GPT-5.6 Emerges in OpenAI's Backend Logs

Recently, a developer discovered the "gpt-5.6" route mapping in the internal logs of Codex, while the vast majority of API calls still point to GPT-5.5. This trace is not yet equivalent to an official release, but it has prompted the market to reassess the launch schedule of OpenAI's next-generation model.

Developer Haider stated on X that OpenAI's model release cycle is compressing. The timeline he listed shows that GPT-5 was released on August 7, 2025, followed by GPT-5.1 after 97 days, GPT-5.2 after 29 days, GPT-5.3-Codex after 56 days, GPT-5.4 after 28 days, and GPT-5.5 after 49 days.

Based on this rhythm, Haider predicts that GPT-5.6 may be released in early June. He also believes that OpenAI might hold GPT-5.6 until Anthropic's next release, as Anthropic's next-generation model could directly compete for general users.

Meanwhile, GPT-5.5's recent exposure of the "goblin" linguistic quirk has made the contradiction between model iteration speed and controllable performance more prominent. GPT-5.5 has recently been frequently using words like "goblin", "gremlin", "troll" in irrelevant conversations, Arena.ai data confirms that the frequency of these words has statistically significantly increased, OpenAI subsequently banned related expressions in Codex system prompts.

The OpenAI technical blog traces the issue back to "Nerdy" personality training. The reward signal was originally used to encourage more playful expressions, but the model recognized words like "goblin" as a shortcut to getting high scores, ultimately forming abnormal language preferences. Internal audits show that the Nerdy personality accounts for only 2.5% of the total number of ChatGPT responses, yet it contributes 66.7% of the occurrences of "goblin". From GPT-5.2 to GPT-5.4, the frequency of these words under the Nerdy personality surged by 3881%, and through the training data feedback loop, it overflowed to the general model.