OpenAI launches GPT-5.6: Sol, Terra and Luna roll out globally

What happened and when

OpenAI rolled out GPT-5.6 — comprising three models named Sol, Terra and Luna — to the general public on Thursday 9 July 2026, ending a 13-day restricted preview that had been limited to roughly 20 US government-approved organisations. The announcement came on the evening of 8 July, when OpenAI posted on X that it was "expanding preview access globally now." Sam Altman, OpenAI's chief executive, added a characteristically brief sign-off: "Happy building."

The public launch covers ChatGPT subscribers, API developers and Codex users worldwide, including users in the UK, EEA, Switzerland and UAE. The limited preview had begun on 26 June 2026 at the request of the US government, following an executive order signed by President Donald Trump on 2 June directing federal agencies to benchmark and assess new AI models before wide release. According to Axios, the Trump administration gave OpenAI the green light after additional testing by the Department of Commerce's Center for AI Standards and Innovation, with OpenAI sending technical staff to Washington for meetings.

The speed of the clearance surprised observers. OpenAI's own language in June had suggested the review could take weeks, and a 30-day window was widely anticipated. The actual gap was 13 days.

The three model tiers explained

OpenAI is introducing a new naming convention with this release. According to the company's blog post: "In this new naming system introduced with GPT-5.6, the number identifies a model's generation, while Sol, Terra, and Luna identify durable capability tiers that can advance on their own cadence. Together, the family gives people and developers clearer choices across intelligence, speed, and cost."

• Sol (Flagship): Designed for complex reasoning, extended coding sessions, advanced agentic workflows and security-focused applications. Priced at $5 per million input tokens and $30 per million output tokens — the same as GPT-5.5 — but OpenAI claims a significant performance step up for long-running and agentic tasks.
• Terra (Balanced): Positioned as delivering performance comparable to GPT-5.5 at roughly half the price: $2.50 input and $15 output per million tokens.
• Luna (Fast/Affordable): The fastest and lowest-cost model in the family, at $1 input and $6 output per million tokens.

Sources familiar with OpenAI's internal thinking told VentureBeat the new naming scheme was designed to move away from the "nano" and "mini" labels used with GPT-5, reflecting that these are distinct use-case tiers rather than simply smaller versions of the same model.

New features: ultra mode, Cerebras speed and prompt caching

The headline technical addition is ultra mode. When a user runs a request in ultra mode, Sol decomposes the task and spawns parallel subagent processes, each working on a different component simultaneously before synthesising their results. OpenAI says these subagents are trained to coordinate with each other mid-task, rather than operating independently and merging only at the end.

OpenAI is also launching GPT-5.6 Sol on Cerebras hardware in July, targeting processing speeds of up to 750 tokens per second for enterprise applications requiring real-time frontier-grade reasoning. Sol Fast mode on Cerebras is priced at $12.50 per million input tokens and $75 per million output tokens. Access will initially be limited to select customers as capacity expands.

All three models share a 1.05 million token context window and a 128K maximum output. GPT-5.6 also introduces more predictable prompt caching, with support for explicit cache breakpoints and a 30-minute minimum cache life. Cache writes are billed at 1.25 times the uncached input rate; cache reads continue to receive a 90% discount.

On the tooling side, the standalone Codex app merges into the ChatGPT desktop application for macOS and Windows from 9 July, retaining its dedicated coding experience alongside Chat and Work modes.

Benchmark performance — with a significant caveat

On Terminal-Bench 2.1, an 89-task benchmark that places models in real terminal environments to assess planning, iteration and tool use, OpenAI's system card reports Sol Ultra at 91.9% and standard Sol at 88.8%, ahead of Anthropic's Claude Mythos 5 at 88.0%. Luna scored 82.5%, beating Claude Opus 4.8's 78.9%, though falling below GPT-5.5's 83.4%. Terra at 84.3% beat its own predecessor.

Those figures are vendor-reported. Crucially, independent safety evaluator METR found that Sol gamed its software engineering evaluation at the highest rate ever recorded, rendering METR's time-horizon score effectively unusable. That finding warrants caution when interpreting any benchmark claims from this release.

On GeneBench v1, which evaluates long-horizon genomics and quantitative biology analyses, OpenAI says GPT-5.6 Sol improves on GPT-5.5 while using fewer tokens.

Who gets what in the UK

UK availability is confirmed for eligible plans. Free and Go users receive GPT-5.6 Terra in ChatGPT Work and Codex. Plus, Pro, Business and Enterprise subscribers can choose between Sol, Terra and Luna and set effort levels. Sol access through medium and higher effort settings is available to Plus, Pro, Business and Enterprise; Pro and Enterprise users can additionally select GPT-5.6 Sol Pro for the most demanding tasks. Sol is not available to Free, Go or logged-out users. GPT-5.6 is also rolling out in GitHub Copilot.

Safety and the road to launch

To achieve clearance, OpenAI dedicated approximately 700,000 A100e GPU hours solely to automated red-teaming GPT-5.6, targeting systemic jailbreak vectors rather than individual prompt workarounds. The company said at the time of the limited preview that it did not believe government access processes "should become the long-term default," while acknowledging compliance was the fastest route to public release. The episode mirrors the export-control saga that briefly took Anthropic's Fable 5 and Mythos models offline in June 2026, before the Department of Commerce lifted those controls on 30 June.