GPT-5.2 introduces a three-tier intelligence system—Instant, Thinking, and Pro—that automatically matches response complexity to query demands, optimizing both speed and reasoning depth depending on what the task requires. Built as the flagship model for professional knowledge work, it handles the breadth of enterprise needs: spreadsheet creation, presentation building, code writing, image understanding, and the management of long-running, multi-step projects. Its response compaction feature extends usable context well beyond standard limits, making it practical for sustained analytical work where earlier models would lose coherence. The architecture is designed to be economical at scale, with cached inputs reducing costs by up to 90 percent for repetitive workflows like brand guidelines processing.

The model posts substantial gains across well-specified benchmarks: reaching 70.9 percent expert-level performance on GDPval across 44 occupations, scoring 92.4 percent on GPQA Diamond science questions without tools, and hitting a perfect 100 percent on AIME 2025 competition math. Software engineering shows similar momentum—80 percent on SWE-bench Verified and 55.6 percent on SWE-Bench Pro. Abstract reasoning reaches 86.2 percent on ARC-AGI-1 and 52.9 percent on the harder ARC-AGI-2, the latter jumping dramatically from GPT-5.1's 17.6 percent. A specialized variant, GPT-5.2-Codex, extends these capabilities for agentic coding with improvements on long-horizon refactors and migrations, stronger Windows performance, and enhanced defensive cybersecurity. This positions the base model as a practical workhorse for businesses deploying AI across complex knowledge tasks.