OpenAI unveiled its most capable AI model family yet on June 26, 2026, launching a limited preview of GPT-5.6 across three tiers: Sol, Terra, and Luna. The release is notable not just for the capability jump but for the regulatory wrapper around it: only approximately 20 government-approved partners have access, with wider availability promised “in the coming weeks.”

The rollout reflects the new reality of frontier AI governance following President Trump’s June 2 executive order on AI innovation and security, which requires federal benchmarking of new models before broad release.

Three Models, Three Jobs

The GPT-5.6 family is structured as a deliberate ladder:

Sol is the flagship. OpenAI describes it as its strongest model to date, with meaningful gains in agentic capabilities across coding, biology, and cybersecurity. Sol introduces two new reasoning modes: a max reasoning effort option that gives the model more time to think through complex problems, and an ultra mode that spins up multiple subagents to work on hard tasks in parallel. For businesses running AI-powered workflows, this is the power tier.

Terra sits in the middle. It is designed for everyday work where you want more capability than a lightweight model but do not need Sol’s firepower or price tag. Think day-to-day content generation, summarisation, research, and customer-facing applications.

Luna is the speed and cost tier. Fast, affordable, and aimed at high-volume applications where latency and token economics matter more than maximum reasoning depth.

Pricing and Access

Sol is priced at $5 per million input tokens and $30 per million output tokens, with a 90% discount on cached input reads and cache writes billed at 1.25x the standard rate. Pricing for Terra and Luna was not disclosed in the initial preview announcement.

The limited access model is an unusual feature. Rather than the typical public API launch, access to GPT-5.6 is being handed out to roughly 20 partner organisations that were individually approved by the US government. This reflects both the model’s safety classification (Sol carries a “High” risk designation for cybersecurity and biological and chemical capabilities) and the broader shift toward government oversight of frontier models.

General availability is expected within weeks.

The Safety Picture

GPT-5.6 Sol launches with what OpenAI calls its “most robust safety stack to date,” with strengthened guardrails for higher-risk cybersecurity requests and biological capability queries. The high risk classification is not a limitation so much as an indicator of what the model can do: Sol is capable enough that OpenAI and the US government both wanted extra caution on the rollout.

Separately, GPT-4.5 was retired on June 27, 2026, continuing the platform’s steady transition toward the GPT-5.x generation.

What This Means for Business

The capability ceiling keeps rising. The introduction of ultra mode, where Sol orchestrates multiple subagents to tackle complex multi-step tasks, marks a shift from single-model AI assistants to coordinated AI workforces. Businesses building automation today should think about whether their architecture can take advantage of multi-agent reasoning, not just single-model calls.

Government approval as a new variable. The GPT-5.6 launch is the first high-profile example of a major AI model release being gated by government review rather than just technical readiness. This is likely to become a recurring feature of frontier model deployments. Businesses that rely heavily on cutting-edge AI capability should watch how this review process evolves.

The cost curve is shifting at the top end. Sol’s $30 per million output token pricing is significantly higher than GPT-5.5’s rates. That gap between frontier capability and affordable capability is widening. Most business applications do not need Sol, but the existence of a model this capable raises the bar on what “good enough” means for customers and competitors.

Model selection gets more complex. Three tiers within a single model family means businesses will need to think carefully about which tier fits which job. Using Sol for simple FAQ handling is expensive overkill. Using Luna for strategic analysis is a capability mismatch. Getting this right is an emerging skill in AI-native organisations.

EDNA’s work sits squarely in this space: helping business leaders and data professionals understand not just what AI can do, but which AI to use, at what cost, for which task. As model families grow more layered, that expertise matters more, not less.