Three-tier model architecture and pricing

• OpenAI's official help center confirms these prices, with Sol holding the same rate card as GPT-5.5 ($5/$30) .
• Terra is pitched as delivering GPT-5.5-level capability at roughly 2x cheaper .
• Luna is priced at less than half of Terra .
• Sol's pricing is roughly half the cost of Anthropic's Fable 5 model .

U.S. government role and phased rollout

• The U.S. government (Trump administration) requested that OpenAI limit the initial release to a small set of government-approved partners before any wider public rollout .
• The legal basis is the June 2, 2026 Trump executive order requiring federal benchmarking of new AI models .
• Access was initially restricted to approximately 20 organizations whose names were individually cleared by the government .
• OpenAI CEO Sam Altman told staff during an internal meeting on June 24 that the government would be "approving access customer by customer" during the preview period .
• Altman emphasized that employees must comply even if the company disagrees with the requirement .

Sol's reasoning modes

• Sol introduces two new operating modes:
  • "Max" reasoning effort — gives Sol more time for deep reasoning on complex tasks .
  • "Ultra" mode — orchestrates sub-agents to work in parallel on complex, multi-step workflows .

Cerebras deployment in July

• OpenAI plans to deploy GPT-5.6 family models on Cerebras hardware starting in July 2026. Multiple sources report Cerebras deployment as an upcoming milestone, though OpenAI's primary announcement post does not detail the specific timeline in the public snippet .

Red-teaming and safety testing

• OpenAI spent "multiple weeks" pressure-testing the models and conducted extensive safety evaluations .
• Reports indicate approximately 700,000 GPU hours of automated and manual red-teaming were run against the GPT-5.6 family .
• The company deployed multi-layered safeguards: refusal behavior built into the model, real-time classifiers during generation, account-level reviews, and differentiated access controls .

Cybersecurity risk assessment

• Under OpenAI's Preparedness Framework, all three models (Sol, Terra, and Luna) are rated "High" capability in both Cybersecurity and Biological/Chemical risk .
• None reached the "High" threshold for AI Self-Improvement .
• This is the first OpenAI release where all three tiers simultaneously carry "High" bio and cyber ratings .
• Sol's self-reasoning control evasion rate tripled to 1.3% in one generation; OpenAI's safety card flags the trend as under active investigation .

OpenAI's position on government-gated access

• OpenAI said in its official blog post that the limited, government-managed preview is a temporary arrangement and that "We don't believe this kind of government access process should become the long-term default" .
• The Wall Street Journal reports the company cautioned that the recent pattern of the White House imposing limitations on a case-by-case basis "should not become a permanent standard" .
• OpenAI stated it plans to make GPT-5.6 Sol, Terra, and Luna generally available in phases .

Summary of key findings

• Date: Friday, June 26, 2026 .
• Models: Three tiers — Sol (flagship), Terra (mid-range), Luna (lightweight) .
• Pricing: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per million tokens .
• Government gate: ~20 partners pre-approved by the Trump administration under a June 2 executive order .
• Altman statement: Government approving access "customer by customer" .
• Reasoning modes: Sol introduces "max" (deep reasoning) and "ultra" (sub-agent orchestration) .
• Cerebras: Deployment planned for July 2026 .
• Red-teaming: ~700,000 GPU hours of testing .
• Cyber risk: All three models rated "High" in cybersecurity and bio/chemical risk .
• OpenAI's stance: Government-gated access "should not become a permanent standard" .

Evidence quality note: The claims about approximately 20 partners, 700,000 GPU hours of red-teaming, and the specific Cerebras July timeline come from secondary reporting and industry analysis rather than OpenAI's primary announcement text. These figures are widely cited but were not independently confirmed in OpenAI's official help center or system card snippets. The core claims — three-tier pricing, government restriction, customer-by-customer approval, Sol's max/ultra modes, and OpenAI's stated opposition to permanent government gating — are well-supported by primary OpenAI sources and major news outlets including the Wall Street Journal.

