The results of the latest ARC-AGI-3 benchmark are a cold shower for those who still believe in the magic of pure parameter counts. OpenAI's report on the GPT-5.6 Sol model clearly demonstrates that the era of mindless weight scaling has given way to an infrastructure war. On the standard ARC-AGI-3 public set, the model scored a modest 13.3%; however, once the Responses API—featuring intermediate reasoning persistence and memory compaction—was activated, the score skyrocketed to 38.3%. This brings the system within striking distance of the 48% baseline human level (the RHAE metric), which OpenAI calculated based on gaming logs.
Infrastructure breakthrough over scaling
The gap between versions shows that a raw model today is effectively handicapped without architectural support. While the previous GPT-5.5 hovered helplessly at 0.4% under the same conditions, the Sol version's success in solving ARC-AGI-3’s 2D puzzles is due entirely to intelligent context management. The core of the benchmark requires an agent to calculate unfamiliar rules on the fly without prompts. Traditional setups typically discard a model’s private thought chains after every action, essentially turning the AI into a goldfish with a three-second memory.
With reasoning persistence and memory compaction enabled, GPT-5.6 Sol clears all six difficulty levels, while no other frontier model on the global leaderboard can break past the first.
OpenAI researchers are confirming the obvious: without retaining an internal monologue, a model is forced to analyze the situation from scratch after every move. Using the Responses API allowed this context to become part of the history, while compaction technology solved the sliding window problem where old but critical actions were simply dropped from memory. Ironically, this architectural hygiene didn't just make the model smarter—it reduced output token volume sixfold. Efficiency, it turns out, smells not just of intelligence, but of significant savings on inference costs.
Infrastructure over prompts
For businesses and CTOs, this case is a signal for an immediate strategy shift. It is time to stop investing in "prompt shamans" who tweak word choices and start building sophisticated agentic middleware. GPT-5.6 Sol is already proving the viability of this approach, cracking open math problems like the Double Circuit Conjecture and playing through Pokémon FireRed. This isn't prompt magic; it is the work of external infrastructure capable of managing an agent's long-term memory.