OpenAI Launches GPT-5.6: Three Tiers, Government Review, and a Cyber Variant

On 9 July 2026, OpenAI ended a two-week restricted preview and released GPT-5.6, a family of three models that share a 1.05-million-token context window but are priced and positioned for very different workloads. The launch followed an unusual pre-release review requested by Washington: the model was announced on 26 June but initially limited to a small set of US-government-preapproved partners, then cleared for public release after technical testing and meetings with US Commerce Department officials.

The release matters because it marks a structural shift in how frontier AI is shipped. Instead of a single flagship, OpenAI now offers Sol, Terra and Luna as explicit cost-routing options, while adding multi-agent reasoning modes. More consequentially, the staggered rollout and the later launch of GPT-5.6-Cyber raise a governance question that will define the rest of 2026: who decides when a model is safe enough to ship, and who gets access first.

A three-tier family built for cost routing

OpenAI did not release one model but three. GPT-5.6 Sol is the flagship, aimed at complex reasoning, coding, science, cybersecurity and agentic work. Terra is the mid-tier, positioned by OpenAI as “GPT-5.5-level capability at roughly half the cost.” Luna is the fast, low-cost tier for high-volume, routine workloads. All three share a 1.05M-token context window and a 128K maximum output, with a knowledge cutoff of 16 February 2026. The API alias gpt-5.6 routes to Sol by default.

The family also introduces two new reasoning modes. Max triggers deeper single-model reasoning, while ultra enables multi-agent orchestration with four agents by default. For enterprise buyers, the practical consequence is that a single model family can now be routed by task complexity and budget, rather than forcing every request through the most expensive tier.

Pricing, benchmarks and the economics of Sol, Terra and Luna

The API pricing underscores the tiering strategy. Per 1 million tokens, input and output prices are:

• Sol: $5.00 input / $30.00 output
• Terra: $2.50 input / $15.00 output
• Luna: $1.00 input / $6.00 output

Cached reads receive a 90% input discount, while cache writes cost 1.25 times the uncached input rate. OpenAI CEO Sam Altman claimed Sol is “54% more token efficient on tasks” than competing models, and the company cited a score of 52.7% on Agents’ Last Exam for Sol, compared with 46.9% for GPT-5.5.

“54% more token efficient on tasks” — Sam Altman on GPT-5.6 Sol

On 21 August 2026, OpenAI cut GPT-5.6 Sol API and credit pricing by more than 20% for three months, citing competition from Anthropic and Chinese models. Independent analysts caution that the benchmark numbers are vendor-reported and advise running task-level evaluations before migrating production traffic. The price cut, however, signals that frontier capability is becoming a commodity faster than many expected.

Government pre-release review and the Cyber variant

GPT-5.6 arrived under an unusual policy shadow. In June 2026, US President Donald Trump signed an executive order creating a “voluntary” 30-day pre-release review process for frontier AI models. OpenAI announced GPT-5.6 on 26 June but restricted access to a small set of US-government-preapproved partners, at Washington’s request. The Trump administration green-lit the public launch on 8 July after technical testing and meetings with US Commerce Department officials. General availability followed on 9 July across ChatGPT, Codex and the OpenAI API, rolling out over 24 hours.

The White House denies that the process amounts to a de facto approval regime. OpenAI itself has been careful, stating publicly that it does not believe the pre-approval process “should become the long-term default,” calling it a short-term step. Still, the sequence established a precedent: a frontier model was held back from general release until government officials had reviewed it.

In August 2026, OpenAI launched GPT-5.6-Cyber, an “offense-grade” vulnerability-finding model under its Daybreak program, with reduced safety refusals for vetted defenders. OECD.AI flagged GPT-5.6-Cyber as an AI hazard due to dual-use risk. Researchers warn that the model’s ability to find code vulnerabilities could be exploited, while OpenAI says internal testing shows it is better at finding flaws than executing full attacks.

Competitive pressure and the road ahead

The release did not happen in a vacuum. Anthropic faced a similar export-control directive weeks earlier, and its Mythos series drew parallel security concerns. Meta released Muse Spark 1.1 and Muse Image the same week, intensifying competition. Meanwhile, the AI Now Institute argues that AI’s benefits are “overstated and underproven,” with gains accruing to companies rather than workers or the public.

For executives and founders, the immediate takeaways are concrete: Terra offers near-flagship performance at half the price, Luna enables high-volume automation at $1 per million input tokens, and the August price cut signals that cost pressure is accelerating. For security and policy professionals, GPT-5.6-Cyber and the White House’s informal review process raise a harder question: if pre-release government review becomes routine, who decides what is safe enough to ship, and who gets access first?

The next months will test whether tiered model families and government pre-release review become standard practice across the industry. OpenAI has positioned GPT-5.6 as a platform for ambition, but the unresolved governance question may prove more consequential than any single benchmark score.