AI
OpenAI launches GPT-5.6 in three tiers, competing on cost and speed
The company rolled out Sol, Terra, and Luna models globally, emphasizing token efficiency and benchmark performance against rivals like Anthropic's Fable 5.
The New Stack
OpenAI has released its GPT-5.6 family of models to app and API users around the world, introducing a three-tier lineup designed to serve different performance and budget needs. The release marks the first time the company has simultaneously launched distinct variants of a new flagship model.
The three models#
The new lineup comprises Sol, a high-end model built to compete with Anthropic's Fable 5; Terra, positioned for mainstream use; and Luna, a lighter option prioritizing speed and affordability over raw capability.
Pricing reflects this hierarchy. Sol costs $5 per million input tokens and $30 per million output tokens, while Terra runs $2.50/$15, and Luna $1/$6 on the API. Access tiers vary across platforms: free ChatGPT users get Terra only, while Plus and higher subscribers can select among all three and adjust reasoning effort levels.
Efficiency as the main selling point#
Rather than simply claiming superiority, OpenAI is framing GPT-5.6 around token efficiency—getting more useful output from fewer tokens. The company notes that Sol matches or exceeds Fable 5's performance on the Artificial Analysis Intelligence Index while costing half as much and running in slightly more than half the time.
This efficiency advantage extends to specialized tasks. On the Agents' Last Exam benchmark, which measures long-horizon agentic workflows, Sol outpaces Fable 5 by 13.1 points. Coding benchmarks show similar patterns, with Sol posting stronger scores on Terminal-Bench 2.1 and DeepSWE 1.1 at meaningfully lower cost.
New capabilities for complex tasks#
OpenAI added built-in features allowing GPT-5.6 models to write and execute lightweight programs internally, coordinating tools and processing intermediate results without constant user input. This enables "tool-heavy tasks" to progress with fewer round-trips and less guidance. A new "ultra" mode, comparable to Anthropic's parallel-agent approach, dispatches multiple agents simultaneously for demanding jobs.
In knowledge work, Sol scored 92.3% on the BrowseComp agentic browsing benchmark and 62.6% on OSWorld 2.0. Notably, the model has improved at inferring design systems from templates and applying them consistently when generating presentations and documents.
