just three weeks after the official release of gpt-5.5, several developers accidentally discovered an unpublished new model—gpt-5.6, internally codenamed iris-alpha—in the backend logs of openai’s codex service. this revelation quickly sparked intense discussion within the tech community, yet as of now, openai has still not issued any official statement on the matter.
the key highlight of this disclosure lies in a dramatic leap in context length: gpt-5.6 supports a context window of up to 1.5 million tokens, nearly a 43% increase compared to the 1.05 million tokens offered by the gpt-5.5 api. real-world testing shows that, even when processing inputs of 900,000 tokens, the model continues to deliver stable responses in tools like opencode. even more astonishing is that, when faced with extremely long texts exceeding 1.05 million tokens, its output quality and latency remain well-controlled. this breakthrough significantly expands the model’s applicability across highly complex tasks such as legal document analysis, full‑scale codebase comprehension, and end-to-end project tracking.
the front-end generation capabilities have also undergone a critical transformation. according to leaked screenshots, gpt-5.6 can autonomously construct a lightweight note-taking app interface called lumen notes using only very simple prompts—the layout employs a precise grid system, with a low-saturation lavender as the primary color, clear typographic hierarchy, and intuitive navigation logic. industry experts describe this advancement as “ui de‑slopping,” marking a shift from previously crude, unusable front-end code generated by large models toward a mature stage where such code is ready to use out of the box and conforms to industrial design standards.
the logs also reveal additional codenames, such as ember-alpha and beacon-alpha, suggesting that gpt-5.6 is not a single version but rather a suite of technologies tailored for different scenarios. preliminary speculation indicates that this series may be divided into standard and pro editions, each enhancing capabilities like multi-hop reasoning, intelligent agent coordination, and advanced code diagnostics. notably, this technological race is far from being a solo effort by openai: anthropic is set to launch claude sonnet 4.8, google’s gemini 3.5 pro is poised to debut, and xai’s grok 5 is also targeting the same timeframe—june 2026. as leading global ai companies collectively unveil their next-generation flagship models, the arms race among large models has entered a feverish phase.
