GPT-5.6 slips to July as OpenAI goes quiet
The model that was supposed to land this week didn't. Here is what slipped, and what is reportedly in it.
GPT-5.6 was widely tipped for a late-June release. It didn't arrive. Prediction-market odds for a 22-28 June launch fell from around 83% to 18% as the week closed. The new expectation is July.
Chief scientist Jakub Pachocki reportedly told staff internally that 5.6 will be a meaningful improvement over 5.5. Leaked details point to a 1.5 million token context window, up from 400K, plus improved agentic workflows, faster Codex responses and better front-end code generation.
There is also a redesigned reward-audit pipeline, reportedly a response to what insiders have called the Goblin Incident: a reward-model miscalibration in GPT-5.5 that produced problematic behaviour patterns. A flagship model shipping with a known reward-model fault, then quietly being patched, is the kind of thing that deserves more attention than a release-date wobble.
The likely reason for the delay is dull but plausible. OpenAI filed its S-1 on 8 June and is now in an IPO quiet period. Marketing a headline model launch while under communication restrictions is awkward, so 5.6 may ship as a low-key technical release rather than a stage show.
A 1.5M context window is the headline most developers will care about. It is the difference between feeding a model a few files and feeding it a whole codebase or a stack of long documents in one go, without the awkward chunking dance. If the agentic and Codex improvements are real, that is a genuinely useful release whenever it lands.
For now it is wait and see, but with a date attached. July is close. The more interesting question is whether OpenAI uses the launch to address the Goblin Incident openly, or hopes everyone has moved on by the time the IPO roadshow starts.