What is GPT-5.6 and when was it released?

GPT-5.6 is OpenAI's latest frontier model family, comprising three tiers: Sol (flagship), Terra (balanced), and Luna (fast and affordable). It was previewed on June 26, 2026 for a limited group of approximately 20 trusted partner organisations, at the request of the US government, due to its powerful cybersecurity capabilities. OpenAI announced general availability on July 8, 2026, and publicly launched all three models on July 9, 2026 across ChatGPT, ChatGPT Work, Codex, and the OpenAI API. It was released alongside GPT-Live 1, OpenAI's new real-time bidirectional voice model.

What are the three GPT-5.6 models: Sol, Terra and Luna?

GPT-5.6 comes in three capability tiers. Sol is the flagship: OpenAI's most capable model to date, designed for complex reasoning, long-horizon coding, biology, cybersecurity, computer use, and professional knowledge work. It supports a new 'max reasoning effort' mode and an 'ultra mode' that uses subagents to accelerate complex work. Terra is the balanced tier: competitive with GPT-5.5 performance at half the cost, designed for everyday production work. Luna is the fast and affordable tier: strong capability at OpenAI's lowest cost, designed for high-volume, lower-complexity tasks. All three share a 1.05M token context window and 128K max output.

How much does GPT-5.6 cost?

GPT-5.6 is priced per 1 million tokens across the three model tiers. Sol: $5 input / $30 output. Terra: $2.50 input / $15 output. Luna: $1 input / $6 output. GPT-5.6 also introduces more predictable prompt caching: cache writes are billed at 1.25× the uncached input rate, while cache reads continue to receive a 90% discount. The system also supports explicit cache breakpoints and a 30-minute minimum cache life. OpenAI also announced GPT-5.6 Sol on Cerebras hardware at up to 750 tokens per second, initially limited to select customers.

What are GPT-5.6 Sol's benchmark results?

GPT-5.6 Sol achieves 88.8% on TerminalBench 2.1, the benchmark for command-line workflows requiring planning, iteration, and tool coordination. Sol Ultra (the enhanced subagent mode) achieves 91.9%, beating Claude Mythos 5 at 88.0%. Luna scores 82.5% on TerminalBench 2.1, beating Claude Opus 4.8 (78.9%) though below GPT-5.5 (83.4%). Terra scores 84.3%, beating its predecessor. On GeneBench v1 (long-horizon genomics and quantitative biology), Sol achieves stronger results than GPT-5.5 while using fewer tokens. On ExploitBench 2 (cybersecurity), Sol is competitive with Anthropic's Mythos Preview using only approximately one-third of the output tokens.

GPT-Live 1 (formerly known as GPT-Bidi 1 during development) is OpenAI's new real-time voice AI model family, launched simultaneously with GPT-5.6 on July 9, 2026. It features bidirectional audio capability, the models can listen and speak at the same time, making interaction feel 'much more like having a real conversation,' according to OpenAI's blog post. It represents a significant upgrade from the turn-based voice interaction of previous ChatGPT voice modes, where the model had to finish speaking before it could listen.

What does GPT-5.6 mean for the US-China AI race?

GPT-5.6 Sol's cybersecurity capability is the direct technical trigger for the escalating US-China AI restrictions dynamic. The model is competitive with Anthropic's Mythos on vulnerability research and exploitation benchmarks, the same category of capability that led the US government to restrict Mythos access to foreign nationals. That Sol achieved this at roughly one-third of Mythos's token cost makes it both more accessible and more strategically significant. The government-gated preview — OpenAI's first — establishes a new precedent: frontier AI models with advanced cybersecurity capabilities may now require government review before public release, just as they do for export to adversarial nations.

GPT-5.6 Is Here: Sol, Terra and Luna, and Why the US Government Had to Approve It First

Q: Why did the US government have to approve GPT-5.6 before public release?

The US government requested a controlled preview period before GPT-5.6's public release due to its significantly enhanced cybersecurity capabilities. GPT-5.6 Sol is OpenAI's most capable model yet for long-horizon security tasks including vulnerability research and exploitation. OpenAI had previewed GPT-5.6's plans and capabilities with the US government for approximately one month before the June 26 preview launch, including in meetings CEO Sam Altman had at the White House in early June. The company had not anticipated that the government would limit the launch to only around 20 trusted partners, it had expected a normal staggered rollout. OpenAI said clearly: 'We don't believe this kind of government access process should become the long-term default.'

