GPT-5.6 has been leaked.

[Introduction] Recently, GPT-5.6 has been leaked! With 1.5 million Tokens and a god-level minimalist UI, it's set to launch urgently next month. Will Altman's "super intelligent agent" shake up the entire Silicon Valley? The AI battle in June has already erupted ahead of schedule.

The long-awaited new OpenAI model, GPT-5.6, is finally coming!

Just three weeks after the release of GPT-5.5, the entire AI community started spreading this news like wildfire yesterday.

Multiple developers discovered a mysterious, unannounced model, gpt-5.6, in OpenAI's Codex backend logs. Its internal development code name is iris-alpha.

Yes, it's still another "slip of the hand," but it reveals a wealth of signals.

If GPT-5.5 three weeks ago was a significant step forward for OpenAI in the programming field, then this time, GPT-5.6, codenamed "iris-alpha," showcases a powerful upgrade in "violent aesthetics."

With the internal code name exposed, successful testing of a 1.5 million context window, and the zero-instruction generation of a god-level minimalist UI... Clearly, GPT-5.6, which we thought we'd have to wait half a year for, and the legendary GPT-5.6 Pro, are set to arrive in June.

Even scarier is that it's not just OpenAI flexing its muscles. Anthropic's Sonnet 4.8, Claude Mythos 1, Google's Gemini 3.5 Pro, and even Musk's Grok 5 are all secretly gearing up, targeting June.

This is no longer just a simple "large model iteration." It's the most intense arms race in the AI field in 2026!

Aesthetic Nuclear Explosion: Will Not Only Programmers but Also Designers Lose Their Jobs?

This time, GPT-5.6's breakthrough in the front-end is a complete qualitative change.

For a long time, there has been a fatal pain point in AI-generated code: "Slop" (garbage code).

The back-end logic generated by large models is often excellent, but when it comes to generating front-end UIs, the results are often less than satisfactory - filled with cumbersome CSS, uncoordinated color combinations, and formulaic "AI-like" grids.

But all of this ends in the face of GPT-5.6.

Tech blogger Leo recently shared a screenshot of an application interface generated by GPT-5.6.

In this screenshot, GPT-5.6 demonstrated what developers call a god-level ability.

In the default state without any detailed prompts for UI guidance, GPT-5.6 independently generated a minimalist note-taking application called "Lumen Notes."

It showcases extremely high-level design aesthetics:

1. Mature grid layout: The spacing between modules is controlled to pixel-perfect precision.

2. Restrained color philosophy: It abandons the high-saturation colors favored by early AIs and uses a lavender-tinted key component combination. Clear color-coded labels make the visual focus extremely clear.

3. Font weight hierarchy and intuitive navigation: The font size and weight have a strong sense of hierarchy, and the navigation design is so intuitive that it doesn't require any instructions.

"UI de-slopification" - this is the most shocking term for front-end developers in this leak.

Just 10 days ago, Leo was complaining about the poor front-end generation results of GPT-5.6. Just a week later, the results have changed dramatically.

This disruptive UI de-slopification effect shows that OpenAI is concentrating its computing power to target and overcome the shortcomings in front-end code generation.

Obviously, they are precisely targeting a long-complained pain point.

Previously, Claude 4.5's Artifacts was considered god-level because it could generate previewable code components in real-time during conversations. Now, GPT-5.6 has brought the aesthetic level to the same level.

Soon, as soon as the GPT-5.6 interface is opened, it can be directly used to generate high-quality commercial-grade application front-ends.

Will front-end programmers lose their jobs? Maybe not. But those lacking top-level design aesthetics will have limited survival space in the face of GPT-5.6.

The "Canary" That Broke the Internet: 1.5M Context Dominates the Charts

This isn't the first time GPT-5.6 has made an appearance.

Well-known developer Haider, while routinely reviewing Codex's routing logs, keenly noticed -

Among the numerous call records pointing to gpt-5.5, there was a clearly mapped entry for gpt-5.6.

At first, Haider thought it was just a trace left by OpenAI during a "canary test" or a system bug because it disappeared quickly.

But in the second week of May, the situation completely changed.

Last week, when developers tried to forcefully specify the call to gpt-5.6 in the code, the system would coldly pop up an error message: "model is not supported."

