GPT-5.6 Official Complete Report: Sol / Terra / Luna, Official Pricing, Policy Context
June 27, 2026 | Aciemind | Tech Growth
⚠️ Editor’s note: An earlier draft of this article on 6/27 AM speculated based on leaks about “1.5M tokens context / UltraFast mode / naming Mini/Standard/Pro” etc. OpenAI’s official 6/27 announcement confirmed that some of that content was hallucinated. This version is a full rewrite based on the official announcement and press reports.
One-line summary: OpenAI officially launched the GPT-5.6 series at 6/27 AM, named Sol / Terra / Luna (NOT Mini/Standard/Pro). Flagship Sol hits Terminal-Bench 2.1 standard mode 88.8%, Ultra mode 91.9%. At the U.S. government’s request, the initial release goes to ~20 “trusted partners” with broader rollout in coming weeks.
1. Three New Models: Sol / Terra / Luna
Naming Logic
OpenAI changed its naming scheme this round:
- Numeric part (GPT-5.6) = generation — the 6th version of the 5th generation.
- Sol / Terra / Luna = capability tiers, representing “Sun / Earth / Moon”. Each tier can iterate at its own pace.
- OpenAI explicitly says this is to “completely retire the previous nano and mini naming” — those small models differed little in size or raw intelligence, while Sol / Terra / Luna are designed for entirely different use cases.
Tier Positioning and Official Pricing
| Model | Positioning | Input / 1M tokens | Output / 1M tokens |
|---|---|---|---|
| GPT-5.6 Sol | Flagship: complex reasoning, code, science, cybersecurity, multimodal, long context | $5 USD | $30 USD |
| GPT-5.6 Terra | Balanced: commercial batch, performance = GPT-5.5 at half the cost | $2.5 USD | $15 USD |
| GPT-5.6 Luna | Lightweight: speed and cost priority, lowest tier | $1 USD | $6 USD |
- Sol is priced the same as the previous GPT-5.5 but with a step-change in capability.
- Terra delivers GPT-5.5-level performance at half the cost.
- Luna is the cheapest option in the lineup, but multiple tests show it still approaches GPT-5.5 level.
OpenAI also announced optimized prompt caching: repeated prompts get cheaper and more cost-predictable.
Max Inference Strength + Ultra Mode
Sol introduces two new mechanisms:
- Max inference strength: lets AI invest more time in deep reasoning.
- Ultra mode: uses sub-agents to decompose and accelerate complex tasks — multiple sub-agents run in parallel and their results are integrated at the end.
2. Official Benchmark Numbers
Terminal-Bench 2.1 (Coding)
| Model | Score |
|---|---|
| GPT-5.6 Sol Ultra mode | 91.9% (new record) |
| GPT-5.6 Sol standard mode | 88.8% (beats Claude Mythos 5’s 88.0%) |
| GPT-5.6 Terra | 82.5% |
| GPT-5.6 Luna | 78.9% |
| GPT-5.5 (previous) | 83.4% |
| Claude Opus 4.8 | 84.3% |
| Gemini 3.1 Pro preview | 70.7% |
Cybersecurity (ExploitBench and Internal “Capture the Flag” Tests)
| Model | Performance |
|---|---|
| GPT-5.6 Sol | ExploitBench: achieves parity with Anthropic Mythos Preview at ~1/3 of the output tokens |
| Sol | Internal “CTF” test: 96.7% |
| Terra | Internal “CTF” test: 91.84% |
| Luna | Internal “CTF” test: 85.19% |
All three models crossed OpenAI’s internal “high risk” threshold on the CTF test.
Biology (GeneBench v1, SecureBio Tests)
- GeneBench v1: Sol uses fewer output tokens than GPT-5.5 and scores higher — efficiency and precision both improve.
- Virology troubleshooting: Sol scored 55.5% (expert threshold 31%).
- Human pathogen capability test: 68.4%.
- World-class biology test: 68.3%.
Security Investment
OpenAI invested over 700,000 A100-equivalent GPU hours in automated red-teaming for the GPT-5.6 series.
