On June 26, 2026, OpenAI officially released the GPT-5.6 family — flagship Sol, balanced Terra, and lightweight Luna — introducing a solar-system naming scheme for the first time. For AI developers and engineering leads, this article covers everything you need: pricing at a glance, Max/Ultra reasoning modes, TerminalBench 2.1 world #1 (91.9%), CTF hit rate 96.7%, Cerebras 750 token/s acceleration, the U.S. government's first limited-release review, a vs Claude Mythos 5 comparison table, a six-step access guide, and a full FAQ. Only about 20 vetted partner organizations can preview the models today; broad availability is expected within weeks.
June was supposed to be AI's "super launch month," but all three top labs had their flagship releases blocked at the door. Developers currently face three core pain points:
Restricted access: At U.S. government request, GPT-5.6 is limited to about 20 vetted partner organizations — ordinary users cannot access it via ChatGPT or the public API
Competitors forced offline: Claude Mythos 5 was shut down June 12 under export controls; Gemini 3.5 Pro slipped to July — leaving a vacuum in the coding agent market
Policy uncertainty: President Trump's June 2 executive order set a precedent for government intervention in AI releases, making future model timelines harder to predict
| Model | Tier | Input Price | Output Price | Highlight |
|---|---|---|---|---|
| GPT-5.6 Sol | Flagship / strongest | $5 / 1M tokens | $30 / 1M tokens | TerminalBench 2.1 world #1 (91.9%) |
| GPT-5.6 Terra | Balanced / workhorse | $2.50 / 1M tokens | $15 / 1M tokens | Near GPT-5.5 performance, 50% lower cost |
| GPT-5.6 Luna | Lightweight / fast | $1 / 1M tokens | $6 / 1M tokens | Best for high-frequency tasks, 80% price advantage |
Current status: At U.S. government request, access is limited to about 20 vetted partner organizations. Broad release is expected within weeks. Polymarket assigns roughly an 87% probability to full release by July 31.
In the early hours of June 27, 2026 (Beijing time), OpenAI officially released the GPT-5.6 series, introducing a solar-system naming scheme for the first time — Sol (Sun), Terra (Earth), Luna (Moon) — mapping to flagship, balanced, and lightweight tiers respectively.
The launch was far from smooth. Following President Trump's June 2 executive order, OpenAI was required to undergo a government security review before broad release — the first time the U.S. government has mandated a limited release of a frontier AI model. CEO Sam Altman complied but publicly stated:
We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.
Sol is OpenAI's most capable model to date, built for the hardest tasks: advanced programming, long-horizon cybersecurity research, and multi-step autonomous agentic workflows.
Two new reasoning modes:
Pricing: $5 / 1M input tokens, $30 / 1M output tokens (same as GPT-5.5)
Terra is the core workhorse for everyday enterprise workloads — high-volume customer support, internal tools, document analysis, and other frequent business scenarios. Performance is close to GPT-5.5 but costs 50% less, making it the best value for large-scale deployments. Pricing: $2.50 / 1M input, $15 / 1M output.
Luna is optimized for high-frequency, low-latency tasks — summarization, drafting, and routine automation. Notably, Luna is also OpenAI's first non-flagship model to earn a High capability rating in both cybersecurity and biology. Pricing: $1 / 1M input, $6 / 1M output.
| Model | Best For | Context Window | Cybersecurity Rating |
|---|---|---|---|
| Sol | Complex coding, security research, long-horizon agents | ~1.5M tokens | High |
| Terra | Enterprise document analysis, support, high-volume API | ~1.5M tokens | High |
| Luna | Summarization, drafting, routine automation | ~1.5M tokens | High |
TerminalBench 2.1 is one of the most authoritative code-agent benchmarks, with 89 complex command-line planning problems testing multi-step tool use, iterative repair, and task coordination in realistic conditions.
| Model | Score | Mode |
|---|---|---|
| GPT-5.6 Sol | 91.9% — New #1 globally | Ultra (multi-agent) |
| GPT-5.6 Sol | 88.8% | Standard |
| Claude Mythos 5 | 88.0% | Standard |
| GPT-5.5 | 83.4% | Standard |
| Gemini 3.1 Pro Preview | 70.7% | Standard |
Sol dethroned Claude Mythos 5 after just 17 days — Mythos 5 had claimed the top spot on June 9. See our earlier pre-release leak roundup for background.
| Model | Task Completion Rate (Code Mode) |
|---|---|
| GPT-5.6 Sol | 50.9% — only model to cross 50% |
| GPT-5.6 Luna | Slightly above GPT-5.5 |
GPT-5.6 is the first OpenAI product line where all three models triggered a "High" cybersecurity risk classification.
| Model | CTF Hit Rate |
|---|---|
| Sol | 96.7% |
| Terra | 91.84% |
| Luna | 85.19% |
ExploitBench: Sol's performance is nearly identical to Anthropic's Mythos Preview, but uses roughly one-third the output tokens — dramatically lowering the cost of enterprise security research.