Q: How does GPT-5.6 Sol compare to Claude Mythos and Fable 5?

On TerminalBench 2.1, Sol Ultra (91.9%) leads Claude Mythos 5 (88.0%), and Sol itself (88.8%) also beats Mythos. On ExploitBench 2 (cybersecurity and exploitation tasks), Sol is competitive with Mythos Preview while using approximately one-third fewer output tokens, a significant efficiency advantage. Anthropic's Fable 5 leads on SWE-Bench Pro at 80.3% (software engineering), while Sol leads on terminal and agentic coding workflows. The two models trade the lead across different task domains, consistent with the broader pattern of the Stanford 2026 AI Index finding that the US-China (and increasingly US-US) performance gap shifts by benchmark.

Q: What is Sol Ultra and what does it do?

Sol Ultra is an enhanced mode within GPT-5.6 Sol that goes beyond the capabilities of a single agent by deploying subagents to accelerate complex work in parallel. It is available in Codex and is positioned for the hardest long-horizon coding, multi-repository work, computer use, and extended agentic sessions. Sol Ultra achieves 91.9% on TerminalBench 2.1, the highest score in the benchmark, above GPT-5.6 Sol (88.8%) and Claude Mythos 5 (88.0%). OpenAI also introduced a new 'max reasoning effort' setting for Sol that gives the model maximum time to reason deeply before generating output.

Graphic artwork with bold typography and abstract geometric shapes in black, white, and red. On the upper left, the text reads 'US GOVERNMENT APPROVES' with 'APPROVES' in a large, stylized red serif font. In the center, a circular framing reveals a monochromatic illustration of the US Capitol dome merged with a complex, wireframe polygonal mesh. To the right, additional text reads 'ChatGPTSol, Terra and Luna' in varying fonts.

When a new AI model requires government approval before the public can use it, you know something has changed.

Not just in what the model can do. In what it means to release one.

On June 26, 2026, OpenAI previewed GPT-5.6 to roughly 20 carefully selected organisations, a list approved and shared with the US government before a single line of the announcement went live. The reason was not a production bottleneck or a typical staggered rollout. It was the model's cybersecurity capability: GPT-5.6 Sol is good enough at finding and exploiting software vulnerabilities that the government wanted to see it before the world did.

On July 9, 2026, OpenAI CEO Sam Altman wrote "Happy building" on X and GPT-5.6 went public for everyone. Three models. New naming conventions. New pricing. New benchmarks. And a new precedent for what it takes to ship frontier AI in 2026.

Here is everything that matters.

Three Models, One Family: What Sol, Terra and Luna Actually Are

GPT-5.6 is not a single model. It is a family, OpenAI's most significant naming restructure in years. The three tiers are named after celestial bodies, each with a distinct position in the capability-cost spectrum:

Sol, the flagship. OpenAI's most capable model ever. Built for complex reasoning, long-horizon coding, biology, cybersecurity, computer use, and professional knowledge work. Introduces "max reasoning effort" mode and "ultra mode" using subagents. · Terra, the balanced tier. Competitive with GPT-5.5 performance at half the cost. Designed for everyday production work where cost and quality both matter. · Luna, fast and affordable. Strong capability at OpenAI's lowest cost. The workhorse for high-volume, lower-complexity tasks.

All three models share a 1.05 million token context window and a 128,000 token maximum output. The API alias gpt-5.6 routes to Sol. Developers who want deterministic routing should specify gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna explicitly.

GPT-5.6 launched simultaneously with GPT-Live 1, formerly GPT-Bidi 1, OpenAI's new real-time bidirectional voice model. Unlike previous voice modes where the model had to finish speaking before it could listen, GPT-Live 1 can do both at once. OpenAI says it makes conversation feel "much more like having a real conversation."

The Benchmarks: What GPT-5.6 Can Actually Do

The performance claims are specific, and the benchmarks are the ones that matter in 2026.

