OpenAI Launches GPT-5.6 With Three Models: Sol, Terra and Luna

OpenAI rolled out general availability of the GPT-5.6 model family on July 9, introducing an entirely new naming system along the way. Instead of the familiar mini and nano suffixes, the company introduced three proper names, Sol, Terra and Luna, which will now denote fixed capability tiers that develop independently of the model's generation number.

Three models, three use cases

Sol is built for the toughest tasks, such as complex programming and security research. Terra is meant to handle large-scale business workloads, including customer service, internal tools and document analysis. Luna was designed as a fast, inexpensive option for everyday work like summarizing text, drafting content and routine automation.

OpenAI says the new naming scheme is meant to bring order to what had become a confusing version history. The number in a model's name, such as 5.6, indicates its generation, while the names Sol, Terra and Luna are meant to identify durable capability tiers that the company can update at different paces, independently of one another.

New tool-calling feature

Alongside GPT-5.6, OpenAI introduced programmatic tool calling, a feature that lets the model run its own hand-written JavaScript code inside an isolated V8 sandbox with no network access. According to figures the company shared, the feature cut token usage by as much as 63.5 percent for select customers.

The model also comes with an updated prompt caching mechanism, featuring explicit cache checkpoints, a minimum cache lifetime of 30 minutes, and discounts of up to 90 percent on cache reads.

Performance against rivals

On the AA Coding Agent Index benchmark, Sol scored 80 points, overtaking the previous leader, Claude Fable 5, which scored 77.2 points. On the Agents' Last Exam test, Sol scored 52.7 percent, also ahead of GPT-5.5's result of 46.9 percent.

The new model does not dominate everywhere, though. On the SWE-Bench Pro benchmark, which measures coding skills in real-world scenarios, Sol scored 64.6 percent, about 15 points behind rival model Claude Mythos 5, which reached 80.3 percent. The cheapest tier, Luna, performs worse on very long context work, scoring 41.3 percent on the MRCR v2 test for the 256,000-to-1,000,000-token range.

Safety and availability

OpenAI says Sol launched with the most extensive safety stack in the company's history. According to the company's own description, teams spent several weeks hunting for weaknesses and stress-testing the system's resilience against real-world attacks, strengthening defenses against high-risk activity, sensitive cybersecurity queries, and repeated abuse.

For Polish companies using the OpenAI API, weighing cost against capability will be key. The three-tier pricing structure, ranging from $1 per million input tokens on the Luna tier to $30 per million output tokens on the Sol tier, lets businesses match spending to project scale, which could matter for vendor selection after a string of reports about rising AI bills.

The GPT-5.6 launch fits into OpenAI's accelerating release cadence, coming just a day after the company shipped its GPT-Live voice model. Competition with Anthropic's Claude Fable 5 and Mythos 5 models, as well as with xAI's Grok 4.5, remains tight, with each company leading on a different set of benchmarks.

Sources: OpenAI Developer Community (community.openai.com), MarkTechPost (marktechpost.com), VentureBeat (venturebeat.com), TechCrunch (techcrunch.com)