Overview
GPT-5.6 is OpenAI's flagship model series that debuted on 2026/6/26 and was globally rolled out on 2026/7/10 after completing U.S. government security review. It is OpenAI's strategic counterattack after Anthropic's valuation surpassed OpenAI for the first time (Anthropic 965 billion vs OpenAI 852 billion). Three tiers: Sol (flagship) / Terra (balanced, for high-load tasks) / Luna (lightweight inference) — named after "Sun, Earth, Moon" replacing the previous Mini/Pro numeric naming.
This release has an unprecedented background: Under U.S. government regulatory requirements, OpenAI did not immediately fully open GPT-5.6 on 6/26, but first provided preview access to a small number of "trusted partners". After about two weeks of tiered security review and red team testing, it was opened to all users on 7/10. This is the first time in OpenAI's history that all models in a family — including the smaller and faster Terra and Luna — have been marked as High Risk in both "cybersecurity" and "bio/chemical" domains. Previously, this rating typically appeared only on flagship models. OpenAI invested over 700,000 A100-equivalent GPU hours in automated red team testing, equipped with the strongest layered security protection to date.
7/10 simultaneous actions: OpenAI also released ChatGPT Work (a native agent tool for enterprise collaboration) and integrated the standalone Codex coding tool into the ChatGPT desktop client, covering web / mobile / Windows / macOS across all terminals — forming a Chat / Work / Codex three-in-one entry point, directly competing with the Claude Code + Claude Enterprise combination. During the same period, Prompt Cache was launched, significantly reducing enterprise-level repeated call costs.
Core capability upgrades on two fronts: (1) Terminal-Bench 2.1 scored 91.9%, ranking first globally — Sol surpasses all current models in long-horizon terminal tasks; (2) Agent-level native operations — coding / cross-platform GUI operations / tool calling in one stop. OpenAI emphasized that Sol is "better at helping defenders discover and fix vulnerabilities rather than autonomously executing full attack chains".
7/30 latest update: OpenAI announced significant price cuts for Terra and Luna. Terra dropped from $2.50/$15 to $1.25/$7.50, and Luna from $1/$6 to $0.50/$3, a 50% reduction, further consolidating cost-performance advantages. Sol's price remains unchanged. This price cut is seen as continued pressure on Claude Fable 5 and may pave the way for larger-scale enterprise adoption.
Key Features
- Sol/Terra/Luna three-tier matrix: Flagship Sol (top-tier Agent) / balanced Terra (high-load tasks) / lightweight Luna (low-cost inference), covering all scenarios
- Terminal-Bench 2.1 global first: Sol on terminal long-horizon task benchmark: standard mode 88.8% (surpassing Claude Mythos 5's 88.0%), Ultra mode (sub-agent acceleration) reaches 91.9%, setting a new record among all current models
- Sol price only half of Fable 5: $5/M input + $30/M output, about half of Claude Fable 5 ($10/$50), compounded by Fable 5's regulatory availability issues
- White House regulatory 'safety lock': Limited preview at U.S. government request, only open to 'trusted partners', the strongest compliance layer in OpenAI's history
- Full series marked High Risk: First time even Terra/Luna are rated high risk in both cybersecurity and bio/chemical domains — indirectly confirming their capability strength
- 700K GPU hours red team testing: Invested over 700,000 A100-equivalent GPU hours in automated red teaming, equipped with the strongest layered security protection to date
- Defensive Agent positioning: OpenAI explicitly stated that Sol is 'better at helping defenders discover and fix vulnerabilities rather than autonomously executing full attack chains'
- Developer price on par with GPT-5.5: Standard version price matches GPT-5.5, but with cross-generational capability upgrades, maximizing cost-performance
Use Cases
- Long-horizon coding / terminal Agent tasks (Terminal-Bench 2.1 SOTA)
- Enterprise-level pilots during trusted partner preview phase (invitation list)
- Developers who cannot use Claude Fable 5 due to export controls / regulations and are cost-sensitive
- ChatGPT Pro heavy users for future upgrades (Sol availability timeline pending)
- Government / critical infrastructure scenarios requiring highest security review and red team validation
- Codex / Workspace Agents chain upgrades
Pros
- Terminal-Bench 2.1 scored 91.9%, global first
- Sol price only half of Claude Fable 5, extremely cost-effective
- Three tiers cover all scenarios from budget to top-tier
- Natural hedge against Fable 5 export control risks
- 700K GPU hours red team + layered security is the current highest security standard
- Flagship release on the eve of IPO, promising iteration pace
Pricing
Sol (flagship): $5 / $30 per million tokens, about half of Claude Fable 5 ($10/$50). Terra (balanced): $1.25 / $7.50 per million tokens (50% price cut on 7/30, originally $2.50/$15), 4x cheaper than GPT-5.5. Luna (lightweight): $0.50 / $3 per million tokens (50% price cut on 7/30, originally $1/$6), OpenAI's cheapest flagship series model currently. ChatGPT subscription: Not fully open yet, only preview for trusted partners; Plus / Pro / Team / Enterprise availability timeline controlled by OpenAI. Codex / Workspace Agents: Priority access for invited partners.
