OpenAI GPT-5.6 Three Models Officially Released: Sol, Terra, Luna Launched Simultaneously

On July 9, 2026, OpenAI announced the official release of the GPT-5.6 series models, along with the launch of ChatGPT Work agents. The series includes three models: Sol, Terra, and Luna, with Sol as the flagship, Terra as the balanced option for daily work, and Luna as the low-cost version. Sol scored 80 on the Artificial Analysis Coding Agent Index v1.1, leading Claude Fable 5 by 2.8 points; scored 53.6 on the Agents’ Last Exam, surpassing Claude Fable 5 by 13.1 points; and achieved 64.6% on SWE-Bench Pro, trailing Claude Mythos 5 by about 15 points.

Model Tiers and Access Permissions

The GPT-5.6 series adopts a three-tier pricing and access design. Sol costs $5 per million input tokens and $30 for output; Terra costs $2.5 for input and $15 for output; Luna costs $1 for input and $6 for output. Plus and Business users can access Sol at medium or higher reasoning intensity, while Pro and Enterprise users additionally get the Sol Pro option. Free and Go users can access Terra via ChatGPT Work, while paid users can choose all three models in both ChatGPT Work and Codex.

ChatGPT Work includes Codex, supporting multi-step tasks across web, mobile, and desktop applications, running independently in the background and accessing local files or the built-in browser. The plugin system connects to external applications, and the Sites beta feature can generate interactive reports and dashboards. Programmatic Tool Calling runs model-generated JavaScript code in the Responses API, using an isolated V8 environment with no network access.

Performance Benchmarks and Real-World Gaps

Official benchmarks show Sol achieving 88.8% on Terminal-Bench 2.1, 72.7% on DeepSWE v1.1, 92.2% on BrowseComp, and 62.6% on OSWorld 2.0. Compared with Claude Opus 4.8, Sol surpasses it on OSWorld while using 85% fewer output tokens. The Agents’ Last Exam covers 55 specialized domains, with Sol scoring 53.6, higher than Claude Fable 5’s 40.5.

In multi-agent parallel mode, the ultra configuration boosts Terminal-Bench 2.1 from 88.8% to 91.9%. Prompt caching supports explicit breakpoints and a minimum cache lifetime of 30 minutes. Cache writes are billed at 1.25x the model input price, while reads receive a 90% discount.

Impact on Developers and Enterprises

Developers can directly call Programmatic Tool Calling and the multi-agent beta feature through the Responses API, enabling parallel agent execution without additional deployment. Enterprise users under Pro and Enterprise plans gain access to Sol Pro and higher reasoning intensity, suitable for code review, threat modeling, and long-duration workflows. Free users remain at the GPT-5.5 level and need a subscription for GPT-5.6 access.

Compared with Anthropic’s Claude Mythos 5 and Fable 5, GPT-5.6 Sol leads in coding agent benchmarks but trails by about 15 points on SWE-Bench Pro. The Claude series retains advantages in some specialized workflow benchmarks, while OpenAI differentiates itself in token consumption and cost.

Follow-Up Signals to Monitor

Developers can track the call success rate and cache hit ratio changes in the Responses API; enterprises can observe token consumption and error rates of Sol Pro in long-duration workflows.