An identifier gpt-5.6 showed up in OpenAI Codex backend traces over the weekend, alongside three internal codenames — iris-alpha, ember-alpha, and beacon-alpha. Polymarket is currently sitting at 85%+ odds for a public release before June 30. Here is what was reportedly leaked, what to take with a grain of salt, and what it means if any of it ships.
The GPT-5.6 Leak
From Codex traces and a handful of corroborating Discord screenshots:
| Field | Reported Value |
|---|---|
| Identifier | gpt-5.6 |
| Internal codenames |
iris-alpha, ember-alpha, beacon-alpha
|
| Context window | 1.5M tokens (+43% over GPT-5.5) |
| Tiers | Standard + GPT-5.6 Pro |
| Pricing rumor | 2-3x cheaper than Anthropic Mythos at same tier |
| Focus | Agentic workflows + front-end generation |
The three codenames suggest a routed-ensemble approach — different sub-models for different task classes — though that is interpretation, not leaked text.
The Claude Mythos Counter-Leak
Anthropic apparently ran Mythos against an 18-benchmark internal suite. Mythos won 17 of 18:
| Benchmark | Mythos | Comparison |
|---|---|---|
| SWE-bench Verified | 93.9% | Opus 4.6: 80.8% |
| SWE-bench Pro | 77.8% | GPT-5.4: 57.7% |
| Terminal-Bench 2.0 | 82.0% | — |
| USAMO 2026 | 97.6% | — |
| GPQA Diamond | 94.5% | — |
| BrowseComp | 86.9% | — |
| Cybench | 100% | saturated |
Mozilla reportedly found 231 zero-day vulnerabilities testing Mythos against Firefox — 10x more than the previous Claude model could surface in the same harness.
The Microsoft Slip
A slide at Build 2026 briefly showed Mythos training compute at approximately 6.1 × 10²⁷ FLOPs — roughly 300x what GPT-4 was trained on. The slide was pulled from the recording shortly after, but several attendees screenshotted it.
How To Read All Of This
- Codex backend identifiers are real signals — they have predicted releases before (GPT-4o, GPT-5-mini).
- The Mythos scorecard is directional, not definitive — internal benchmarks are always optimized for internal context.
- The Microsoft slide is the weakest signal — a single unverified screenshot.
- Polymarket odds above 85% are meaningful — that market has been well-calibrated on OpenAI releases.
What To Watch
- Any OpenAI API pricing page update showing a
gpt-5.6SKU - Anthropic's model card for Mythos (expected alongside release)
- Codex UI changes in the next two weeks
This will resolve quickly. Neither company has a reason to delay past Q2 at this point.
All benchmark figures above are from leaked documents, not official model cards. Treat them as directional until vendor APIs or official releases confirm them.
Want more AI news like this in 60 seconds? Watch the 60-second walkthrough
If you want to stay ahead of every frontier-model leak, benchmark drop, and release signal, gptaiclips.com aggregates the signal from developer Discords, Polymarket, and API changelogs into one daily digest — free to follow.
Top comments (0)