OpenAI Releases GPT-5.6 With Three Tiers and Tool Calling

OpenAI released its GPT-5.6 family of models to general availability this week, following a limited preview period. This new generation introduces a three-tier structure, comprising Sol as the flagship model, Terra as a balanced everyday option, and Luna as the most cost-efficient choice. These models are priced per 1 million tokens, with Sol costing $5 for input and $30 for output, Terra at $2.50 input and $15 output, and Luna at $1 input and $6 output.

Sol has achieved a score of 80 on the Artificial Analysis Coding Agent Index, placing it 2.8 points above Claude Fable 5. The release also features Programmatic Tool Calling, which executes model-written JavaScript within an isolated V8 runtime environment that lacks network access. This capability, when used with the 'ultra' setting that runs four agents in parallel, has improved the Terminal-Bench 2.1 score from 88.8% to 91.9%. However, Sol's performance on SWE-Bench Pro remains a gap, with a score of 64.6%, trailing Claude Mythos 5's 80.3% by approximately 15 points.

GPT-5.6 models are accessible across various surfaces. ChatGPT Plus, Pro, Business, and Enterprise users can access Sol with medium to higher effort settings, and Pro and Enterprise users can also select GPT-5.6 Sol Pro. Free and Go users of ChatGPT Work and Codex can access Terra, while paid users have the choice of all three models and can set effort levels per model. The 'max' setting is available to all users with GPT-5.6 access and can be toggled in settings.

The API now supports all three GPT-5.6 tiers, along with Programmatic Tool Calling and a multi-agent beta within the Responses API. Prompt caching has also been updated, with GPT-5.6 supporting explicit cache breakpoints and a minimum cache life of 30 minutes. Cache writes are billed at 1.25 times the model's uncached input rate, while cache reads benefit from a 90% cached-input discount. In terms of performance, OpenAI reports that Sol achieved a new high of 53.6 on Agents' Last Exam, a benchmark evaluating long-running professional workflows across 55 fields. This score is reported to surpass Claude Fable 5's adaptive reasoning score by 13.1 points.