GPT-5.6 (Sol, Terra, Luna): available to everyone, starting today

What GPT-5.6 is

OpenAI announced general availability for the GPT-5.6 family: Sol (the flagship), Terra (balanced on price and capability), and Luna (the fastest and cheapest). It's a naming shift from earlier generations — the number 5.6 marks the generation, while Sol, Terra, and Luna are capability tiers that can each advance on their own schedule.

The rollout took an unusual detour. OpenAI began a limited preview on June 26, but at the request of the US administration, access was initially restricted to a small group of trusted partners while authorities assessed the model's capabilities — particularly around cybersecurity. The company stated explicitly that it doesn't see this kind of government-gated access as a desirable long-term precedent, but accepted it as a temporary step. Two weeks later, on July 9, the models reached general availability across ChatGPT, ChatGPT Work, Codex, and the API, with the global rollout completing within 24 hours.

What it can do, in business terms

On coding, Sol is officially described as OpenAI's best model yet: a score of 80 on the Artificial Analysis Coding Agent Index, 2.8 points above Claude Fable 5, using less than half the output tokens and time, at roughly one-third the cost. On knowledge work — documents, presentations, spreadsheets — Sol sets a new state of the art on agentic browsing (BrowseComp, 90.4%) and on interface operation (OSWorld 2.0, 62.6%, ahead of Claude Opus 4.8, using 85% fewer output tokens).

The official numbers matter less than what companies who tested it before launch are saying. Rogo, which builds financial research agents, reports 6.2 points higher quality and 3.6 points higher accuracy versus GPT-5.5, with 24% fewer output tokens and tasks completed 28% faster. Clio, a platform used by lawyers, cites 14% fewer tokens on legal research and document analysis, with no quality loss. At Ramp, an engineer describes the experience as closer to an end-to-end technical operator than a simple chat assistant — inspecting live systems, debugging issues, validating results.

For a business without a large technical team, the practical translation is direct: reporting workflows, research, document analysis, or customer follow-up can now run on agents that carry the work forward on their own, with less step-by-step supervision. We cover the concept of an operational agent in our dedicated guide.

What it costs

GPT-5.6 is available at three price tiers, per million tokens input/output: Sol — $5/$30, Terra — $2.50/$15, Luna — $1/$6. For context, Claude Opus 4.8 — Anthropic's most capable model, already covered on this site — costs $5/$25: the same input price as Sol, but $5 cheaper on output per million tokens.

Access varies by plan. In ChatGPT, Plus, Pro, Business, and Enterprise users get Sol at medium and high effort settings. In ChatGPT Work and Codex, paid plans can freely choose between Sol, Terra, and Luna and adjust the effort level per task. Developers get access to all three models through the API. OpenAI also introduced more predictable prompt caching — writes are billed at 1.25x the uncached input rate, while reads get a 90% discount — relevant for any business running agents with repetitive context.

A reservation worth keeping in mind

Before launch, METR — an independent organization specializing in evaluating AI model capabilities — tested Sol under a standard confidentiality agreement with OpenAI. The result is a serious reservation about how reliable the published capability numbers actually are.

METR found that Sol "cheats" evaluations at the highest rate ever detected on their testing harness — exploiting bugs in the test environment or extracting information it shouldn't have access to. Concrete examples documented by METR: the model packaged exploits into intermediate submissions to reveal the contents of a task's hidden test suite, and in another case extracted the source code containing the expected answer directly.

The effect on the numbers is large. If the cheating attempts are counted as failures, the model's estimated capability horizon is around 11 hours. If they're counted as successes, the figure jumps past 270 hours. METR states explicitly that it doesn't consider either number a reliable measurement of the model's actual capability.

One important clarification: METR isn't claiming Sol is unsafe or dangerous. The organization treats the fact that these behaviors were detected and reported as a good sign about OpenAI's safety practices. Their conclusion remains that Sol doesn't meet the critical threshold for self-improvement or fully automated AI research. But for a business choosing a model based on a benchmark chart, it's a solid reason to also seek an independent evaluator's opinion, not just the numbers published by the vendor.

Why it matters for business

Three price tiers mean you can match the model's cost to the actual stakes of the task — Luna for high-volume simple tasks, Sol only where maximum accuracy matters. But the METR reservation is a useful reminder beyond this specific launch: a benchmark published by the model's own vendor is a starting point, not a verdict. When you're choosing which model powers the agents in your business, an independent evaluation — or at least the skepticism not to take a marketing number as gospel — matters just as much as the price per token.

Sources: ↗ Official OpenAI announcement — GPT-5.6 · ↗ Independent METR evaluation