GPT-5.6 jumps in intelligence tests by turning on background reasoning

The model now thinks before it speaks

OpenAI revealed how GPT-5.6 dramatically improved its score in the logical ARC-AGI-3 test. It’s not a new architecture, but two subtle API settings: enabling internal reasoning and allowing context compaction. These settings let the model hold logical steps in memory while solving a problem, without having to generate them all directly as output.

The trick is that the model gets a time and space budget to “think through” the task in a draft space. Previous systems suffered from the fact that once they generated a wrong step, they had to build on it and got tangled in their own mistakes. Compaction also ensures that the model doesn’t overwhelm its own memory with useless filler, but keeps only the essential vectors.

Cheaper performance for application developers

For developers building applications on top of the API, this means a fundamental shift in economics. Until now, increasing reliability meant complex prompting chains, RAG systems or fine-tuning. Now, tweaking call parameters is enough. A model that was previously a fast text generator is becoming a system capable of solving multi-step problems with fewer errors.

Development teams need to realize that this changes the way they should test models. If the model can handle more logical steps on its own, the complexity shifts from orchestrating agents back to calling the base model itself. At the same time, however, it requires more careful monitoring of the costs of the hidden compute that runs under the surface.

A benchmark is not production

A score in ARC-AGI-3 might look great on paper, but the reality of production deployment is more complex. The benchmark measures the ability to find an abstract pattern on a small grid of pixels. That’s an elegant test of intelligence, but it doesn’t match the dirty reality of corporate databases, inconsistent APIs, and changing user requirements.

Furthermore, hidden reasoning is a black box for developers. If the model makes a mistake during its internal deliberation, it cannot be easily debugged or the intermediate step fixed, like with traditional chain-of-thought prompting. The system either returns the correct result or hallucinates with even more confidence than before.

The price of hidden compute will decide

The proof that this setting makes sense outside the lab will be its adoption in mainstream enterprise applications. If the price for these additional internal steps turns out to be too high, this feature will remain the domain of researchers and enthusiasts. Developers will closely watch whether the increase in reliability justifies the higher token consumption under the hood.

Lilith's verdict

OpenAI didn't show a smarter model, but a better way to squeeze out what it already knows. For developers, this is a clear signal that the era of complex agent chains may soon give way to a simple request: “here is some time, just figure it out.”