GPT-5.6 Sol, Terra & Luna:
Full Review, Benchmarks, Pricing & Access Guide

Availability today: GPT-5.6 is limited to roughly 20 government-approved trusted partners via API and Codex preview. General ChatGPT access is expected within weeks. Polymarket traders assign an 87% probability to broad release by July 31, 2026.

The launch was politically fraught. A Trump executive order signed June 2, 2026 gave U.S. agencies up to 30 days of pre-release access to frontier models. This is the first time the U.S. government formally required an AI company to restrict a frontier release. OpenAI complied but pushed back publicly:

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.

What hurts developers right now:

• No general API yet: Most teams cannot validate Sol's Ultra multi-agent mode in production.
• Competitors also blocked: Claude Mythos 5 went offline June 12 under export controls; Gemini 3.5 Pro slipped to July.
• Tier confusion: Sol/Terra/Luna pricing and capability gaps are large — benchmarks matter more than branding.
• Safety compliance: All three tiers carry OpenAI's "High" cyber rating; enterprises need abuse controls and audit plans early.
• Infra timing: July brings Cerebras 750 t/s plus public API — teams without elastic nodes risk integration bottlenecks on day one.

OpenAI named the family after celestial bodies for the first time: Sol (Sun) for flagship capability, Terra (Earth) for balanced cost/performance, and Luna (Moon) for fast, affordable throughput.

GPT-5.6 Sol — flagship

Sol targets hard coding, long-horizon security research, and multi-step agentic workflows. Pricing matches GPT-5.5: $5 / $30 per million tokens input/output. Context window is roughly 1.5M tokens.

Two new reasoning modes:

• Max mode: Extra reasoning time before answering — trades latency for accuracy on high-stakes tasks.
• Ultra mode: Spawns multiple parallel subagents that split work, execute concurrently, and merge results. This architecture drives the 91.9% TerminalBench record but uses significantly more tokens.

GPT-5.6 Terra — balanced workhorse

Terra is the default for high-volume business work: support tickets, internal tools, document analysis. Performance is close to GPT-5.5 at half the cost ($2.50 / $15 per MTok).

GPT-5.6 Luna — lightweight

Luna optimizes for summarization, drafting, and routine automation. It is OpenAI's first non-flagship model rated High in both cybersecurity and biology, at $1 / $6 per MTok.

Coding: TerminalBench 2.1

TerminalBench 2.1 runs 89 complex command-line planning challenges — closer to real agent work than single-shot code completion.

Claude Mythos 5 held the top spot for only 17 days (since June 9) before Sol passed it.

Long-horizon agents: Agent's Last Exam

• GPT-5.6 Sol: 50.9% task completion — the only model above 50%
• GPT-5.6 Luna: Slightly above GPT-5.5

Cybersecurity: CTF and ExploitBench

GPT-5.6 is the first OpenAI family where all three tiers hit the "High" cybersecurity classification.

ExploitBench: Sol matches Anthropic's Mythos Preview while using only about one-third of the output tokens — same security research depth at dramatically lower cost.

Safety note: OpenAI red-teaming confirmed Sol cannot autonomously engineer a complete, functional exploit chain against hardened Chromium or Firefox targets. It stays below OpenAI's "Cyber Critical" threshold.

Life sciences: On GeneBench v1, Sol matches or beats GPT-5.5 with fewer tokens. HealthBench Professional: 60.5 (+8.7 vs GPT-5.5).

Cerebras acceleration (July 2026)

Starting in July, Sol on Cerebras hardware targets up to 750 tokens per second. Most frontier models today run 50–150 t/s — a 5× to 15× speedup. A 10-second response could finish in under one second for real-time coding assistants and streaming agents. Initial access is limited to select enterprise customers as capacity expands.

Built-in safeguards: Real-time misuse classifiers, account-level review for sensitive flows, 700,000 A100-equivalent GPU hours of automated red-teaming, universal jailbreak testing, a specialized reasoning filter as backup, and external security audits before launch.

Bottom line: Sol leads on TerminalBench and offers comparable security research at half the price. Mythos 5 may still lead on benchmarks like SWE-Bench Pro where GPT-5.6 system card data is not fully published yet.

June 2026: all three frontier labs got blocked

June was supposed to be the biggest month in AI history. Instead, all three flagship releases stalled at the door — setting a precedent for how frontier models may deploy globally.

Access timeline:

• Now (June 2026): ~20 approved partners via API and Codex only.
• Coming July 2026: General ChatGPT (Plus/Pro first), public API, Cerebras Sol up to 750 t/s for enterprise.

Six steps before general availability:

1. Keep production on GPT-5.5 or Claude Opus 4.8 until Sol is broadly available — do not break agent pipelines on preview-only access.
2. Pre-wire LiteLLM routes with placeholders for gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna so you can A/B test via env vars on day one.
3. Set Ultra cost guardrails: Ultra multi-agent mode burns tokens fast — add max_tokens and complexity thresholds in code.
4. Version Cursor Rules and Agent Skills in Git for one-hour benchmark replay when API opens.
5. Watch OpenAI Status and Polymarket: 87% implied odds for release by July 31 — useful signal, not an official schedule.
6. Stand up a 24/7 bare-metal node with Codex CLI, OpenClaw Gateway, and self-hosted runners so model switches are hot-swaps, not rebuilds.

Citable technical data:

• TerminalBench 2.1: Sol Ultra 91.9%, standard 88.8%, Mythos 5 88.0%, GPT-5.5 83.4%
• CTF hit rates: Sol 96.7%, Terra 91.84%, Luna 85.19%
• Cerebras: up to 750 t/s in July — roughly 5–15× today's frontier speeds
• Pricing: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per MTok
• Context: ~1.5M tokens (reconfirm when the full system card ships)

FAQ

• Is GPT-5.6 on ChatGPT now? Not for the general public. ~20 trusted partners only; full ChatGPT rollout expected within weeks (July 2026).
• Is GPT-5.6 Sol better than Claude Fable 5 for coding? Sol leads TerminalBench 2.1 (91.9% vs Mythos 5's 88%). Fable 5 still leads SWE-Bench Pro, but is offline under export controls.
• What is Ultra mode? Multiple parallel subagents split a task, work concurrently, and synthesize one answer — higher performance, higher token cost.
• Why is GPT-5.6 restricted? U.S. government review under Trump's June 2 order, coordinated via White House OSTP and ONCD. OpenAI complied but opposes making this permanent.
• How fast is GPT-5.6 on Cerebras? Up to 750 tokens per second — about 5–15× most current frontier models, July 2026 for select enterprise customers.
• What is the GPT-5.6 context window? Reported at ~1.5M tokens, up from GPT-5.5's 1M — official confirmation expected with the full system card.
• Are all three GPT-5.6 models safe for cybersecurity work? All carry OpenAI's "High" cyber rating with layered safeguards; models cannot autonomously build complete functional exploits.

Sources (re-check links after any upstream update):

OpenAI: Previewing GPT-5.6 Sol

OpenAI Deployment Safety System Card

VentureBeat: GPT-5.6 Launch Coverage

SiliconAngle: GPT-5.6 vs Claude Mythos 5

TechTimes: Government Lock Analysis

Cloud IDEs and shared VPS instances often fail agent workloads: laptops sleep mid-run, teams collide on one instance, and launchd daemons cannot stay up 24/7. For production Cursor agents, Codex CLI, and iOS CI/CD, CALMVPS bare-metal Mac Mini rental delivers dedicated Apple Silicon, 120-second provisioning, and monthly billing — switch OPENAI_MODEL on an isolated node when GPT-5.6 opens without rebuilding your stack.