GPT-5.5 is barely three weeks old, and outsiders are already poking around its successor.
Multiple developers using ChatGPT Pro OAuth have successfully called an unreleased gpt-5.6 model inside the Codex environment. Probe tests show a context window of 1.5 million tokens — a 43% jump from GPT-5.5's 1.05M on the API.
https://twitter.com/synthwavedd/status/2054594392552255933
The first hint of GPT-5.6 appeared on April 28. Developer Haider, combing through Codex routing logs, noticed that nearly all calls went to gpt-5.5 — except one mapping explicitly pointing to gpt-5.6. He initially chalked it up to a canary test or a bug, since the entry vanished quickly.
But this week, the signals got louder. Developers report that last week, specifying gpt-5.6 returned a "model is not supported" error. This week, with a Pro OAuth token, it just works. The real tell is the context window: GPT-5.5's API context sits at 1.05M tokens (and just 400K through Codex OAuth), but GPT-5.6's probe hits 1.5M — nearly 1.5x the capacity. Tests in OpenCode confirm the model handles over 900,000 tokens reliably, accepting requests exceeding 1.05M. The model identifies itself as openai/gpt-5.6, offers reasoning levels up to xhigh, and fast mode is snappy.
Blogger Leo reports that GPT-5.6 development is in full swing, with initial checkpoints tested internally over the past few days. He expects a release next month, revealing two internal codenames: ember-alpha and beacon-alpha.
Haider analyzes OpenAI's release cadence: from yearly to semi-annual, to quarterly, to bimonthly, and now every 30–45 days. Based on that, he predicts a GPT-5.6 launch in early June. He also believes GPT-5.6 will surpass GPT-5.5 on several benchmarks where it still trails Mythos — arguing that GPT-5.5 is already close, and more reinforcement learning would widen the gap, as OpenAI's RL loops for coding, math, and research are stronger.