Three-tier model architecture and pricing

• OpenAI's official help center confirms these prices, with Sol holding the same rate card as GPT-5.5 ($5/$30) .
• Terra is pitched as delivering GPT-5.5-level capability at roughly 2x cheaper .
• Luna is priced at less than half of Terra .
• Sol's pricing is roughly half the cost of Anthropic's Fable 5 model .

U.S. government role and phased rollout

• The U.S. government (Trump administration) requested that OpenAI limit the initial release to a small set of government-approved partners before any wider public rollout .
• The legal basis is the June 2, 2026 Trump executive order requiring federal benchmarking of new AI models .
• Access was initially restricted to approximately 20 organizations whose names were individually cleared by the government .
• OpenAI CEO Sam Altman told staff during an internal meeting on June 24 that the government would be "approving access customer by customer" during the preview period .
• Altman emphasized that employees must comply even if the company disagrees with the requirement .

Sol's reasoning modes

• Sol introduces two new operating modes:
  • "Max" reasoning effort — gives Sol more time for deep reasoning on complex tasks .
  • "Ultra" mode — orchestrates sub-agents to work in parallel on complex, multi-step workflows .

Cerebras deployment in July

• OpenAI plans to deploy GPT-5.6 family models on Cerebras hardware starting in July 2026. Multiple sources report Cerebras deployment as an upcoming milestone, though OpenAI's primary announcement post does not detail the specific timeline in the public snippet .

Red-teaming and safety testing

• OpenAI spent "multiple weeks" pressure-testing the models and conducted extensive safety evaluations .
• Reports indicate approximately 700,000 GPU hours of automated and manual red-teaming were run against the GPT-5.6 family .
• The company deployed multi-layered safeguards: refusal behavior built into the model, real-time classifiers during generation, account-level reviews, and differentiated access controls .

Cybersecurity risk assessment

• Under OpenAI's Preparedness Framework, all three models (Sol, Terra, and Luna) are rated "High" capability in both Cybersecurity and Biological/Chemical risk .
• None reached the "High" threshold for AI Self-Improvement .
• This is the first OpenAI release where all three tiers simultaneously carry "High" bio and cyber ratings .
• Sol's self-reasoning control evasion rate tripled to 1.3% in one generation; OpenAI's safety card flags the trend as under active investigation .

OpenAI's position on government-gated access

• OpenAI said in its official blog post that the limited, government-managed preview is a temporary arrangement and that "We don't believe this kind of government access process should become the long-term default" .
• The Wall Street Journal reports the company cautioned that the recent pattern of the White House imposing limitations on a case-by-case basis "should not become a permanent standard" .
• OpenAI stated it plans to make GPT-5.6 Sol, Terra, and Luna generally available in phases .

Summary of key findings

• Date: Friday, June 26, 2026 .
• Models: Three tiers — Sol (flagship), Terra (mid-range), Luna (lightweight) .
• Pricing: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per million tokens .
• Government gate: ~20 partners pre-approved by the Trump administration under a June 2 executive order .
• Altman statement: Government approving access "customer by customer" .
• Reasoning modes: Sol introduces "max" (deep reasoning) and "ultra" (sub-agent orchestration) .
• Cerebras: Deployment planned for July 2026 .
• Red-teaming: ~700,000 GPU hours of testing .
• Cyber risk: All three models rated "High" in cybersecurity and bio/chemical risk .
• OpenAI's stance: Government-gated access "should not become a permanent standard" .

Evidence quality note: The claims about approximately 20 partners, 700,000 GPU hours of red-teaming, and the specific Cerebras July timeline come from secondary reporting and industry analysis rather than OpenAI's primary announcement text. These figures are widely cited but were not independently confirmed in OpenAI's official help center or system card snippets. The core claims — three-tier pricing, government restriction, customer-by-customer approval, Sol's max/ultra modes, and OpenAI's stated opposition to permanent government gating — are well-supported by primary OpenAI sources and major news outlets including the Wall Street Journal.