Model	TerminalBench 2.1	Key Strength	Price (Input / Output per 1M)
GPT-5.6 Sol Ultra	91.9%	Agentic coding, multi-repo, subagent orchestration	Sol tier ($5 / $30) + usage
GPT-5.6 Sol	88.8%	Coding, biology, cybersecurity, reasoning	$5 / $30
Claude Mythos 5	88.0%	Cybersecurity, vulnerability research	Restricted access
GPT-5.6 Terra	84.3%	Everyday production work, cost-efficiency	$2.50 / $15
GPT-5.5	83.4%	Previous OpenAI flagship (April 2026)	Previous pricing
GPT-5.6 Luna	82.5%	High-volume, fast, affordable tasks	$1 / $6
Claude Opus 4.8	78.9%	Anthropic's balanced frontier model	Anthropic pricing

Three additional benchmark results stand out:

GeneBench v1 (genomics and quantitative biology): Sol achieves stronger results than GPT-5.5 while using fewer tokens, efficiency and capability improving simultaneously.
ExploitBench 2 (cybersecurity and exploitation): Sol is competitive with Anthropic's Mythos Preview using only approximately one-third of the output tokens. This is the benchmark that triggered government review.
Coding efficiency: OpenAI says GPT-5.6 is 54% more token-efficient on agentic coding tasks compared to previous models, the stat Sam Altman cited to CNBC as the most practically significant improvement.

The Government Approval That Changed Everything

The story of GPT-5.6's release is as much about what happened before launch as what happened on launch day.

OpenAI had been previewing GPT-5.6 with the US government for approximately one month before the June 26 announcement, including in meetings Sam Altman had at the White House in early June. The concern was Sol's cybersecurity performance, specifically its ability to help with vulnerability research and exploitation. This is precisely the category of capability that led the US government to restrict Anthropic's Mythos model earlier the same month.

OpenAI had not anticipated a government-gated launch of this kind. The company expected a normal, if careful, staggered rollout. Instead it found itself limited to approximately 20 pre-approved partner organisations at launch, a list the government had to sign off on before a single API call could go through.

OpenAI's response was careful but pointed: "We believe in broad access, and we plan to make GPT-5.6 Sol, Terra, and Luna generally available in the coming weeks. 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."

The company said the US government was aware of its plans to launch broadly and had "expressed support for those plans barring any concerns in the additional testing period." On July 8, OpenAI announced it was expanding preview access globally. On July 9, the models went live.

The two-week delay established something important: the US government now considers frontier AI cybersecurity capability to be a matter of national security review, not just export control, but domestic release gating. This is new. It will not be the last time it happens.

What This Means For How You Use AI

The GPT-5.6 family is not just a capability upgrade. It is a pricing and architecture signal that tells you how OpenAI thinks about the next phase of AI deployment.

Terra is the most strategically revealing model in the family. At $2.50 input / $15 output, half the cost of Sol, it delivers performance competitive with GPT-5.5. That means the previous generation's flagship capability is now the middle tier. For most enterprise use cases, Terra is the right default: it gives you near-frontier performance at a price that doesn't require a tokenmaxxing governance framework to control.

Luna at $1 / $6 changes the economics of high-volume AI deployment. At those prices, tasks that were previously expensive to automate at scale, customer service triage, document classification, structured data extraction, become trivially affordable. The Jevons Paradox warning applies: low token prices do not mean low bills if you use them for everything.

Sol is the research and specialist tool. The 54% agentic coding efficiency improvement is real and significant, it means fewer tokens per completed task, not just better output per token. For teams running Codex-powered development workflows, Sol Ultra's subagent orchestration changes what is possible in a single session.

The Cerebras integration, GPT-5.6 Sol at up to 750 tokens per second, is the most forward-looking element of the launch. At that inference speed, real-time agentic workflows become practically viable in a way they weren't before. Speed, not just capability, is now a frontier metric.

GPT-5.6 is the first model family in OpenAI's history to require US government approval before public release. That fact tells you more about where AI is heading than any benchmark number. The tools are now capable enough that governments want to see them before the world does. Whether you find that reassuring or alarming probably depends on which government you trust more than the others.

The Silicon Curtain

The same cybersecurity capability that triggered GPT-5.6's government review is why Beijing is discussing restricting access to China's own frontier models. The two stories are connected.

China Declares Claude Code a National Security Threat

While GPT-5.6 was going through US government review, China was declaring Anthropic's Claude Code a national security risk. The AI security story of July 2026, explained.

