Breaking

OpenAI Ships GPT-5.6 in Limited Preview Vetted by US Government

June 26, 2026 at 13:40 EDT

Foundation Models
Security
Policy & Regulation

On June 26, 2026, OpenAI released its new "GPT-5.6" model family as a limited preview restricted to trusted partners, saying its cybersecurity capabilities had reached the level of Anthropic's Mythos Preview. The lineup spans the flagship "GPT-5.6 Sol," the balanced "GPT-5.6 Terra," and the fast, low-cost "GPT-5.6 Luna."

June 26, 2026 · OpenAI

GPT-5.6 "Sol" Hits Parity in the Cyber-Capability Race

OpenAI's new flagship reaches Mythos-Preview-level efficiency on long-horizon security tasks — vulnerability research and exploitation — shipping only as a limited preview to trusted partners.

10m 22s

Reverse-engineering task solved by the model

~$1.73

Cost for that same task

~12 hrs

Estimated time for a human expert

One Task: AI vs Human Expert

Time to solve a reverse-engineering task (each block ≈ 1 hour)

~0.2 hr

GPT-5.6 Sol-class

~12 hrs

Human expert

Capability Benchmarks: A Dead Heat

OpenAI vs Anthropic across two cyber evaluations

Expert-level tasks (AISI)

GPT-5.5

71.4%

Mythos Preview

68.6%

CyberGym

GPT-5.5-Cyber

85.6%

Mythos 5

83.8%

The GPT-5.6 Lineup

Sol

Flagship; internal SOTA across benchmarks

$5 / $30

in / out per 1M tokens

Terra

Balanced; ~2× cheaper than GPT-5.5

$2.50 / $15

in / out per 1M tokens

Luna

Fast, low-cost, high-volume

$1 / $6

in / out per 1M tokens

WELCOMED

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.