Honestly, I don't care for wading through press releases disguised as technical documents, but here we are. This is clearly just marketing fluff celebrating the release of GPT-5.2.
Abstract:
OpenAI announces the release of GPT-5.2, positioning it as their most advanced frontier model for professional knowledge work, available immediately in the API and rolling out to paid ChatGPT tiers. The core claims revolve around significant performance gains over GPT-5.1 across benchmarks measuring knowledge work (GDPval), coding (SWE-Bench Pro), complex reasoning (GPQA, ARC-AGI), long-context understanding (MRCRv2), and vision tasks. Notably, GPT-5.2 Thinking reportedly achieves or exceeds human expert levels on 70.9% of GDPval tasks while being significantly faster and cheaper. Improvements in factuality (30% relative reduction in error rate) and substantially better long-context retention (near 100% accuracy on 4-needle MRCR at 256k tokens) are highlighted as key differentiators for professional application. Safety measures are also reportedly strengthened, particularly concerning mental health-related prompts.
Introducing GPT-5.2: Performance and Availability
16:08 Model Lineup: GPT-5.2 series introduced, including Instant (fast, everyday work), Thinking (deeper work, better reasoning/coding), and Pro (smartest/most trustworthy). Instant and Thinking roll out today to paid ChatGPT users; all are available in the API.
Economic Value Claims: Average ChatGPT Enterprise user saves 40–60 minutes/day. GPT-5.2 designed to unlock more value, showing better performance in spreadsheets, presentations, coding, and long-context tasks.
16:08 Benchmark Superiority (GPT-5.2 Thinking vs. GPT-5):
GDPval (Knowledge Work): 70.9% win/tie rate vs. industry professionals (up from 38.8% for GPT-5).
SWE-Bench Pro (Coding): 55.6% accuracy (up from 50.8%).
GPQA Diamond (Science): 92.4% (up from 88.1%).
ARC-AGI-2: 52.9% (significant leap from 17.6%).
Spreadsheet/Presentation Improvement: GPT-5.2 Thinking scores 9.3% higher than GPT-5.1 on internal junior investment banking analyst spreadsheet modeling tasks (68.4% vs 59.1%).
18:29 Coding Performance: SOTA at 55.6% on SWE-Bench Pro (multi-language, contamination resistant). Reportedly much stronger in front-end development, including complex UI/3D elements.
19:45 Factuality: 30% relative reduction in responses containing at least one error compared to GPT-5.1 Thinking (6.2% error rate vs 8.8%).
20:03 Long Context: Sets SOTA on OpenAI MRCRv2, achieving near 100% accuracy on the 4-needle variant up to 256k tokens.
22:40 Vision: Error rates cut roughly in half on chart reasoning (CharXiv Reasoning) and software interface understanding (ScreenSpot-Pro) compared to GPT-5.1. Shows better spatial understanding of image elements (e.g., motherboard labeling).
23:25 Tool Calling: Achieves 98.7% on Tau2-bench Telecom, indicating reliable multi-turn tool use for complex workflows (e.g., coordinating flight rebooking, compensation, and seating changes).
27:10 Safety: Improved performance in sensitive conversations (suicide, self-harm, mental health distress), with lower undesirable response rates than GPT-5.1 models.
28:20 Availability/Pricing:
ChatGPT: Instant, Thinking, Pro rolling out to paid plans. GPT-5.1 legacy support for three months.
API: GPT-5.2 Thinking available as gpt-5.2. Pricing is higher than GPT-5.1: GPT-5.2 is $1.75/$14 per million input/output tokens (GPT-5.1 was $1.25/$10). GPT-5.2 Pro is significantly higher at $21/$168.
30:05 Partner Acknowledgement: Built in collaboration with NVIDIA (GPUs including H200, GB200-NVL72) and Microsoft (Azure).