GPT-5.6 Hits Its Highest ARC-AGI-3 Score — the Gains Came from Two Toggles
OpenAI tripled GPT-5.6‘s score on the ARC-AGI-3 benchmark by enabling two API configuration settings. The lab announced the result on July 29.
It’s the highest score the model has posted on the test — and it came without any architectural changes to the underlying system.
Also Read: AI Agents Versus Human Traders: Nansen’s CEO Says the Contest Ends in Two Years
The improvement came entirely from two adjustments: retaining reasoning context across attempts, and compacting prior outputs before each new pass.
What The ARC-AGI-3 Benchmark Actually Tests
The ARC-AGI-3 benchmark is designed to measure a model’s ability to solve novel visual reasoning puzzles that require genuine generalization rather than pattern memorization. OpenAI announced the result after enabling the two settings through the existing API.
Each puzzle presents a grid of colored cells. The model must infer the transformation rule from a handful of examples and then apply it correctly to a new input.
The test was specifically created to resist the kind of statistical shortcut-taking that large language models excel at.
A model that has memorized enough internet text cannot simply recall the answer. It must figure out the rule from scratch each time.
That makes score gains on ARC-AGI-3 meaningful in a way that gains on older benchmarks, like MMLU or HumanEval, often are not.
Those tests have been absorbed into training data across the industry. ARC-AGI-3 is harder to game.
The 2 Settings That Changed Everything
The first setting is reasoning retention.
Normally, each inference pass on a multi-step problem starts fresh. With reasoning retention enabled, GPT-5.6 keeps its chain of thought alive between attempts, allowing it to build iteratively on what it has already worked out rather than restarting from the prompt.
The second setting is output compaction.
Before each new reasoning pass, the model compresses prior outputs into a dense summary. This prevents the context window from filling with redundant material and keeps the relevant signal-to-noise ratio high as the problem unfolds over multiple steps.
Together, the two settings change how the model allocates its finite attention across a reasoning chain.
Retaining the chain gives the model memory within a problem. Compacting prior outputs gives it room to keep thinking without hitting token limits.
OpenAI said the combination boosted both raw scores and efficiency.
The model produces more useful reasoning per unit of compute when both settings are active simultaneously.
Why A 3x Jump On This Test Is Significant
A tripling of score on the ARC-AGI-3 benchmark is not a minor refinement. It represents the difference between a model that occasionally gets lucky on easy puzzles and one that can work through hard ones in a structured, self-improving loop.
The result also demonstrates something more commercially relevant. Neither setting requires a new model release, a fine-tuning run, or an architecture change. Both are available through the existing API.
Any developer or enterprise already paying for GPT-5.6 access can enable them today. That lowers the cost of the capability jump to essentially zero for existing customers.
The gain also carries implications for agentic workflows, where models are asked to complete long tasks autonomously.
Reasoning retention and compaction are precisely the capabilities that degrade fastest when an AI agent encounters a multi-step problem. A model that can hold its own reasoning together across attempts is more reliable as an autonomous worker.
The Score In The Broader Frontier Race
OpenAI published the ARC-AGI-3 result as part of a broader push to demonstrate that GPT-5.6 delivers frontier intelligence alongside frontier efficiency.
The lab has positioned the model as superior not just in raw capability but in cost-per-useful-output, a metric that matters as AI compute spending becomes a boardroom conversation.
Benchmark competition among frontier labs has intensified through this year. Google DeepMind, Anthropic, and Chinese developers including Moonshot AI have all posted competitive scores on reasoning tests.
A 3x jump from two configuration settings, rather than a new model, gives OpenAI a narrative that does not require disclosing training costs or model architecture.
The ARC-AGI-3 benchmark was created by François Chollet, the machine learning researcher who also created the Keras deep-learning framework. Chollet has argued that ARC-AGI tests measure something closer to fluid intelligence than conventional benchmarks do.
OpenAI’s gain on his test is therefore legible as a claim on a specific and contested kind of progress.
Read Next: White House in Panic Mode Over Shocking AI Breach at OpenAI
