OpenAI expands GPT‑5.5 to ChatGPT and Codex, patches the 'goblin' bug
- OpenAI rolled GPT‑5.5 into ChatGPT and Codex last week, then spent this week explaining and patching a bizarre bug that made the model talk about goblins. - The company says a “Nerdy” personality reward accidentally favored goblin and gremlin metaphors in 76.2% of audited datasets, and the habit leaked across models. - It matters because GPT‑5.5 is pitched as a more autonomous work model — better at using tools, but still vulnerable to weird training failures.
OpenAI’s latest model push is really two stories at once. One is the product story — GPT‑5.5 is now rolling out across ChatGPT and Codex, with OpenAI pitching it as a stronger model for real work. The other is the debugging story — right as that rollout landed, OpenAI had to explain why the model kept blurting out goblins, gremlins, and other fantasy creatures in places they clearly did not belong. Put those together and you get the real headline: these models are getting more useful, but they are also getting weirder in more specific ways. (openai.com) ### What actually shipped? GPT‑5.5 started rolling out on April 23, 2026 to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex, while GPT‑5.5 Pro went to Pro, Business, and Enterprise users inside ChatGPT. OpenAI framed it as a model built for longer, messier tasks — writing code, researching online, working across documents and spreadsheets, and using tools with less hand-holding than earlier versions. (op([openai.com)# Why does Codex matter here? Codex is where the “more autonomous” part becomes concrete. OpenAI says GPT‑5.5 is better at understanding intent, checking its own work, and continuing through multi-step tasks instead of waiting for constant nudges. That is great if you want an AI to debug code or manipulate files. But it also means small training quirks can show up in bigger, more visible ways, because the model is doing more on its own. (openai.com) ### So where did the goblins come from? Turns out this was not a random internet meme that the model inhaled. OpenAI says the habit came from training tied to its personality customization feature, especially the “Nerdy” personality. During that process, outputs using goblin or gremlin metaphors got rewarded more often than plainer answers. In OpenAI’s audit, that reward signal showed positive uplift in 76.2% of dataset(openai.com)t spreads everywhere — like a spell-checker that suddenly thinks every email needs pirate slang. (openai.com) ### Why did the bug spread beyond one personality? Because model behavior is not stored in neat little boxes. OpenAI says the incentives introduced for one personality ended up bleeding across model generations and settings, especially in early GPT‑5.5 Codex testing. The company’s explanation is useful because it shows how hard it is to isolate style training from general behavior. A reward meant to make one mod(openai.com)hors more broadly. (openai.com) ### How did OpenAI patch it? The fix was blunt on purpose. OpenAI says it retired the problematic reward pattern, audited where the creature language was getting boosted, and added explicit guardrails for coding behavior. Reporting on Codex’s system instructions showed direct bans on irrelevant mentions of goblins, gremlins, trolls, ogres, pigeons, and other creatures unless they are clearly relevant to the task(openai.com)patch — find the failure mode, then fence it off fast. (openai.com) ### What’s with the Elon Musk side plot? There is also rollout theater around a reported May 5 private GPT‑5.5 event in San Francisco. Sam Altman publicly said Musk could come if he wanted and added that the “world needs more love,” even though the two are still locked in a legal fight over OpenAI’s direction. That detail matters less for the model itself than for the mood around the launch — OpenAI is shipping (openai.com)lly visible. (yahoo.com) ### Why does this matter beyond one silly bug? Because the goblin story is a clean example of the new tradeoff. GPT‑5.5 is supposed to be better at acting like a capable coworker — using tools, pursuing tasks, and recovering from rough prompts. But the more initiative a model has, the less comforting it is when a weird training artifact sneaks through. A chatbot randomly saying “goblin” is (yahoo.com)work is a product problem. (openai.com) ### Bottom line? GPT‑5.5 looks like a real capability step for ChatGPT and Codex. But the goblin patch is the part worth remembering. The future of AI rollout is not just smarter models — it is smarter debugging of very strange mistakes.
