GPT-5.6

What it is

GPT-5.6 is OpenAI’s flagship large language model, released on 9 July 2026 and still the top of the line as of early August 2026, after a limited preview that began on 26 June. That preview was unusual. OpenAI ran it at the request of the US government, giving a small group of trusted partners early access before the general release. The company was careful to stress that it does not want such a government-brokered arrangement to become the default for future launches, presenting gpt-5.6 as an exception rather than a new template for how frontier models reach the public.

The three tiers

The model comes in three tiers, and the naming deserves explaining because OpenAI has changed how it labels its releases. The number, 5.6, marks the generation. The tier names, Sol, Terra and Luna, advance on their own separate cadence, so the version number no longer tries to capture everything in a single figure. In practice a user picks a tier by capability and cost rather than decoding a long string. Sol sits at the top: OpenAI describes it as the workhorse of the family and its best coding model yet. Terra and Luna step down from there for lighter or more price-sensitive work. The effect is a clearer menu, at the cost of asking users to learn a new vocabulary. On 30 July 2026 OpenAI cut the two lower tiers hard, Luna by 80% to $0.20 in and $1.20 out per million tokens and Terra by 20% to $2 and $12, while Sol held at $5 and $30, a move aimed squarely at the price competition from open-weight rivals.

Security and access controls

Positioning for gpt-5.6 is squarely around coding, knowledge work, cybersecurity and science. OpenAI calls it its strongest cybersecurity model so far, and it has wrapped that capability in controls that are unusual for a consumer product. A Trusted Access for Cyber programme restricts the most sensitive work to vetted users engaged in defensive security, a tacit admission that a model skilled at finding software flaws is as useful to an attacker as to a defender. From 1 September 2026, users will need hardware passkeys to keep access to the most cyber-capable tiers. Requiring a physical security key sits closer to the practice of sensitive enterprise software than of an everyday chat assistant, and it shows how OpenAI now thinks about the risk of misuse.

Coding and knowledge work

Beyond security, gpt-5.6 is aimed at knowledge work and science, the tasks where a careful, well-sourced answer earns its keep, and OpenAI has leaned on those strengths to justify the flagship label. The way the model reached the market is part of the story too. A preview run at a government’s request and restricted to trusted partners is a marked change from the open betas of earlier years, and it hints at how much scrutiny the most capable systems now attract before they become widely available.

Coding is the clearest commercial pitch. By naming Sol its best coding model yet, OpenAI is competing directly for developers, the segment where rival labs have pushed hardest this year. gpt-5.6 did not arrive on its own. Alongside it OpenAI shipped ChatGPT Work, an agent product aimed at business tasks; a rebuilt ChatGPT desktop application; hosted sites; and GPT-Live voice models. Taken together, that is a move to carry ChatGPT from a single chat box towards a set of working tools, with the flagship as the engine underneath. Readers tracking that shift can follow it on our large language models hub.

What to watch

The government-brokered preview, the passkey requirement and the gated cyber programme point to a maturing set of expectations around frontier releases: that the strongest models now ship with access controls closer to those of sensitive software than of an ordinary app. Whether gpt-5.6 keeps its lead will depend on independent testing and on how the rival flagships released within days of it hold up in real use. For now it is OpenAI’s headline model and the yardstick against which this summer’s launches are measured, and a useful anchor point on our wider AI coverage.