OpenAI will release its GPT-5.6 family of models on Thursday, splitting its flagship offering into three capability tiers, Sol, Terra, and Luna, in a structural shift from the company’s previous single-model approach.
CEO Sam Altman confirmed the July 10 release date in a post on X on Wednesday. The models were first previewed on June 26, and OpenAI has been slow-rolling access to a small group of trusted partners and U.S. government reviewers.
The three tiers
Sol is the flagship model, designed for deep reasoning, agentic tasks, and complex multi-step problem solving. OpenAI says Sol makes fewer factual errors, hallucinations, compared to any previous OpenAI model. It supports a new “ultra” reasoning mode that can spawn subagents for complex workflows.
Terra is the mid-tier option, positioned as competitive with GPT-5.5 at roughly half the price. OpenAI pitches it as the Goldilocks choice for production workloads that need strong quality without the cost of full frontier reasoning.
Luna is the fastest and cheapest tier, built for high-volume tasks where speed matters more than peak reasoning ability.
Pricing per million tokens is $5 input / $30 output for Sol, $2.50 / $15 for Terra, and $1 / $6 for Luna. All three tiers support a 90% discount on cached input reads, and OpenAI confirmed a deployment on Cerebras hardware in July that pushes Sol to 750 tokens per second for select customers.
Safety and government review
The release comes as the U.S. government’s stance on AI oversight shifts. President Trump’s June 2 executive order created a voluntary system for frontier AI developers to share access to new models with the Defense Department, giving officials 30 days to raise concerns before public release.
OpenAI said it voluntarily shared early access with the government and “trusted partners” but noted it does not believe government review should become the long-term default. A White House official told CNET that the government did not give OpenAI a green light to release GPT-5.6, nor was one required under the executive order.
The models include a five-layer safety stack: model-level refusal training, real-time misuse classifiers that can pause generation, account-level review, differentiated access for sensitive capabilities, and monitoring with rapid-response jailbreak remediation. OpenAI dedicated over 700,000 A100-equivalent GPU hours to automated red-teaming.
Cyber capabilities
OpenAI’s safety evaluations show that Sol is competitive with Anthropic’s Mythos Preview on the ExploitBench cybersecurity benchmark while using roughly one-third of the output tokens. However, the company’s Preparedness Framework assessment found that Sol “does not cross the Cyber Critical threshold”, it can identify bugs and exploitation primitives but cannot autonomously produce a full-chain exploit.
Sol also set a new state of the art on Terminal-Bench 2.1 for command-line coding workflows and outperformed GPT-5.5 on GeneBench v1 for biology tasks while using fewer tokens.
Naming convention
The numbering (5.6) identifies the generation, while Sol, Terra, and Luna are durable capability tiers that can each advance at their own cadence. OpenAI plans broader availability via ChatGPT, Codex, and the API in the coming weeks.
Sources: CNET (July 8); OpenAI announcement (June 26)