OpenAI announces GPT-5.6 model series with Sol, Terra, and Luna variants in limited preview

OpenAI has announced a limited preview of its next-generation GPT-5.6 model series, introducing the Sol, Terra, and Luna variants. At the request of the U.S. government, following a June 2, 2026 executive order, initial access is restricted to a small group of trusted partners. The flagship Sol model features advanced reasoning and multi-agent capabilities, setting new benchmarks in coding and cybersecurity. To secure these models, OpenAI dedicated over 700,000 GPU hours to automated red-teaming, implementing a layered safeguard stack that may introduce operational friction for legitimate dual-use security work.

GPT-5.6 model series launch

• ▪GPT-5.6 Sol features a new max reasoning effort mode for deep reasoning and an ultra mode that deploys subagents to accelerate complex work.
• ▪OpenAI plans to launch GPT-5.6 Sol on Cerebras hardware in July 2026, targeting processing speeds of up to 750 tokens per second.
• ▪OpenAI announced a limited preview of the GPT-5.6 model series, which includes three capability-tiered models named Sol, Terra, and Luna.

Government coordination process

• ▪OpenAI previewed the capabilities of the GPT-5.6 models to the U.S. government prior to launching the limited preview.
• ▪At the request of the U.S. government, OpenAI restricted the initial preview of GPT-5.6 to a small group of trusted partners.
• ▪The limited preview coordination follows an executive order on artificial intelligence innovation and security issued by President Donald J. Trump on June 2, 2026.

Cybersecurity benchmark performance

• ▪In evaluations involving Chromium and Firefox, GPT-5.6 Sol identified bugs and exploitation primitives but did not autonomously produce a functional full-chain exploit.
• ▪GPT-5.6 Sol (Ultra) achieved a state-of-the-art score of 91.91% on Terminal-Bench 2.1, which tests command-line workflows requiring planning and iteration.
• ▪On ExploitBench, GPT-5.6 Sol demonstrated performance competitive with Mythos Preview while utilizing approximately one-third of the output tokens.

Layered safeguard architecture

• ▪GPT-5.6 models utilize real-time cyber and biology misuse classifiers that can pause generation for review by a larger reasoning model if a potential violation is detected.
• ▪The real-time safety stack of GPT-5.6 may cause false positives, localized latency spikes, and intermittent request refusals for legitimate dual-use security work.

Automated red-teaming process

• ▪OpenAI dedicated over 700,000 A100-equivalent GPU hours to automated red-teaming to identify universal jailbreaks in the GPT-5.6 models.
• ▪OpenAI complemented its automated red-teaming by collaborating with third-party experts to conduct human expert red-teaming during the preview period.

Model pricing tiers

• ▪GPT-5.6 Sol is priced at $5.00 per million input tokens and $30.00 per million output tokens.
• ▪GPT-5.6 API introduces prompt caching where cache writes are billed at 1.25x the uncached input rate, and cache reads receive a 90% discount.
• ▪GPT-5.6 Terra is priced at $2.50 per million input tokens and $15.00 per million output tokens, while GPT-5.6 Luna is priced at $1.00 input and $6.00 output.

2 sources