GPT-5.6 'Iris' Leak: 1.5M Context, June Rumored

GPT-5.6, the anticipated next-generation model from OpenAI, has emerged within the internal logs of the company's Codex programming tool. In early May, developers successfully called the model using Pro-account authentication tokens, uncovering both its codename, 'iris-alpha,' and a key specification: a 1.5-million-token context window. These findings have fueled industry speculation of a release as early as June.

Crucially, OpenAI has made no official announcement regarding the model. GPT-5.6 does not appear on the company's public model directory, and its existence has been reconstructed solely from development environment footprints and backend routing logs. This potential successor surfaces a mere three weeks after the public release of GPT-5.5.

The First Traces in Codex Logs

The first trace appeared on April 28, when developer Haider observed a discrepancy in the Codex routing logs. While nearly all user requests were directed to gpt-5.5, a single query mapped to gpt-5.6. At the time, Haider dismissed the anomaly as a routine canary test or a minor backend bug, and the entry was promptly removed.

However, the signal reappeared this week with greater clarity. While queries targeting gpt-5.6 returned a 'model not supported' error last week, subsequent calls authenticated with Pro-account OAuth tokens successfully returned active responses. The model identified itself as 'openai/gpt-5.6,' offered reasoning configurations up to 'xhigh,' and demonstrated low latency in its fast-response mode.

Such inadvertent disclosures are not unprecedented for OpenAI. Indeed, GPT-5.4 itself surfaced the same way — through an employee's accidental post on X (formerly Twitter) and corresponding traces in the Codex backend. OpenAI appears to be repeating its established pattern of quietly validating new models using canary traffic before a formal release.

1.5-Million-Token Context and 'Slop' Reduction

Regarding technical specifications, the most notable detail is the context capacity. The leaked GPT-5.6 parameters indicate a 1.5-million-token context window, representing a 43% expansion over the 1.05-million-token API limit of GPT-5.5. Compared to the 400,000-token cap previously enforced on the Codex OAuth pathway, the increase is even more pronounced.

Performance data suggests this capacity expansion is functional rather than superficial. According to evaluations published by TestingCatalog, tests conducted on OpenCode demonstrated that the model reliably processed up to 900,000 tokens and successfully accepted inputs exceeding 1.05 million tokens. This indicates the model can analyze extensive documents in a single query without degradation.

Another key development is the model's direct interface generation capabilities. A prominent industry blogger shared a demonstration showing GPT-5.6 rendering a minimalist, functional note-taking application named 'Lumen Notes' without any specific design prompts. Observers noted that the generated interface largely avoided the generic, machine-stamped 'slop' aesthetic common in AI-designed software.

Three Distinct Codenames: Iris, Ember, and Beacon

Intriguingly, the backend logs revealed multiple codenames. Alongside 'iris-alpha,' testers identified tags for 'ember-alpha' and 'beacon-alpha,' suggesting that OpenAI is concurrently evaluating several variants of the GPT-5.6 architecture. Of the three, 'iris-alpha' is seen as the leading candidate for the production model's name, and the industry increasingly treats iris-alpha as effectively another label for GPT-5.6.

This discovery has fueled industry speculation that a 'GPT-5.6 Pro' model may be launched alongside the standard edition. The Pro tier is anticipated to target developers, enterprises, and advanced technical workflows, with optimizations focused on enhanced reasoning capabilities and autonomous agentic performance.

Recent statements from OpenAI researchers provide further context. Internal communications hint that the model responsible for a recent mathematics breakthrough has already become the primary tool for internal debugging and technical development. If this development model is indeed GPT-5.6, it indicates that OpenAI is already using the system to build its own software.

Three Major Models Set to Collide in One Month

The remaining question centers on the launch timeline. Haider observes that OpenAI's release cycle has compressed from annual to semi-annual, quarterly, bimonthly, and now approximately every 30 to 45 days. Based on this accelerating cadence, projecting an early June debut is a plausible estimate.

Significantly, competitors are also preparing major releases for the same month. Anthropic is weighing updates with Claude 4.8 Sonnet, Claude 4.8 Opus, and its frontier model, Mythos, while Google is expected to debut Gemini 3.5 Pro. Mythos, registered by Anthropic in April under the 'Mythos Preview' identifier, is currently active in a restricted cybersecurity program called Glasswing. Crucially, industry reports suggest Mythos still leads in the very benchmarks where GPT-5.6 reportedly lags.

Ultimately, current evidence is limited to routing logs and circumstantial developer reports. Until GPT-5.6 is officially added to OpenAI's public directory, both its release date and the existence of a Pro variant remain speculative. The industry will likely receive definitive answers as June approaches.