The Tokenpocalypse

GPT-5.6 Luna at $1 per million input tokens. The Jevons Paradox warning: cheaper tokens do not mean lower bills. The corporate AI cost crisis that's happening right now.

GPT-5.6: FAQ

GPT-5.6 is OpenAI's latest frontier model family, comprising three tiers: Sol (flagship), Terra (balanced), and Luna (fast and affordable). It was previewed on June 26, 2026 for approximately 20 government-approved trusted partner organisations due to its advanced cybersecurity capabilities. OpenAI announced general availability on July 8, 2026, and publicly launched all three models on July 9, 2026 across ChatGPT, ChatGPT Work, Codex, and the OpenAI API. It launched alongside GPT-Live 1, OpenAI's new real-time bidirectional voice model.

GPT-5.6 comes in three tiers. Sol is the flagship: OpenAI's most capable model ever, designed for complex reasoning, long-horizon coding, biology, cybersecurity, computer use, and professional knowledge work. It supports a new "max reasoning effort" mode and "ultra mode" using subagents. Terra is the balanced tier: competitive with GPT-5.5 at half the cost, for everyday production work. Luna is fast and affordable: strong capability at OpenAI's lowest cost, for high-volume tasks. All three share a 1.05M token context window and 128K max output.

GPT-5.6 pricing per 1 million tokens: Sol, $5 input / $30 output. Terra, $2.50 input / $15 output. Luna, $1 input / $6 output. GPT-5.6 introduces more predictable prompt caching: cache writes billed at 1.25× the uncached input rate; cache reads receive a 90% discount. Supports explicit cache breakpoints and a 30-minute minimum cache life. GPT-5.6 Sol is also launching on Cerebras hardware at up to 750 tokens per second in July 2026, initially limited to select customers.

The US government requested a controlled preview due to GPT-5.6 Sol's enhanced cybersecurity capabilities, specifically its performance on vulnerability research and exploitation tasks. OpenAI had previewed the models with the government for approximately one month before launch, including in White House meetings with CEO Sam Altman in early June. The government requested that OpenAI start with a limited preview of approximately 20 approved partner organisations before broader release. OpenAI cooperated but stated clearly: "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."

On TerminalBench 2.1 (command-line agentic workflows): Sol Ultra scores 91.9%, Sol scores 88.8%, Terra scores 84.3%, Luna scores 82.5%. Claude Mythos 5 scores 88.0% and Claude Opus 4.8 scores 78.9% for comparison. On GeneBench v1 (genomics and quantitative biology): Sol beats GPT-5.5 while using fewer tokens. On ExploitBench 2 (cybersecurity): Sol is competitive with Anthropic's Mythos Preview using only approximately one-third of the output tokens. OpenAI also reports a 54% improvement in token efficiency on agentic coding tasks compared to previous models.

On TerminalBench 2.1, Sol Ultra (91.9%) and Sol (88.8%) both lead Claude Mythos 5 (88.0%). On ExploitBench 2 (cybersecurity), Sol is competitive with Mythos Preview using approximately one-third fewer output tokens — a significant efficiency advantage. Anthropic's Fable 5 leads on SWE-Bench Pro at 80.3% (software engineering). The two model families trade benchmark leads across different task domains, consistent with the Stanford 2026 AI Index finding that no single model holds a dominant cross-domain lead, capability at the frontier is now task-specific rather than categorical.

Sol Ultra is an enhanced mode within GPT-5.6 Sol that deploys multiple subagents working in parallel to accelerate complex work. Available in Codex and the ChatGPT desktop app, it is positioned for long-horizon coding, multi-repository work, computer use, and extended agentic sessions. Sol Ultra achieves 91.9% on TerminalBench 2.1, the highest published score on the benchmark, above Sol (88.8%) and Claude Mythos 5 (88.0%). GPT-5.6 also introduces a "max reasoning effort" setting for Sol that gives the model maximum time to reason before generating output — separate from the Ultra subagent mode.

GPT-Live 1 (formerly GPT-Bidi 1) is OpenAI's new real-time voice AI model family, launched simultaneously with GPT-5.6 on July 9, 2026. It features bidirectional audio capability, the models can listen and speak at the same time, rather than requiring turn-based interaction. OpenAI says it makes conversation feel "much more like having a real conversation." It represents a significant upgrade from previous ChatGPT voice modes and is positioned as the foundation for natural AI voice interaction in consumer and enterprise applications.