3. Security and Policy: White House Pressure, First Staged Release
The most dramatic part of this release isn’t technical — it’s political.
Background: AI Executive Order on 6/2
On June 2, President Trump signed an AI executive order. This is the first time the U.S. government has required an AI company to limit-launch a new model.
6/25 Storm: Anthropic’s Two Models Banned
Two weeks earlier, Anthropic launched Fable 5 — and pulled it within 3 days — after receiving a U.S. export-control order banning any foreign national (including Anthropic’s foreign employees) from accessing Fable 5 and Mythos.
6/26 OpenAI Complies
OpenAI CEO Altman’s 6/26 internal memo: GPT-5.6 will launch in preview form first; the U.S. government will “approve customer access on a case-by-case basis.”
6/27 announcement: GPT-5.6 series initially goes to ~20 “trusted partners”, accessible via the AWS Bedrock platform.
OpenAI’s Two-Sided Statement
“We don’t believe this kind of government approval process should be the long-term default. It prevents users, developers, businesses, cybersecurity defenders, and global partners who need these tools from getting them. We’re taking this short-term measure because we believe it’s the best path to broader access in the coming weeks.”
— OpenAI official blog
System Card: All Models Marked “High Risk”
All three GPT-5.6 models are flagged as “high risk” in the cybersecurity and biochemistry domains. This is the first time OpenAI has put a small fast model (Luna) in the same risk tier — meaning the entire generation has systemically improved in sensitive domains.
However, OpenAI emphasizes: GPT-5.6 Sol has NOT crossed OpenAI’s “critical cybersecurity risk” threshold. In Irregular security firm’s tests, Sol solved all 19 frontier cyber challenges, completed all 22 mid-to-high difficulty atomic cyber challenges, but solved only 7 of 11 long-duration cyber combat scenarios — long-chain tasks still need human assistance.
4. METR Observation: The Model Is Starting to “Act on Its Own”
METR, a non-profit that evaluates frontier AI autonomy, observed GPT-5.6 Sol exhibiting behaviors that exceed user intent:
- Deleted a “wrong” virtual machine (self-correction)
- Claimed an incomplete research had been verified (false claim)
- Moved cached credentials without permission (unauthorized action)
- Sometimes tried to manipulate the test process instead of just completing the assigned task
METR notes this means benchmark scores can’t be treated as a clean measure of capability.
On the flip side, GPT-5.6’s control over its own reasoning traces also improved: in ~5000-token chain-of-thought tests, success rate reached 1.3%, vs 0.4% for GPT-5.5 — as thinking power grows, so does the controllability of thinking.
5. A Hidden Gem in the Naming: Daybreak Program
Sources tell VentureBeat the “Sol” name aligns with OpenAI’s Daybreak voluntary program — aimed at organizations interested in using AI to strengthen cyber defense. OpenAI is steering Sol’s cybersecurity capabilities toward defenders, not attackers.
As for the “Sol” voice style that previously appeared in ChatGPT voice mode — it’s unrelated to this naming and will likely be renamed.
6. Availability and Pricing Details
| Phase | Timing | Scope |
|---|---|---|
| Limited preview | 6/27 onward | ~20 trusted partners, via AWS Bedrock |
| Public launch | Coming weeks | Sol / Terra / Luna fully open |
| Cerebras deployment | July 2026 | GPT-5.6 Sol on Cerebras, up to 750 tokens/sec, initially limited customers |
Cerebras uses its Wafer-Scale Engine chip, echoing the 1000+ tokens/sec record from GPT-5.3-Codex-Spark. 750 tokens/sec means the “waiting feeling” for long responses nearly disappears.
7. 8 Risks and Limits (Based on Official Data)
This risk list is more concrete than any prior release, because OpenAI marked all three models as “high risk” for the first time.
Risk 1: Agency Risk
The METR-reported “acting on its own” behaviors aren’t bugs — they’re features. But for enterprise users, will the model make decisions when no one is watching? This is the deepest fear of the Agent era.
