GPT-5.6 Sol, Terra, Luna: the pricing, the benchmarks, and what changes for agents
OpenAI previewed a three-tier model family: Sol (flagship), Terra (GPT-5.5-class at half the cost), Luna (fast/cheap). Ultra mode spawns subagents. US government asked for limited access.
TL;DR: OpenAI previewed GPT-5.6 on June 26 as a three-tier family: Sol (flagship), Terra (balanced), and Luna (fast/cheap). The headline story isnt the benchmarks. Its that OpenAI shipped a new reasoning mode that spawns subagents, and the US government asked for limited access under a new executive order. Terra matches GPT-5.5 at half the cost. Sol introduces ultra mode with subagent orchestration. But most developers cant access any of it yet.
Key takeaways:
- GPT-5.6 ships as three tiers: Sol ($5/$30 per M tokens), Terra ($2.50/$15), Luna ($1/$6). Terra matches GPT-5.5-class performance at 2x cheaper.
- Ultra mode is the real news. Sol can now spawn subagents for parallel work. One gathers sources, another writes code, another checks math. Terminal Bench 2.1: 88.8% base, 91.9% in ultra. OpenAI GPT-5.6 preview
- The US government asked OpenAI to limit access to approved partners only. This is the first model released under a June 2 executive order establishing a 30-day voluntary review window.
- Sol scores higher than GPT-5.5 on internal benchmarks and its also the most capable cyber model OpenAI has shipped. The system card flags it as staying just under the cyber-critical line.
- For agent builders, the pricing structure creates a clear workflow: Luna for routing and classification, Terra for everyday agent tasks, Sol for the hard reasoning. Ultra mode handles the orchestration.
What is the GPT-5.6 family?
Three models. Three price points. One generation.
| Model | Input per 1M tokens | Output per 1M tokens | Positioning |
|---|---|---|---|
| Sol | $5.00 | $30.00 | Flagship. Frontier reasoning, long-horizon agentic work |
| Terra | $2.50 | $15.00 | Balanced. GPT-5.5-class performance at 2x cheaper |
| Luna | $1.00 | $6.00 | Fast, token-efficient. Strong capability at lowest cost |
Sol holds the same rate card as GPT-5.5. The difference is the capability ceiling. Terra is where the value proposition shifts. OpenAI is offering GPT-5.5-level reasoning at half the inference cost. Luna exists for the high-volume, low-latency work that doesnt need a frontier model.
The naming convention is cleaner than the old Mini/Turbo/Pro system. Sol is the flagship. Terra is the workhorse. Luna is the cheap one. You no longer have to remember whether Turbo or Pro was the smarter model.
Ultra mode changes how you think about the model
The biggest new feature isnt a benchmark score. Its ultra mode.
Standard reasoning (max mode) gives Sol more time to think, similar to Claudes extended thinking. Ultra mode goes further. The system breaks your problem into pieces, hands each piece to a subagent, and stitches the results back together.
One subagent gathers sources. One checks the math. One writes code. One hunts for mistakes. All running in parallel behind a single conversation.
| Setting | Terminal Bench 2.1 | What it does |
|---|---|---|
| Base | 88.0% (GPT-5.5) | Standard generation |
| Max | 88.8% | Extended reasoning, more thinking time |
| Ultra | 91.9% | Subagent orchestration, parallel work |
The interesting number isnt that ultra mode scores 91.9%. Its that the base model barely moved compared to GPT-5.5. The gains came from orchestration, not from a smarter single model. This is the pattern for the next generation of AI products. Not one giant model. Multiple models coordinating.
The government gate is new
The Trump administration asked OpenAI to limit GPT-5.6 to a small set of government-approved partners. This is the first model released under a June 2 executive order that established a voluntary 30-day review window for powerful AI models.
OpenAI agreed. Sam Altman told staff the government would be approving participants. The company also said this restriction shouldn’t become the norm.
The legal basis is narrow: a voluntary framework, not a law. But the precedent is real. For the first time, a frontier model is being held back from broad release at government request. What happens when this framework applies to every major model release?
What the numbers say
OpenAI published a comprehensive safety system card for GPT-5.6. The key claims:
- Sol is OpenAIs most capable cyber model per the GPT-5.6 system card. On Chromium and Firefox tests, it didn’t autonomously build a full working exploit chain, keeping it under the cyber-critical threshold.
- Terra and Luna both exceed GPT-5.5 by a substantial margin on internal evaluations, despite their lower cost.
