GPT-5.6 Successfully Tested in Codex, Expected Release in June

According to monitoring by Beating, just three weeks after the release of GPT-5.5, its successor GPT-5.6 has already been successfully tested by external developers. Multiple developers accessed the yet-to-be-announced GPT-5.6 model in the Codex environment through ChatGPT Pro’s OAuth authentication. Probe tests indicate that the context window has reached 1.5 million tokens, an increase of approximately 43% compared to the 1.05 million tokens of the GPT-5.5 API. The first signs of GPT-5.6 appeared on April 28. Developer Haider discovered while reviewing Codex routing logs that most calls pointed to GPT-5.5, but one entry explicitly mentioned GPT-5.6. He later revised his assessment, suggesting it was more likely a canary test or a bug, as the entry quickly disappeared. However, significant changes have been noted since this week. Some developers found that last week, specifying GPT-5.6 resulted in an error stating ‘model is not supported,’ but this week it successfully passed using Pro’s OAuth. More critically, the context window has dramatically increased: the API context for GPT-5.5 is 1.05 million tokens, while Codex OAuth only allowed 400,000 tokens, whereas GPT-5.6’s probe directly reached 1.5 million tokens, nearly a 1.5-fold increase. Developers’ tests in OpenCode also confirmed that the model responds normally above 900,000 tokens, and requests over 1.05 million tokens are also accepted. The model identifies itself as running on openai/gpt-5.6 during conversations, with inference levels adjustable to xhigh, and fast mode available with quick speeds. Blogger Leo stated today that the development of GPT-5.6 is progressing rapidly, with the first batch of checkpoints beginning internal testing in the past few days, and a release is expected next month, alongside the disclosure of two internal codenames: ember-alpha and beacon-alpha. Haider analyzed OpenAI’s iteration pace: moving from annual updates to semi-annual, quarterly, bi-monthly, and now down to 30-45 days, suggesting that GPT-5.6 should be released in early June. He also predicts that GPT-5.6 will surpass GPT-5.5 on several benchmarks where it currently lags behind Mythos, reasoning that GPT-5.5 is already very close, and with more reinforcement learning, the gap can be widened, as OpenAI’s RL cycles in coding, mathematics, and scientific research are stronger. The probability of GPT-5.6 being released before June 30 on Polymarket is currently around 85%.