Inside the OpenAI Codex Leak: GPT‑5.5, Glacier, Heisenberg and What They Reveal

In the early hours, OpenAI suffered a major leak on its Codex platform when an internal staging environment was mistakenly pushed to production, revealing a list of previously unseen "ghost models" such as GPT‑5.5, oai‑2.1, Arcanine, Glacier and Heisenberg.

The leak occurred because the internal testing system used for dog‑fooding was exposed to a few Pro users, showing the internal dropdown menu that is normally hidden. OpenAI quickly patched the vulnerability, but developers had already recorded the full video and captured the model tooltips.

GPT‑5.5 appears as the most prominent entry, paired with the internal identifier oai‑2.1 and described as the "Latest frontier agentic coding model." The use of the term "Agentic" rather than "Language Model" supports rumors that OpenAI’s next‑generation models will move beyond chat interfaces toward autonomous agents capable of coding, debugging and full‑stack deployment.

Rumors suggest GPT‑5.5 may be 3–4× faster than the publicly known GPT‑5.4, and the newly released Image 2 feature could further aid web development tasks.

The Glacier series, listed as glacier-alpha and its variants, carries the tagline "Intelligence that moves continents." This bold phrasing fuels speculation that Glacier represents a massive parameter‑scale model or a fundamentally new architecture, possibly built from modular "cy blocks" hinted at by its naming convention.

Heisenberg is labeled as the "Latest frontier life science research model," indicating OpenAI’s deliberate expansion into a vertical life‑science product line that could compete with DeepMind’s AlphaFold in protein folding, drug discovery and genomics.

Arcanine, named after the legendary Pokémon, bears the description "Frontier model with legendary appetite for starches." While its exact capabilities remain unknown, the whimsical naming suggests an internal engineering joke rather than a serious product announcement.

These revelations echo Sam Altman’s interview six weeks earlier, where he claimed that a new architecture could deliver gains comparable to the Transformer’s impact on LSTM. The leaked models—especially Glacier and oai‑2.1—may be early manifestations of such a self‑evolving architecture.

The technical community quickly split into two camps: one viewing the incident as an inevitable engineering accident given OpenAI’s rapid growth, the other suspecting a deliberate leak to generate hype, noting a pattern of frequent "accidental" disclosures across AI firms.

Ultimately, the Codex page has been restored, but the leaked model lineup hints at a broader, multi‑track AI development strategy at OpenAI, raising the final question: will GPT‑5.5 arrive tomorrow?