Safety note: OpenAI testing shows Sol can identify vulnerabilities and exploit primitives in Chromium and Firefox codebases, but cannot autonomously construct complete, functional exploit chains — keeping it below OpenAI's "Cyber Critical" threshold.
Starting in July, GPT-5.6 Sol will deploy on the Cerebras hardware acceleration platform for select customers, reaching up to 750 tokens per second. For context: most flagship models today output at 50–150 token/s. At 750 token/s, response times could shrink to one-fifth to one-fifteenth of current models — a step change for real-time coding assistants and streaming AI applications.
President Trump signed an executive order allowing U.S. government agencies up to 30 days of pre-release access to review frontier AI models for national security. The order is not legally mandatory, but it had real constraining effect. On June 26, coordinated by the White House Office of Science and Technology Policy (OSTP) and the Office of the National Cyber Director (ONCD), OpenAI agreed to limit GPT-5.6's launch to approximately 20 pre-approved "trusted partner" organizations.
| Company | Model | Status |
|---|---|---|
| OpenAI | GPT-5.6 Sol/Terra/Luna | Limited preview for ~20 partner orgs |
| Anthropic | Claude Fable 5 / Mythos 5 | Forced offline June 12 under export controls |
| Gemini 3.5 Pro | Delayed to July; originally planned for June |
| Dimension | GPT-5.6 Sol | Claude Mythos 5 |
|---|---|---|
| TerminalBench 2.1 | 91.9% (Ultra) / 88.8% | 88.0% |
| ExploitBench | Near-identical to Mythos Preview, 1/3 the tokens | Data not publicly released |
| Input pricing | $5 / M | Originally $10/M (currently offline) |
| Availability | Limited preview; broad release within weeks | Offline due to export controls |
| Context window | ~1.5M tokens | 200K tokens |
Sol leads Mythos 5 on coding and cybersecurity benchmarks at half the input price with comparable security research capability. Fable 5 still holds advantages on dimensions like SWE-bench Pro; full GPT-5.6 System Card data is pending for a complete comparison. Background: Claude Fable 5 export control analysis.
Monitor OpenAI's official status page: Set alerts for GPT-5.6 general availability so you don't miss the API launch window
Audit your current model stack: Until GPT-5.6 is broadly available, keep GPT-5.5 or Claude Opus 4.8 as your production baseline
Pre-select models by scenario: Reserve Sol for complex agent tasks; Terra for high-volume business API; Luna for lightweight high-frequency workloads
Priority-test after API opens: TerminalBench-style multi-step coding, CTF security research, and long-context document analysis
Compare token costs: Ultra mode delivers peak performance but consumes significantly more tokens — enable only for genuinely complex tasks
Plan Cerebras acceleration integration: After July, evaluate 750 token/s ROI for real-time coding assistants and contact OpenAI enterprise channels
| Your Need | Recommended Model |
|---|---|
| Complex code generation, debugging, multi-step agent tasks | Sol |
| Enterprise document analysis, support, high-volume API calls | Terra |
| High-frequency summarization, drafting, routine automation | Luna |
| Tight budget but need GPT-5.5-level capability | Terra (same-tier performance, 50% lower cost) |
| Latency-critical real-time apps (after July) | Sol on Cerebras |
Pure cloud APIs make model switching easy, but they carry risks: policy shocks, soaring long-context costs, and unpredictable Ultra-mode token consumption. Full self-hosting demands A100/H100-class GPUs and ongoing ops overhead. For production environments that need stable 24/7 AI agents, multi-agent coding pipelines, or iOS CI/CD automation, NodeMini's Mac Mini M4 cloud rental offers unified memory architecture and Apple Silicon efficiency — a better balance of performance, compliance isolation, and operational cost. See rental pricing for details.
Not yet for the general public. Currently limited to about 20 government-vetted trusted partner organizations via API and Codex. Full ChatGPT rollout is expected in July 2026; Polymarket assigns roughly an 87% probability to broad release by July 31.
Sol leads on TerminalBench 2.1 at 91.9% (Ultra) vs Mythos 5's 88.0%. ExploitBench performance is comparable but Sol uses roughly one-third the tokens. Mythos 5 still leads on SWE-bench Pro in some dimensions — wait for the full System Card before drawing final conclusions.
Ultra mode uses a multi-agent architecture: Sol breaks complex tasks into subtasks, distributes them to parallel sub-agents, and synthesizes the final output. This is the core reason for its TerminalBench record, but it consumes significantly more tokens — reserve it for genuinely complex tasks.
Following President Trump's June 2, 2026 executive order, the White House coordinated OSTP and ONCD to require OpenAI to conduct a government security review before broad release. OpenAI complied but publicly opposes this becoming permanent industry practice.
Starting July 2026, GPT-5.6 Sol on Cerebras hardware acceleration reaches up to 750 tokens per second — roughly 5 to 15 times faster than current flagship models at 50–150 token/s. Initial access is limited to select enterprise customers.
Choose Sol for complex coding and multi-step agent tasks; Terra for enterprise document analysis and high-volume API calls; Luna for summarization, drafting, and routine automation. For hardware runtime guidance, see the Help Center, or read our four-way coding assistant comparison.