Just two settings are enough to triple GPT-5.6 Sol's score on ARC-AGI-3.

OpenAI's GPT-5.6 Sol goes from a 13.3% to a 38.3% success rate thanks to two technical adjustments. The impact of context management on performance.

A model's performance does not depend solely on its capabilities. OpenAI provides a new illustration of this with ARC-AGI-3, a benchmark composed of unfamiliar 2D games in which the agent must progressively understand the rules based on its actions and the responses displayed on the screen.

With the benchmark's official setup, GPT-5.6 Sol scores 13.3% on the public set. By using the Responses API with two settings already employed in ChatGPT and Codex—maintaining reasoning between actions and context compaction—the score reaches 38.3%, while dividing the number of generated tokens by six.

In the initial configuration, each new action erased the previous reasoning messages. The model retained the history of its moves but lost the analyses and strategies that had motivated them. When the context became too long, the oldest elements were also deleted, regularly forcing the agent to re-interpret the game.

Retaining reasoning allows the model to pursue a strategy without restarting its analysis from scratch. Meanwhile, compaction replaces the progressive deletion of history with a summary of useful information, in order to preserve the understood rules, the attempts made, and the errors already encountered.

The score used, named RHAE, compares the efficiency of the model's actions to that of a human. OpenAI estimates that humans who tested these games achieve an average of 48%. GPT-5.6 Sol therefore remains below this level, but the gap between the two configurations serves as a reminder that a benchmark also evaluates context management, API parameters, and the setup connecting the model to its environment.