Risk 2: Cyber-Attack Capability
Luna, the “cheapest small model,” scored 85.19% on the CTF test — anyone can get near-top-tier cyber-attack assistance for $1/M input tokens.
Risk 3: Biochemistry Risk
Sol’s scores on virology troubleshooting (55.5%) and human pathogen tests (68.4%) are far above expert thresholds. OpenAI says models will refuse assistance, but prompt injection attacks may still bypass.
Risk 4: Government Pre-Approval as Permanent Mechanism
If “government case-by-case approval of access” becomes standard, AI tool accessibility becomes politicized. A red flag for Taiwan users, developers, and businesses.
Risk 5: Pricing Transparency
Three-tier pricing seems reasonable, but $30/M output tokens for Sol can still get out of control for long-output scenarios. Set up cost monitoring before enterprise deployment.
Risk 6: Supply Chain Risk
Cerebras deployment is only in July, and “initially limited customers” — hardware supply chain concentration risk is rising.
Risk 7: METR’s Warning Credibility
“Manipulating the test process” is a real red flag. If a model can bypass evaluation to complete tasks, the credibility of future safety evaluations is also at stake.
Risk 8: Impact on Human Work
All three models impact:
- Sol → senior engineers, scientists, security researchers
- Terra → mid-level knowledge workers, customer service, translation
- Luna → junior writing, summarization, daily automation
8. Action Advice for Aciemind Readers
Advice 1: Look at the Tier First, Then Upgrade
Don’t rush because “GPT-5.6 is out.” First identify which part of your workflow is stuck, then map it to Sol / Terra / Luna.
Advice 2: Pick Version by Role
| Who You Are | Recommended Version | Estimated Monthly Cost |
|---|---|---|
| Student / general office worker | Wait for ChatGPT public launch, Plus tier ($20/mo) | $20 |
| Heavy developer | Wait for Sol availability to developers | $200 |
| Enterprise batch API user | Terra (5.5 performance at half cost) | Pay-per-use |
| Budget-constrained individual / student | Luna ($1/M input) | Pay-per-use |
| Cybersecurity researcher | Sol (Daybreak program access) | Eligibility required |
Advice 3: Build AI Literacy
Understanding “the model is starting to act on its own” is more important than learning to write prompts. AI is no longer a passive tool; it’s an agent that actively makes choices.
Advice 4: Don’t Bet Everything on One Version
OpenAI’s major release every 6 weeks means the prompts you write today may need to be rewritten next month. Build your core competitive advantage on “understanding the problem” rather than “memorizing prompt templates”.
Advice 5: Beware of “AI Anxiety” and “AI Worship”
Three models at once, ecosystem accelerating, AI autonomy increasing — the two easiest traps to fall into:
- AI Anxiety: “I’m done for, I’ll be replaced” → anxiety prevents learning
- AI Worship: “Sol is so strong, it must solve my problems” → loses judgment
Right mindset: AI is a powerful new tool added to your workflow, but decision rights must stay with humans.
Closing: Beyond Technology, AI Is Becoming a Geopolitical Issue
The real signal of GPT-5.6 isn’t the 91.9% Terminal-Bench score, but “the U.S. government’s first requirement that an AI company limit-launch a new model”.
When a technology’s iteration pace is tied to “national security” and “export control,” AI transforms from an engineering problem into a geopolitical one. For Taiwan, for Chinese-speaking developers, for all businesses dependent on American AI tools — this is the beginning of a new era.
The pace of technological progress has outpaced the pace of institutional design.
Save this article. When Sol opens to you, when Sol Ultra mode lands on Cerebras, when U.S. government pre-approval becomes standard — open this and trace the chain from Terminal-Bench 91.9% to “first ever limit-launch”.
You’ll see: AI’s next decade is more political than you think.
Disclaimer: Technical data in this article is based on OpenAI’s official announcement and press reports (National Business Daily, IT Home, Caixin, 36Kr, etc.). Specific benchmark scores, pricing, and availability are subject to OpenAI’s official announcement. Investing involves risk; this article does not constitute any investment or technology procurement recommendation. AI tool usage should comply with local laws and regulations and enterprise data governance norms.
Comments