OpenAI has announced a limited preview of the GPT‑5.6 model family, consisting of GPT‑5.6 Sol (the highest‑capability model), GPT‑5.6 Terra (a balanced model for everyday work), and GPT‑5.6 Luna (a fast, lower‑cost option). The company said it plans to make the models generally available in the coming weeks, but is starting with a small group of trusted partners whose participation has been shared with the U.S. government.
Performance and capabilities
According to OpenAI, Terra delivers performance competitive with GPT‑5.5 while costing roughly half as much. Luna offers strong capability at the company’s lowest cost tier. OpenAI describes GPT‑5.6 Sol as their most capable model to date, showing particular improvements in agent‑style, multi‑step tasks.
The company shared evaluation highlights: Sol sets a new state of the art on Terminal‑Bench 2.1 (command‑line workflows), improves on GeneBench v1 for long‑horizon genomics and quantitative‑biology analyses while using fewer tokens, and advances cybersecurity performance on ExploitBench and ExploitGym. OpenAI said it will publish an expanded suite of evaluation results when the models are broadly released.
New modes and latency handling
GPT‑5.6 introduces a new "max reasoning effort" option to give Sol more time to reason deeply, and an "ultra mode" that uses subagents to accelerate complex tasks.
OpenAI also announced that GPT‑5.6 Sol will be offered on Cerebras hardware in July at up to 750 tokens per second, with initial access limited to select customers as capacity expands.
Safety and phased release
OpenAI emphasizes that GPT‑5.6 Sol ships with the company’s most robust safety stack to date. They spent multiple weeks probing weaknesses, pressure‑testing the system under real‑world attack scenarios, and hardening protections. The company says safeguards are configured to scale with model capability.
As part of ongoing engagement with the U.S. government and at the government’s request, OpenAI previewed its plans and the models’ capabilities ahead of the launch and is beginning with a limited preview for a small group of trusted partners. OpenAI frames this short‑term step as the strongest path to broader availability in the coming weeks while it works with the Administration on a cyber Executive Order framework and a repeatable process for future releases.
Layered defenses
OpenAI notes that no single safeguard is sufficient against determined or adaptive misuse. For the GPT‑5.6 preview they apply layered protections that vary across models: model‑level refusal behavior, real‑time generation checks, account‑level signals, differentiated access, monitoring, enforcement, and ongoing testing.
The models are trained to refuse prohibited cyber assistance, including attempts to disguise intent or jailbreak the model. Real‑time cyber and biology misuse classifiers evaluate output as it is generated; for higher‑risk cases generation may be paused while a larger reasoning model reviews the conversation. If the output is assessed as disallowed, it can be withheld before reaching the user. Flagged activity may also trigger account‑level review across conversations and risk signals, which helps distinguish persistent malicious behavior from legitimate dual‑use security work.
OpenAI warns that during the preview users may encounter blocks or refusals, and some requests may take longer because generation is paused for review. The preview is intended to test whether safeguards limit misuse while still allowing legitimate users to complete normal work; partner feedback will be used to reduce unnecessary blocks and delays and to improve contextual interpretation.
Automated and human red‑teaming
To harden safeguards, OpenAI reports dedicating over 700,000 A100‑equivalent GPU hours to automated red‑teaming aimed at finding universal jailbreaks—attacks that can work across many prompts or contexts. This automated work is complemented by third‑party human expert red‑teaming, which will continue through the preview.
OpenAI acknowledges that no evaluation can cover every product configuration or multi‑step attack, so it maintains a rapid‑response process to reproduce, assess, prioritize and remediate new jailbreaks and then add them into ongoing evaluations.
Access and pricing
During the preview, GPT‑5.6 models will initially be available via the API and Codex to a select group of trusted partners and organizations. OpenAI plans to make them more broadly accessible soon through ChatGPT, Codex, and the API.
Under the new naming scheme, the numeric portion denotes generation while Sol, Terra and Luna denote durable capability tiers that can evolve independently. Pricing is quoted per 1M tokens: Sol is $5 input / $30 output; Terra is $2.50 input / $15 output; and Luna is $1 input / $6 output. GPT‑5.6 also introduces more predictable prompt caching, including support for explicit cache breakpoints and a 30‑minute minimum cache life. Cache writes are billed at 1.25× the uncached input rate, while cache reads retain a 90% cached‑input discount.
Summary
OpenAI presents the GPT‑5.6 family as offering clearer trade‑offs between intelligence, speed, and cost while pairing increased capabilities with strengthened safeguards and intensive red‑teaming. The company is using a phased preview, government coordination, and partner feedback to refine protections before wider release in the coming weeks.