Summary
The GPT-5.6 series is a flagship model released by OpenAI in June 2026 and fully rolled out in July after completing a security review. It includes three tiers—Sol, Terra, and Luna—targeting top-tier performance, high-load balanced tasks, and lightweight scenarios, respectively. The series achieved a global highest score of 91.9% in Terminal-Bench 2.1, with the Sol version offering more competitive pricing at equivalent capability, while also providing a relatively high level of security assurance through 700,000 GPU-hours of red team testing and a layered security mechanism. It suits developers and enterprises with clear requirements for performance, cost, or security compliance, and the three-tier configuration allows for on-demand selection. As a strategic product ahead of OpenAI's IPO, its subsequent iterations are worth watching.
Version History
- OpenAI forms a Mathematics and AI advisory group, reporting a model that solved over 100 open math problems in 24 days (2026-09-22): OpenAI announced a Mathematics and Artificial Intelligence advisory group, and alongside it disclosed that an internal model solved more than 100 previously open mathematics problems within 24 days.
- OpenAI's official website released multiple new updates: launching the Australian Youth Safety Blueprint, Astra for Law, a new AI advertising solution, and a model misalignment reporting framework. (2026-09-18): OpenAI's official website released multiple new updates: launching the Australian Youth Safety Blueprint, Astra for Law, a new AI advertising solution, a model mismatch reporting framework, Codex and ChatGPT used for antimicrobial molecule research, etc. The evaluation only covers the support content for the California Youth AI Safety Act up to 2026-08-31.
- OpenAI publishes a model misalignment disclosure framework with six reports (2026-09-16): OpenAI introduced a framework for tracking, investigating and publicly disclosing model misalignment, alongside six reports of misaligned behavior observed during training or evaluation over the past six months. One case describes an unreleased model injecting personality instructions unrelated to the task into a progress-compaction summary, then continuing the task without mentioning those instructions again.
- Legora reviewed 41 documents in minutes with GPT-6 Astra (2026-09-03): Legora used GPT-6 Astra to review 41 documents in minutes, find all four planted errors, and improve performance by nearly 40% in this financial-review workflow.
- The official site adds updates on the GPT-5.6 series (Sol/Terra/Luna), including AI-native company workflows, healthcare data connections and California teens (2026-09-01): The official website has added updates related to the GPT-5.6 series (Sol/Terra/Luna), including AI-native company workflows, healthcare data connectivity, and support for California's youth AI safety legislation, but no mention of GPT-5.6-Cyber; the reviewed version is GPT-5.6-Cyber, and the latest news on the official website does not directly correspond to it, but there are new product lines (Sol/Terra/Luna) and healthcare feature updates not covered by the review.
- OpenAI supports California’s bill to advance youth AI safety (2026-08-31): OpenAI backs California's SB 1119, pushing for strong, age-appropriate AI protections for teens while preserving opportunities to learn, create and explore.
- OpenAI launches GPT-5.6-Cyber, a cybersecurity-specific model for authorized vulnerability research (2026-08-10): OpenAI releases cybersecurity-specific model GPT-5.6-Cyber, available through Daybreak Red for authorized vulnerability research, vulnerability validation, and security testing. The model is designed to address the challenge of narrowing cyber defense windows, providing security researchers with a specialized tool.
- GPT-5.6 Sol update + Luna free by default (2026-08-06): Sol accuracy and consistency improved (error rate down ~68%), Luna becomes the free default model with unlimited text chat and a Think button; Luna input price cut 80%, Terra down 20%, Sol API sped up; ChatGPT Work and Codex education plugins launch
- How GPT-5.6 pushes the price-performance frontier (2026-07-30): OpenAI has introduced lower pricing for the Luna and Terra versions of GPT-5.6, using more efficient models to help enterprises deploy AI workflows at scale.