But in the past few days, multiple sensitive developers successfully called the unannounced GPT-5.6 model in the Codex environment through ChatGPT Pro's OAuth authentication.

Not only did it work, but probe tests also showed that GPT-5.6's context window reached an astonishing 1.5 million tokens!

You know, the context window of the currently powerful GPT-5.5 API is 1.05 million tokens, and if you go through the Codex OAuth channel, it's limited to a mere 400K. GPT-5.6 directly raises this limit by nearly 43%.

Developers' extreme real-world tests in the auxiliary tool OpenCode confirmed that when the input reaches 900,000 tokens, the model can still respond smoothly, and even requests exceeding 1.05 million can be perfectly handled.

In these secret test conversations, the new model revealed that it runs on openai/gpt-5.6, the inference level can be set up to x high, and it supports an extremely fast mode.

Subsequently, the well-known leaker Leo also came forward to confirm: "The development of GPT-5.6 has been fully promoted. The first batch of checkpoints have started internal testing in the past few days, and it's expected to be officially released next month."

Along with the confirmation came three highly mysterious internal test code names: iris-alpha (Iris), ember-alpha (Ember), beacon-alpha (Beacon).

Currently leaked information shows that OpenAI will adopt a "dual-version" strategy this time: the standard version and GPT-5.6 Pro.

- The standard version focuses on a leap in multi-step reasoning ability.

- The Pro version emphasizes the enhancement of the "intelligent agent" workflow.

Obviously, Altman doesn't just want to create a chat dialog box. He wants to create a "super intelligent agent" that can take over all your digital living spaces.

Polymarket has even put its money where its mouth is: the probability of GPT-5.6 being released before June 30 has soared to over 85%.

GPT-5.6 Pro vs. the Big Three: June Will Be a Bloodbath

If 1.5 million Tokens and the god-level UI are GPT-5.6's "external skills," then its truly terrifying "internal skills" are its upgrades in multi-step reasoning and intelligent agent workflows.

A researcher close to OpenAI's internal circle leaked an important piece of information: "The underlying model that supports a recent major mathematical breakthrough is currently widely used as the main driver for daily debugging and technical work within OpenAI."

A senior industry insider commented: "The version number isn't important. What matters is that OpenAI is clearly using it as a daily debugging tool internally. This is usually the clearest signal. When the engineers building AI start trusting it for actual R & D, you know its capabilities have significantly improved."

Obviously, the evolution direction of GPT-5.6 is very clear: not only to pursue high benchmark scores but also to make the user experience smoother when performing long-term tasks, which is the direction of the Agent workflow.

Moreover, it's not just GPT-5.6. According to the latest supply chain and internal leaks, Anthropic and Google are also ready to make their moves in June.

Obviously, we're going to witness a wave of large model releases in June.

Anthropic Claude Sonnet 4.8 has appeared in the Vertex AI backend list, codenamed "Conway," which focuses on persistent background agents and is specifically designed for enterprise-level long-term tasks.

Google Gemini 3.5 Pro is also rumored to be released in June, aiming to regain ground in the multimodal field.

The three cutting-edge models will go head-to-head in the same month. This June will be a real feast of large models!

OpenAI's Large Model Iteration Is Accelerating

In this frenzy of GPT-5.6 leaks, there's a piece of data that's truly thought-provoking: OpenAI's release cycle.

Let's review the timeline of the GPT series models:

- GPT-5: August 7, 2025

- GPT-5.1: November 12, 2025 - an interval of 97 days

- GPT-5.2: December 11, 2025 - an interval of 29 days

- GPT-5.3-Codex: February 5, 2026 - an interval of 56 days

- GPT-5.4: March 5, 2026 - an interval of 28 days

- GPT-5.5: April 23, 2026 - an interval of 49 days

- GPT-5.6: Expected to be released in early June 2026 - an interval of about 40 days

A few years ago, we waited nearly three years from GPT-3 to GPT-4. From GPT-4 to GPT-5, we waited over a year.

Now, OpenAI has compressed the iteration cycle of its main model versions to just 30 to 45 days!

It's clear that the model release cycle is terrifyingly shortening. Behind this is not just the accumulation of computing power but also a qualitative change in reinforcement learning cycles and model architecture optimization.