 03OpenAI's Romain Huet confirmed on March 30th that using the Codex subscription endpoint in third-party tools is officially supported.
GPT-5.5 is now available inside OpenAI Codex and rolling out to paid ChatGPT subscribers, but the standard API release is being held back while OpenAI addresses safety and security requirements for large-scale serving. The article describes building `llm-openai-via-codex`, a plugin for the LLM tool, by having Claude Code reverse-engineer the open-source `openai/codex` repository to figure out how authentication tokens are stored. The plugin lets users route prompts — including image attachments, chat sessions, tool calls, and logged conversations — through the same `/backend-api/codex/responses` endpoint used by tools like Pi and Opencode.
Ethan Mollick's review is cited as finding that GPT-5.5 remains subject to the "jagged frontier," excelling at some tasks while struggling unpredictably at others.
The legitimacy of this approach is backed by public statements from OpenAI: Romain Huet tweeted on March 30th that OpenAI wants people to use Codex and their ChatGPT subscription "wherever they like," explicitly naming JetBrains, Xcode, OpenCode, Pi, and Claude Code as supported contexts. Peter Steinberger also confirmed that the OpenAI subscription is "officially supported" via this mechanism. The broader context involves ongoing tension around agent harnesses like OpenClaw integrating with provider subscriptions — Anthropic blocked OpenClaw from doing so, while OpenAI took the opposite stance after hiring OpenClaw creator Peter Steinberger.
On the model itself, the article runs a "pelican benchmark" — generating an SVG of a pelican riding a bicycle — and finds that using `reasoning_effort xhigh` (which consumed 9,322 reasoning tokens versus just 39 at default) produced a significantly better result, though at the cost of nearly four minutes of generation time. GPT-5.5 is priced at $5 per 1M input tokens and $30 per 1M output tokens — double GPT-5.4's $2.5/$15 — while GPT-5.5 Pro goes further at $30/$180 per 1M tokens. GPT-5.4 will remain available. Ethan Mollick's review is cited as finding that GPT-5.5 remains subject to the "jagged frontier," excelling at some tasks while struggling unpredictably at others.
Key facts
- 01GPT-5.5 is available in OpenAI Codex and rolling out to paid ChatGPT subscribers, but its API release is delayed for safety and security reasons.
- 02A new plugin, `llm-openai-via-codex`, was built by using Claude Code to reverse-engineer the open-source `openai/codex` repo and extract its authentication mechanism.
- 03OpenAI's Romain Huet confirmed on March 30th that using the Codex subscription endpoint in third-party tools is officially supported.
- 04