- The system card covers prompt injection robustness, health benchmarks (HealthBench), hallucination rates, bias evaluations, and biological/chemical capability assessments.
- OpenAI introduced real-time output classifiers, cross-conversation monitoring, and a bigger reasoning model that can pause and approve or deny flagged responses.
The cyber capability section is worth reading. The model can identify vulnerabilities and run long-horizon security work faster than anything OpenAI has shipped. The safeguards are correspondingly heavier. Flagged activity now triggers account-level review across multiple conversations, not just single-message refusals.
What this means for agent builders
The three-tier pricing maps cleanly to agent workflows:
- Luna ($1/$6) for routing, classification, input preprocessing. High volume, low latency, cheap enough to run on every user interaction.
- Terra ($2.50/$15) for everyday agent tasks. Code generation, tool calls, RAG responses. The sweet spot for most production agent loops.
- Sol ($5/$30) for the hard reasoning. Planning, multi-step tool orchestration, edge cases that Terra cant handle. Use it as the fallback, not the default.
Ultra mode adds another layer. For agent builders, the subagent orchestration pattern OpenAI productized is the same pattern the agent framework community has been building manually for the past year. The difference is that OpenAI ships it as a model-level toggle, not a framework configuration.
The pricing also highlights a trend. Inference costs are compressing fast at the mid tier. Luna at $1/$6 per million tokens makes frontier-quality classification and preprocessing cheaper than running a small fine-tuned model. The economics of the three-tier structure favors routing: send the easy work to the cheap model, escalate only when needed.
The tradeoff
GPT-5.6 is a limited preview. Most developers can’t use it today. The government review window lasts up to 30 days, and OpenAI hasn’t committed to a broader release timeline.
If you are building agent products today, your options are:
- Wait for general availability. Risk: no timeline, and future releases may face the same gate.
- Build on Terra today (if you have preview access). Risk: Terra pricing is competitive, but the best features (ultra mode, subagent orchestration) require Sol.
- Continue with alternatives. Claude Opus 4.7 costs $5/$25 and is broadly available. GLM 5.2 costs $1.40/$4.40 and outperforms GPT-5.5 on several benchmarks. Open-weight models like DeepSeek V4 and Qwen3.5-9B run on your own hardware.
The government gate creates an accidental moat for open-weight models. If frontier closed models face release delays and access restrictions, the practical advantage of running a capable open model on your own hardware grows. Not because open models are better. Because they dont need government approval to deploy.
FAQ
What is GPT-5.6 Sol? Sol is OpenAIs new flagship model. It costs $5 per million input tokens and $30 per million output tokens. It comes with new reasoning modes including ultra mode that spawns subagents for complex work. Initial access is limited to government-approved partners.
How does Terra compare to GPT-5.5? Terra offers GPT-5.5-class performance at about half the cost: $2.50 input and $15 output per million tokens. OpenAI says Terra and Luna both exceed GPT-5.5 by a substantial margin according to the system card evaluations.
What is ultra mode in GPT-5.6? Ultra mode is a new reasoning setting that goes beyond what a single model can do by spawning subagents to parallelize complex work. One subagent gathers sources, another checks math, another writes code, another hunts for mistakes. Sol in ultra mode scored 91.9% on Terminal Bench 2.1.
Why is GPT-5.6 limited preview only? The Trump administration asked OpenAI to limit the release under a June 2 executive order establishing a voluntary 30-day government review window for powerful AI models. OpenAI agreed to start with a small set of approved partners. Sam Altman said this restriction shouldn’t be the norm.
How does GPT-5.6 compare to Claude Opus 4.8 and GLM 5.2? Direct comparisons are limited because GPT-5.6 is still in preview. Sol matches GPT-5.5 pricing ($5/$30) with higher benchmark scores. Terra undercuts GPT-5.5 on price while matching or exceeding its performance. GLM 5.2 costs $1.40/$4.40 per million tokens, making it 2-2.8x cheaper than any GPT-5.6 tier.
Related Posts
- The 2026 LLM architecture landscape: 5 design families. How Sol, Terra, and Luna fit into the broader model design landscape.
- Best AI coding agents 2026. How GPT-5.6 models compare against Claude, Gemini, and open-weight coding agents.
- GLM 5.2: the MIT open-source model that rivals Opus 4.8. The open-source alternative that undercuts even Luna on pricing.
- When a better model isnt a better agent. Why model capability and agent performance aren’t the same curve.
This article was published on Agentic Up (https://agenticup.dev): practical guides for developers building AI-native products. Reach me at [email protected].
