OpenAI releases GPT-5.6 worldwide – with three variants, new Codex app and naming confusion

Three variants with a new naming scheme

GPT-5.6 is available in ChatGPT, Codex and via the OpenAI API. GPT-5.6 Sol is available to Plus, Pro, Business and Enterprise users. According to the company, the rollout is to take place worldwide over the next 24 hours.

Also new is the 'Ultra' level for the Sol model, which coordinates four agents in parallel to complete particularly complex tasks more quickly. All three GPT-5.6 tiers introduce a new mode called 'max reasoning effort', which gives the model more time on difficult problems.

Availability and pricing

Terra costs $2.50 and $15 respectively. This matches the terms already announced in June. For GPT-5.6, OpenAI is also introducing cache write pricing at 1.25x the input price, while cache reads continue to be billed at a 90 percent discount – a structure also used by Anthropic.

The naming scheme is reminiscent of Anthropic's variant designations Opus, Sonnet and Haiku. The company says Sol also offers advantages in estimated costs at medium compute depth. Sol uses fewer tokens than previous models and is therefore expected to provide a better performance-to-cost ratio.

According to the company, Sol is to achieve new best values in areas such as programming, knowledge work, IT security and science. In programming, OpenAI pointed to the Artificial Analysis Coding Agent Index, in which Sol achieves a score of 80 at maximum reasoning depth. On the Terminal-Bench 2.1 coding benchmark, Sol Ultra is said to come in at 91.9 percent, also ahead of Fable 5, which reaches 88 percent.

Benchmarks and performance comparison

On so-called agentic last-exam tests, which map longer workflows across 55 subject domains, Sol is reported to have achieved a score of 53.6 points. OpenAI compared this to the Claude Fable 5 model, which it says scored 13.1 points lower. According to the company, this is achieved with fewer than half as many output tokens as comparable models.

On the Artificial Analysis Intelligence Index, however, Claude Fable 5 (with fallback) edges out GPT-5.6 Sol (max) with 60 points to 59, followed by Grok 4.5 (high) with 54, GLM-5.2 (max) with 51 and Gemini 3.5 Flash with 50 points. On the AA Briefcase benchmark for realistic knowledge work, Fable 5 is also ahead of Sol, although Sol achieves the highest 'Presentation Elo' ever recorded – an indication of particularly well-designed outputs in tools such as PowerPoint and Excel.

In addition, GPT-5.6 is said to be capable of independently writing and running small programs to coordinate tools and process intermediate results. OpenAI also promises improvements for office applications such as presentations, documents and spreadsheets. The model is said to adhere more reliably to templates and reference formats than its predecessor GPT-5.5, for example with regard to layout, typography and recurring design elements.

Safety and protective measures

OpenAI describes Sol as particularly strong in programming, biology and cybersecurity. In cybersecurity, the model is said to be more capable of finding and fixing vulnerabilities than of independently carrying out full attacks on well-protected systems. OpenAI says it introduced additional safeguards for the cybersecurity and biology domains.

Compared with previous models, the protective measures in GPT-5.6 Sol are said to block roughly ten times as many potentially harmful activities. GPT-5.6 does not exceed the internally defined critical threshold in either category.

The Codex app becomes the new central ChatGPT desktop app, bringing Chat, Work and Codex together under one roof. However, it still largely looks like Codex and includes the Codex icon as a selectable option. The previous ChatGPT app will from now on be called ChatGPT Classic. On Hacker News, users expressed confusion: users ask where to simply chat now, and complain that switching between ChatGPT Work and ChatGPT Codex seemingly does nothing.

Codex app as the new desktop hub

On the knowledge benchmark AA-Omniscience, GPT-5.6 shows only a slight improvement over its predecessor GPT-5.5, and the slightly higher hit rate is accompanied by a somewhat increased hallucination rate. The figures come from tests selected by OpenAI under self-defined conditions and can only be independently verified to a limited extent.

Prior to its broad release, GPT-5.6 was only available for several weeks in a heavily restricted preview. GPT-5.6 was the first leading American AI model to undergo a formal government review before general release. The review was conducted through the Center for AI Standards and Innovation at the Department of Commerce; according to Axios, OpenAI sent specialists to Washington to answer the agency's questions.

Government pre-release review in the US

Commerce Secretary Howard Lutnick, Treasury Secretary Scott Bessent and National Cyber Director Sean Cairncross were involved in the proceedings. The underlying framework was introduced by the Trump administration on June 2, 2026 and provides for a voluntary pre-release review of the most capable models by executive order, asking AI firms to give the government up to 30 days of pre-release access to models with advanced cyber capabilities.

For GPT-5.6, the voluntary review escalated into a government-managed access list, and the model was accessible for weeks only to roughly 20 partners whose names were individually approved by the government. OpenAI has made clear that it views this precedent with discomfort. The company participated this time, but does not consider such a government access procedure to be an appropriate permanent state of affairs.