On 9 July 2026, OpenAI announced the general availability of GPT-5.6, its new family of frontier models. Three variants, three names evoking the solar system: Sol, the flagship; Terra, the balanced model for everyday work; Luna, the fastest and most cost-efficient.
All three are already available on AIDeskPro.
Three models, three levels of capability
The logic behind GPT-5.6 is straightforward: not a single model for everything, but a scalable family where you choose the right level for the task at hand.
GPT-5.6 Sol is the frontier model. OpenAI reports that on Agents’ Last Exam - the benchmark measuring professional agentic workflows across 55 domains - Sol reaches 53.6 points, surpassing Claude Fable 5 by over 13 points. It does so using fewer tokens, with lower latency and at an estimated lower cost. On coding, it sets new standards: in the Artificial Analysis Coding Agent Index it scores 80, 2.8 points above Fable 5, using less than half the output tokens at roughly one third of the cost.
GPT-5.6 Terra is the model for those who want competitive performance without paying the flagship price. OpenAI positions it just above GPT-5.5 across major benchmarks, at roughly one quarter of the cost. For everyday business tasks - document analysis, summarisation, drafting, team assistance - it is likely the ideal balance point.
GPT-5.6 Luna is the fast, economical model. Surprisingly, it outperforms Claude Opus 4.8 on several coding and computer use benchmarks, at roughly one sixteenth of the cost of Fable 5. For repetitive tasks, automated pipelines or contexts where response speed matters more than depth of reasoning, Luna is the right choice.
A leap in design judgment and knowledge work
What stands out about GPT-5.6 is not just the benchmarks. It is the ability to see work through to completion without stopping halfway.
OpenAI describes a model that inspects its own rendered output, identifies visual and functional issues, and applies final touches before returning the result. On BrowseComp - the benchmark measuring agentic web navigation - Sol reaches 92.2%. On OSWorld 2.0, which measures autonomous computer use, it scores 62.6%, surpassing Claude Opus 4.8 with 85% fewer output tokens.
For professional knowledge work - presentations, documents, spreadsheets, financial analysis - GPT-5.6 shows a markedly superior ability to follow reference templates and styles compared to previous models. Anyone working with recurring corporate formats will notice the difference.
What this means for AIDeskPro users
AIDeskPro integrates the best models available on the market and updates them as soon as they are ready for production use. GPT-5.6 Sol, Terra and Luna are available today in the model selection menu.
A note on data residency: GPT-5.6 Sol runs on global endpoints. As always, AIDeskPro makes this explicit: you can choose the right model based on the sensitivity of the data you are working with. For more sensitive data, models on European infrastructure remain the recommended choice.
Pricing
OpenAI has published official pricing per million tokens:
- Sol: $5 input / $30 output
- Terra: $2.50 input / $15 output
- Luna: $1 input / $6 output
Terra and Luna in particular offer a quality-to-price ratio that OpenAI has never offered before at these performance levels.
Source: GPT-5.6: frontier intelligence that scales with your ambition - OpenAI , 9 July 2026.