OpenAI just officially launched GPT-5.6 — not one model but three: Sol (flagship), Terra (balanced), and Luna (fast). New “max” and “ultra” reasoning modes. Government-approved customer list. $5/$30 per million tokens. The model race just restarted.
GPT-5.6 — Three Models
New Capabilities
Max
New reasoning level — deepest thinking
Ultra
Beyond single agent — uses subagents
~20
Companies in limited preview (gov-approved)
Safety
“Most robust safety stack to date”
What OpenAI Launched
Sol is OpenAI’s strongest model — with agentic improvements in coding, biology, and cybersecurity. The new “max” reasoning mode gives Sol the most time to think deeply. The new “ultra” mode goes beyond a single agent by leveraging subagents to accelerate complex work. This is OpenAI building loops into the model itself — the same paradigm Boris Cherny called “as big as the step from code to agents.”
Terra matches GPT-5.5 performance at half the cost. Luna is OpenAI’s cheapest model ever. The three-tier structure mirrors what Anthropic has with Opus/Sonnet/Haiku — but OpenAI is pricing aggressively to undercut.
The catch: limited preview to ~20 government-approved companies only. General availability “in the coming weeks.” The government stagger we reported yesterday is now official policy.
The key insight: “Ultra” mode using subagents is OpenAI embedding loops at the model level. The model doesn’t just think harder — it spawns other agents to work in parallel. This is the architecture shift that makes GPT-5.6 structurally different from GPT-5.5, not just incrementally better.
The Competitive Map — Updated
Frontier Models — June 26, 2026
OpenAI GPT-5.6 Sol/Terra/Luna
Just launched. Limited preview (~20 cos). $5/$30. Ultra mode with subagents. Gov-gated.
Anthropic Claude Opus 4.8 / Fable 5
Polymarket 94.8%. Fable 5 behind paywall. Glasswing running. ID verification July 8.
Google Gemini 3.5 Pro
2M context. Deep Think. Still in limited preview. Pichai promised June — 4 days left.
DeepSeek V4-Pro / GLM-5.2
1.6T params on Huawei. GLM beats GPT-5.5 at coding for 1/6th cost. Open weights.
The Structural Read
ULTRA MODE = LOOPS AT THE MODEL LEVEL
Sol’s “ultra” mode spawns subagents. That’s the loops paradigm — agents prompting agents — built into the model itself. OpenAI didn’t just ship a better model. They shipped a model that orchestrates other models. The harness is now inside the model, not just around it.
THREE TIERS = THE INFERENCE ECONOMICS PLAY
Sol for frontier work. Terra for everyday (at half GPT-5.5 cost). Luna for high-volume cheap inference. Combined with Jalapeño chips cutting inference cost 50%, OpenAI is attacking the cost curve from both sides — cheaper models AND cheaper hardware. The $100B ad business needs cheap inference to serve free-tier users profitably.
GOVERNMENT GATING IS NOW THE DEFAULT
~20 companies approved by the government. General availability “coming weeks.” Yesterday’s report about the government asking for staggered release is now confirmed in the official launch. The Permission Layer is baked into the product launch itself — not an afterthought.
The Bottom Line
GPT-5.6 is official. Three models: Sol (flagship with subagent “ultra” mode), Terra (GPT-5.5 quality at half cost), Luna (cheapest ever). Government-approved limited preview. General availability coming. The question is no longer whether OpenAI can compete with Anthropic at 94.8% Polymarket odds — it’s whether GPT-5.6 Sol changes those odds before June 30. Five days to find out.
Sources: OpenAI, MacRumors, VentureBeat — June 26, 2026