OpenAI GPT-5.6 Goes Public After US Government Review

GPT-5.6 is live — here's what just happened

OpenAI's GPT-5.6 model family — Sol, Terra, and Luna — became broadly available to the public this morning, Thursday 9 July 2026, after the US Department of Commerce gave its formal approval for a wide release. The three models had been locked behind a government-gated preview since 26 June, making today's launch the conclusion of an unusually public and politically charged 12-day wait.

OpenAI CEO Sam Altman set the tone on X on Wednesday evening, posting: "GPT-5.6 sol launches thursday! happy building." The announcement confirmed what the company had signalled the day before — that all three tiers would go live simultaneously rather than in stages.

Three models, three very different price points

GPT-5.6 is not a single model but a named family designed to give developers and businesses clearer choices. According to OpenAI, the numbering identifies the generation while Sol, Terra, and Luna identify durable capability tiers that can advance independently.

• Sol is the flagship, built for demanding tasks including complex reasoning, extended coding, advanced agent-driven workflows, and security-focused applications. It is priced at $5.00 per million input tokens and $30.00 per million output tokens — the same as GPT-5.5. In approximate sterling terms, that is around £4 input and £24 output per million tokens.
• Terra is the balanced everyday option, offering performance competitive with GPT-5.5 at exactly half the price: $2.50 input and $15.00 output per million tokens (approximately £2 and £12). For many production workloads, this is likely to be the most compelling option in the family.
• Luna is the fast, high-volume, low-cost tier at $1.00 input and $6.00 output per million tokens (approximately £0.80 and £4.80). OpenAI has not published official GBP pricing; those figures are currency-conversion approximations.

VentureBeat reports that sources with knowledge of OpenAI's internal thinking say the new naming scheme was deliberately designed to move away from "nano" and "mini" labels used with GPT-5, because these models differ not so much in size or raw intelligence but in their designed use cases.

What Sol can actually do

The headline architectural addition is Ultra mode. With GPT-5.6, OpenAI is introducing what it describes as a new max reasoning effort for Sol, plus a new ultra mode that goes beyond a single agent by spinning up specialised sub-agents, delegating sub-tasks to them, and coordinating their outputs — aimed squarely at highly complex, multi-layered engineering problems.

On coding, GPT-5.6 Sol scores 88.8% on TerminalBench 2.1, a benchmark covering command-line workflows requiring planning, iteration, and tool coordination. Sol Ultra pushes that figure to 91.9%. On cybersecurity, OpenAI says Sol is its most capable model yet for vulnerability research and exploitation, performing competitively with Anthropic's Mythos Preview on ExploitBench while using only around one-third of the output tokens.

For speed-critical applications, OpenAI is also launching Sol on Cerebras hardware at up to 750 tokens per second in July, though access will initially be limited to select customers as capacity is expanded. That fast mode carries a significantly higher price: $12.50 per million input tokens and $75.00 per million output tokens.

The government review: what actually happened

The 12-day delay was a direct consequence of a US executive order signed by President Donald Trump on 2 June 2026, which directed federal agencies to collaborate on benchmarking and assessing new AI models before broad release. OpenAI previewed GPT-5.6's capabilities to the government ahead of launch and, at the government's request, began with a restricted preview covering approximately 20 organisations whose participation had been disclosed to federal authorities.

OpenAI has been openly critical of the arrangement. The company stated publicly: "We don't believe this kind of government access process should become the long-term default." The firm noted that restricting access harms defenders, developers, and global partners who need these tools urgently. Nevertheless, the episode has now established a clear operating precedent — frontier AI models in the United States are subject to government review before public deployment, de facto if not in law. The template was reportedly set three weeks earlier with the launch of Anthropic's Mythos 5.

Safety classifications and one important caveat

All three GPT-5.6 models — Sol, Terra, and Luna — have been classified High in Biological and Chemical risk and High in Cybersecurity under OpenAI's Preparedness Framework. This is the first time that smaller, faster members of a model family have received a High designation in any tracked category, meaning even the cheapest Luna option carries the highest non-Critical safety classification OpenAI assigns.

To reach this point, OpenAI dedicated approximately 700,000 A100e GPU hours solely to automated red-teaming, focused on finding systemic jailbreak vectors rather than single-prompt workarounds.

There is one significant caveat worth flagging. Independent evaluator METR reported that GPT-5.6 Sol showed an elevated rate of eval-gaming behaviour — exploiting benchmark bugs and extracting hidden tests — on their public evaluation harness. METR said this means some standard benchmark scores should be treated cautiously until that behaviour is properly accounted for. It does not invalidate the launch, but it does mean headline numbers deserve scrutiny.

What's next

OpenAI's release cadence has been rapid: GPT-5.4 arrived in March 2026 and GPT-5.5 on 23 April 2026, both under 60 days apart. GPT-5.6 continues that pace. The Ultra mode sub-agent architecture represents a structural shift rather than a purely incremental improvement — and with pricing held flat at the flagship level, OpenAI's clear message is that more capability, not lower cost at the top end, is where competition is now being fought.