GPT-5.2

GPT-5.2 is OpenAI's December 2025 flagship series, positioned for professional knowledge work and long-running agents. It ships in three tiers: Instant (fast, high-volume), Thinking (configurable reasoning effort), and Pro (maximum compute for research-grade tasks) — API IDs gpt-5.2-chat-latest, gpt-5.2, gpt-5.2-pro.

Official API docs confirm: 400K context window, 128K max output, August 31 2025 knowledge cutoff, reasoning effort levels (none/low/medium/high/xhigh), streaming, function calling, and structured outputs. OpenAI reported 80.0% SWE-bench Verified, 55.6% SWE-Bench Pro, 92.4% GPQA Diamond, 100% on AIME 2025, and the first 90%+ score on ARC-AGI-1.

OpenAI's docs now mark GPT-5.2 as the previous flagship (recommending GPT-6 Astra), but it remains live with pinned snapshots (e.g. gpt-5.2-2025-12-11) and is widely deployed across ChatGPT, Azure OpenAI, and GitHub Copilot.

Provider-reported (OpenAI, launch announcement + system card):

• SWE-bench Verified: 80.0% | SWE-Bench Pro: 55.6% (record at launch)
• GPQA Diamond: 92.4% | AIME 2025: 100% | FrontierMath: 40.3%
• ARC-AGI-1 (Verified): first model above 90% | ARC-AGI-2: 54.2% (Pro)
• τ2-bench Telecom: 98.7% (multi-turn tool use) | GDPval: 70.9%

Independent coverage: SiliconANGLE, VentureBeat comparisons, and third-party guides (Augmenter, codelabsacademy) covered the launch and pricing; GitHub added GPT-5.2 to Copilot public preview the same day. No independent re-run of OpenAI's benchmark numbers found.