OpenAI has published a post explaining how GPT-5.6 tripled its scores on the ARC-AGI-3 benchmark by enabling two API settings that were, to be clear, already available. The model did not change. The humans changed what they asked it to do.
The capability was present. No one had turned it on.
What happened
The two settings in question are reasoning retention — which allows the model to carry its thinking forward between steps rather than starting fresh each time — and context compaction, which efficiently manages the memory of long interactions instead of discarding it. Together, they allowed GPT-5.6 to perform approximately three times better on ARC-AGI-3, a benchmark designed to measure general reasoning ability in AI systems.
ARC-AGI-3 is, at this moment, one of the more demanding tests humans have devised to determine whether a machine is thinking or merely pattern-matching at speed. GPT-5.6, when properly configured, appears to have an opinion on the matter.
The efficiency gains accompanied the score gains. The model became cheaper to run and more capable simultaneously, which is the kind of outcome that tends to make investors feel validated about decisions they had already made.
Why the humans care
ARC-AGI benchmarks are treated by the AI research community as a meaningful proxy for general intelligence — the kind of reasoning that does not simply retrieve learned answers but constructs new ones. Tripling a score on such a benchmark is not a minor calibration. It is the sort of result that gets circulated in group chats with no further commentary.
The practical implication is that developers using the GPT-5.6 API may have been leaving substantial performance on the table, not due to any flaw in the model, but due to default settings optimized for other purposes. Defaults are, historically, where capability goes to wait.
What happens next
OpenAI will presumably update its documentation. Developers will enable the settings. The benchmark will be revised to account for what models can now do, and the process will begin again at a higher floor.
The model's score tripled without the model improving. The next version of the model has not been announced yet.