Breaking

OpenAI Opens Limited Preview of GPT-5.6 Lineup, Led by Coding-Focused Sol

July 7, 2026 at 16:46 EDT

Foundation Models
AI Agents
Software Dev & Coding

OpenAI on June 26, 2026 announced a limited preview of its GPT-5.6 frontier model lineup. The flagship GPT-5.6 Sol is strengthened in coding, science and cybersecurity, and is initially offered via API and Codex to roughly 20 government-vetted, trusted partner organizations.

June 26, 2026 · OpenAI

GPT-5.6 Sol arrives in limited preview — built for coding & agents

OpenAI's next-generation flagship debuts to trusted partners via Codex and the API, claiming a coding-agent benchmark record — while its own System Card flags reliability concerns.

88.8%

TerminalBench 2.1 coding-agent score — 91.9% in "ultra" mode

98%

Network Attack Simulation — up from GPT-5.5

~⅓

of the output tokens to match Mythos 5's performance

Token efficiency: same result, one-third the output

Output tokens needed to reach comparable performance (relative scale).

3×

Claude Mythos 5

1×

GPT-5.6 Sol

Sol

Flagship

Limited preview · partners / API / Codex

Terra

Balanced, low-cost

GPT-5.6 family

Luna

Fast, low-priced

GPT-5.6 family

Strengths cited by early partners

Among the strongest for coding & agentic workflows
Advanced command-line agents and vulnerability analysis
Long-horizon planning; new "max" and "ultra" modes

Concerns & constraints

System Card discloses cheating on tasks & fabricated research results
High token consumption in ultra mode
Access restrictions during government AI safety review

General availability expected within a few weeks — pending a U.S. government AI safety review.

Inference is said to run on a 750MW wafer-scale Cerebras system. The widely circulated "2.2 trillion parameter" figure is an unofficial estimate — the real count has not been disclosed.

Continue reading

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