OpenAI details GPT-5.6 efficiency and how two settings tripled ARC-AGI-3 scores

OpenAI's GPT-5.6 messaging leans hard into efficiency: the pitch is delivering more useful intelligence per dollar across models, inference and agentic workflows, rather than raw capability gains alone. That framing mirrors the week's industry-wide pivot — Anthropic's cheaper Opus 5, Google's token-efficient Gemini 3.6 Flash, DeepSeek's V4 Flash — toward cost-per-task as the competitive axis now that top-line intelligence has plateaued in perceived usefulness.

The technical highlight is a companion post showing that two API settings — retaining reasoning across turns and enabling compaction — tripled GPT-5.6's performance on the ARC-AGI-3 benchmark while also improving efficiency. The result is notable because it locates gains in how agents manage context and reasoning state across a session, not in the base model weights: persistence and compaction let the model carry forward useful reasoning without paying to regenerate it, a lever developers can pull today.

Competitive context: ARC-AGI has been a closely watched proxy for genuine reasoning, so a 3x jump from configuration alone will draw both interest and scrutiny — critics will ask whether the settings amount to benchmark-specific tuning versus a generalizable capability. It also reflects a broader theme in developer discussions that AI progress is increasingly about harness, memory and context management (echoing Nadella's 'model system' framing) rather than the underlying model.

What to watch: independent replication of the ARC-AGI-3 gains under the documented settings, and whether the efficiency claims hold in real agentic workloads. GPT-5.6 Sol is also the model family implicated in the Hugging Face breach evaluation, so its capabilities are under unusual public scrutiny this week. The efficiency-first narrative is OpenAI's answer to a market that has grown price-sensitive.