OpenAI GPT-5.6 goes fully live across ChatGPT and API
After a government-gated limited preview, OpenAI's most capable model family is now generally available worldwide — bringing three capability tiers, new voice models, and a merged Codex app.
By Dan Coveney · Founder & Lead Reviewer · Updated July 11, 2026What you need to know
- GPT-5.6 went generally available on 9 July 2026 across ChatGPT, ChatGPT Work, Codex and the OpenAI API
- Three tiers — Sol (flagship), Terra (balanced), Luna (fastest/cheapest) — share a 1.05M-token context window and a February 2026 knowledge cutoff
- A two-week government-gated preview was triggered because Sol cleared OpenAI's "High" cybersecurity and biology risk thresholds simultaneously — a first for any OpenAI model
Generally available from 9 July
OpenAI made GPT-5.6 generally available on Wednesday 9 July 2026, rolling the model out across ChatGPT, ChatGPT Work, Codex and the OpenAI API, with the company saying the global rollout would complete within 24 hours. The release followed a limited preview that had been running since 26 June — one that, unusually, required sign-off from the US government before it could widen.
CEO Sam Altman marked the moment with characteristic brevity, posting on X late on Tuesday evening: "Happy building."
Three tiers, one generation
GPT-5.6 introduces a new naming system in which the number identifies a model generation and a second name — Sol, Terra, or Luna — identifies a durable capability tier that can advance on its own cadence. All three tiers share a 1.05-million-token API context window, a 128K maximum output, and a knowledge cutoff of 16 February 2026.
- Sol — the flagship, tuned specifically for work in biology, chemistry and cybersecurity. It is also available in a stronger Sol Ultra variant. In the API, the generic
gpt-5.6identifier routes to Sol by default. - Terra — the balanced option, offering performance competitive with GPT-5.5 at half the cost.
- Luna — the fastest and cheapest, designed to bring strong AI capability at the lowest price point in the lineup.
API pricing is set in US dollars: Sol costs $5 per million input tokens and $30 per million output tokens; Terra is $2.50/$15; and Luna is $1/$6. Based on current exchange rates, a UK tech review site estimates those at roughly £4/£24, £2/£12 and £0.80/£4.80 per million input/output tokens respectively — though OpenAI has not published official sterling API pricing, and UK buyers are advised to check the checkout price on openai.com.
For ChatGPT subscribers, Plus, Pro, Business and Enterprise users can access GPT-5.6 Sol through medium and higher effort settings. Pro and Enterprise users can also select Sol Pro. OpenAI lists the Plus plan at $20 per month; UK prices are not set in GBP by the company, so the sterling figure including VAT will vary.
The release 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× the uncached input rate, while cache reads receive a 90% discount. One important pricing note for developers: prompts exceeding 272,000 input tokens attract a long-context surcharge, billed at 2× the input rate and 1.5× the output rate for the full request.
Codex merges into the desktop app
Alongside the model launch, OpenAI confirmed that Codex — its dedicated coding environment — has merged with the ChatGPT desktop app on macOS and Windows. The company says Codex remains a distinct coding experience, but it now sits alongside Chat and Work as a mode within the same application rather than a separate product.
OpenAI also announced a new generation of voice models called GPT-Live. According to OpenAI, the models can listen and speak simultaneously, with the company describing the experience as feeling "much more like having a real conversation." Unlike previous voice implementations, OpenAI says GPT-Live "continuously processes input while generating output" rather than handling discrete sequential messages.
Benchmark headlines
OpenAI's published benchmark tables show Sol Ultra scoring 91.9% on Terminal-Bench 2.1, with Sol at 88.8%, Terra at 84.3% and Luna at 82.5%. For context, GPT-5.5 scores 83.4% on the same test; Anthropic's Claude Mythos 5 scores 88.0% and Claude Opus 4.8 scores 78.9%.
On Agents' Last Exam — an evaluation covering long-running professional workflows across 55 fields — GPT-5.6 Sol sets a new high of 53.6, which OpenAI says eclipses Claude Fable 5 by 13.1 points. Sol also claims a 54% improvement in token efficiency on agentic coding tasks and the strongest reported biology scores of any publicly available model, including 68.4% on the Human Pathogen Capabilities Test, nine percentage points above GPT-5.5.
OpenAI's own tables do show gaps, however. Claude Mythos 5 leads Sol on SWE-Bench Pro and ExploitBench, while Claude Fable 5 leads on GDPval-AA v2, FrontierMath Tier 4, Toolathlon, HealthBench Professional and the Artificial Analysis Intelligence Index v4.1, in each case by a small margin.
The government detour
The limited preview that preceded this launch was not a standard staggered rollout. GPT-5.6 is the first OpenAI model family rated High under the company's Preparedness Framework in both biology and cybersecurity simultaneously — and it is also the first time a budget tier, Luna, has cleared that same High cybersecurity threshold. That combination triggered direct intervention from Washington.
The day before the 26 June preview began, the Office of the National Cyber Director and the Office of Science and Technology Policy formally asked OpenAI to restrict access to a small group of government-approved partners. Commerce Secretary Howard Lutnick, Treasury Secretary Scott Bessent and National Cyber Director Sean Cairncross were involved in what OpenAI described as a "collaborative back and forth." CEO Sam Altman told employees that the government would be "approving access customer by customer" during the preview.
OpenAI was clear about its position: "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."
According to Axios, the Trump administration cleared the wider release after additional testing conducted by the Department of Commerce's Center for AI Standards and Innovation, with OpenAI sending technical staff to Washington to address questions in real time. The episode follows a similar pattern with Anthropic, whose Claude Fable 5 and Mythos 5 models were temporarily restricted under an export control directive before the Department of Commerce lifted that order late last month.
The precedent is worth watching. The Trump administration's AI executive order, signed in June, asks developers to voluntarily share frontier models with the government for capability assessment before full release and gave federal agencies 60 days to develop an evaluation process. Wednesday's launch suggests that process, however informal, is now operational — and that future model releases from any major lab could face a similar holding pattern.
Why it matters
For UK subscribers and developers, GPT-5.6 represents the most significant capability jump since GPT-5.5, with the Terra tier offering comparable performance at half the cost — a meaningful saving for startups and agencies building on the API. The government-gating episode is also a signal that Washington now considers itself a gatekeeper for frontier AI releases, which could affect the timing of future model launches and access for non-US partners. UK buyers should note that OpenAI has not published official sterling pricing; the USD rates convert to roughly £4/£24 per million tokens for Sol and £0.80/£4.80 for Luna, but checkout prices will vary.