GPT-5.6 iris-alpha surfaces in OpenAI Codex backend logs hinting at a 1.5M token context window
OpenAI Codex backend logs reveal GPT-5.6 iris-alpha with a 1.5M token context window and direct UI generation via Lumen Notes.
What just surfaced
Developers digging through OpenAI Codex backend telemetry surfaced references to an unannounced model line internally codenamed **iris-alpha**, almost certainly GPT-5.6, alongside sibling variants called **ember-alpha** and **beacon-alpha**. The headline number from the discovery: a native **1.5 million token context window**, roughly 43% larger than the 1.05M ceiling on the current GPT-5.5 API.
A single number rarely matters on its own, but this one does. A 1.5M-token window means a model can hold an entire codebase, a long legal contract, or weeks of agent trajectory in working memory without summarization passes. For agentic workflows, persistent context is the bottleneck right now, not raw capability. Pushing past 1M without falling off a quality cliff is the harder problem.
Lumen Notes and AI-generated frontends
A second piece is more interesting than the context number. Screenshots reference a feature called **Lumen Notes**, in which the model emits a production-grade rendered UI rather than just code or markdown. If that holds up, GPT-5.6 is being positioned to ship usable software artifacts directly, not just snippets for a developer to assemble.
That is the actual frontier right now. Long context is table stakes. Generating a coherent, styled, interactive surface in one shot is the capability that turns a model into a product.
Why this leaked now
The timing matters. Anthropic Claude Sonnet 4.8 and Google Gemini 3.5 Pro are both rumored for June launches. OpenAI dropping iris-alpha references into Codex telemetry, intentionally or not, primes the market for a competitive response window. Expect a formal announcement inside the next four to six weeks.
- 1.5M token native context window
- Three internal variants tested simultaneously
- Direct UI generation, not code emission
- Discovered via Codex infrastructure logs, not a press leak
- No official OpenAI confirmation yet
What to watch
The real test for GPT-5.6 will not be the 1.5M number. It will be effective context: does the model actually reason over the full window, or does quality degrade past 500K like most current long-context systems. Watch for needle-in-a-haystack benchmarks and real agent traces when the model lands, not the spec sheet.
Source: Wavespeed