As of July 9, 2026, GPT-5.6 is available across ChatGPT, ChatGPT Work, Codex, and the OpenAI API. In ChatGPT: Plus, Pro, Business, and Enterprise users access Sol through medium and higher effort settings; Pro and Enterprise can also select Sol Pro. In ChatGPT Work and Codex: Free and Go users start with Terra; Plus and higher plans choose Sol, Terra, or Luna depending on rollout and workspace settings. In the API: model IDs are gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna. The alias gpt-5.6 routes to Sol. Codex is also moving into the ChatGPT desktop app with this release.

GPT-5.6 Sol's cybersecurity capability is the direct technical trigger for the escalating US government control dynamic. The model is competitive with Anthropic's Mythos on vulnerability research at one-third the token cost, and Mythos's cybersecurity capability was precisely what led the US to restrict its access to foreign nationals. That GPT-5.6 achieves comparable performance more efficiently makes it both more accessible and more strategically sensitive. The government-gated preview establishes a new precedent: frontier AI cybersecurity capabilities may now require government review before domestic public release, the same logic as export control, applied domestically. Beijing's simultaneous discussions to restrict its own frontier model exports make this a symmetrical pattern, not a unilateral one.

Jans Bock-Schroeder

Publisher & Founder of AI Angst

Coming from the world of art, photography, and the luxury market, Jans launched AI Angst in 2025 to explore the cultural, ethical, and psychological impacts of artificial intelligence. His work bridges creative vision with critical technology analysis, offering clarity in an era of rapid technological change.

Sources and Citations

This article is based on the following primary sources, published June 26 – July 9, 2026:

OpenAI: "Previewing GPT-5.6 Sol: a next-generation model" (June 26, 2026)
Official primary source for all model specifications: Sol/Terra/Luna tier structure, pricing ($5/$30, $2.50/$15, $1/$6 per 1M tokens), 1.05M context window, 128K max output, Sol Ultra subagent mode, max reasoning effort, Cerebras at 750 tokens/sec, GeneBench v1 results, ExploitBench 2 results, government preview rationale, and safety stack description.
https://openai.com/index/previewing-gpt-5-6-sol/
Wikipedia: "GPT-5.6" (July 9, 2026)
Source for general availability date (July 9, 2026), TerminalBench 2.1 scores (Sol 88.8%, Sol Ultra 91.9%, Terra 84.3%, Luna 82.5%, Claude Mythos 5 88.0%, Claude Opus 4.8 78.9%, GPT-5.5 83.4%), and GPT-Live 1 / GPT-Bidi 1 naming history.
https://en.wikipedia.org/wiki/GPT-5.6
CNBC: "OpenAI to publicly release GPT-5.6, rolls out conversational AI models" (July 8, 2026)
Source for Sam Altman "Happy building" post, "expanding preview access globally," 54% agentic coding efficiency improvement quote, GPT-Live announcement, and Anthropic context.
https://www.cnbc.com/2026/07/08/openai-expanding-gpt-5point6-ai-model-release-ending-government-limits.html
Axios: "OpenAI releases powerful new GPT-5.6 model under restrictions" (June 26, 2026)
Source for approximately 20 trusted partner organisations detail, White House meeting context (early June), government sign-off on partner list, and "not anticipate severe restrictions" characterisation.
https://www.axios.com/2026/06/26/openai-gpt-sol-terra-luna-trump
Nextgov/FCW: "OpenAI's advanced GPT-5.6 models to be publicly released" (July 8, 2026)
Source for government partners safety evaluation framing, "real-world adversarial pressure" quote, and "expanding preview access globally now" X post text.
https://www.nextgov.com/artificial-intelligence/2026/07/openais-advanced-gpt-56-models-be-available-public/414651/
Coursiv: "GPT-5.6 Sol: ChatGPT Release Date, Price & Review" (July 9, 2026)
Source for ChatGPT access tier detail (Plus/Pro/Business/Enterprise), API model IDs (gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.6 alias), and Codex desktop app integration.
https://coursiv.io/blog/chatgpt-5-6-sol

Published: July 9, 2026. All sources verified at time of publication. All external links open in a new tab.