OpenAI merged Codex into the main model. Stop waiting for a specialized coding brain.

The dedicated OpenAI coding model is dead. It has been dead for a while, but nobody bothered to say it out loud until today.

In a brief quote captured by Simon Willison, OpenAI's Romain Huet confirmed that the company will not be releasing a GPT-5.5-Codex model. The reason is structural. Since GPT-5.4, OpenAI quietly unified Codex and the main flagship model into a single system. There is no separate coding line anymore. GPT-5.5 simply takes that unified architecture further, with a focus on agentic coding and general computer use.

Launch day theater trained us to expect specialized model variants. We got used to the idea that there is a general brain for writing polite emails and a specific, highly-tuned brain for writing Python. Translation: that bifurcation was a temporary artifact of older, weaker infrastructure. As models scale, separating logical reasoning from natural language generation no longer makes sense. Code is just another language the main model speaks natively.

What changed: you are no longer routing between specialized intelligence tiers.

For teams building agentic workflows, this simplifies the architecture board. You do not need to maintain separate evaluation suites to figure out if you should hand a task to the generalist or the specialist. The generalist is the specialist. When GPT-5.5 executes a script or attempts a complex “computer use” task, it is not switching tracks. It is applying its primary reasoning engine directly to the problem.

The catch: when the coding model and the main model are identical, regressions in one domain can quietly infect the other. If a future fine-tuning pass makes the model safer or more conversational, you have to verify that it did not simultaneously make its Python output more bloated.

This is a reality check for anyone still waiting for a magical, code-only breakthrough from OpenAI. The breakthrough is that they stopped drawing a boundary between writing a sentence and compiling a function. Stop looking for a specialized endpoint. Point your evaluation scripts at GPT-5.5 and accept that this is the only brain you get.