GPT-5.2 launches with major benchmark gains; Oracle delays OpenAI data centers to 2028
Dec 12, 2025
Key Points
- OpenAI's GPT-5.2 posts major benchmark gains on reasoning tasks, but the improvements hinge on maximum inference-time effort rather than model architecture alone, a standard practice across all labs that complicates performance comparisons.
- Reasoning models now dominate hard abstract reasoning problems, jumping from 12% to 90% on ARC-AGI, but add computational latency that reshapes production cost-benefit calculations.
- Oracle delays OpenAI data center completion from 2027 to 2028 citing supply chain constraints, affecting the $300 billion contract signed this summer.
Summary
OpenAI's GPT-5.2 achieves significant benchmark gains across reasoning tasks, but the gains depend heavily on computational effort rather than architectural innovation alone. GPT-5.2 thinking posts jumps on SWE-Bench, GPQA Diamond, Frontier Math, and AIME, winning nearly all benchmarks except advanced mathematics, where Gemini 3 Pro leads. OpenAI disclosed that its models ran with maximum available reasoning effort. Anthropic and Google use the same practice, making peak reasoning settings now standard across all labs, though the disclosure framing raised fairness questions about result presentation.
Greg Brockman, OpenAI's president, called reasoning models systemically underrated. On ARC-AGI, a long-standing abstract reasoning test that humans solve easily but AI has struggled with for years, GPT-5.2 non-reasoning achieves only 12 percent while reasoning variants jump to 90 percent. That gap marks the current frontier: raw scale alone no longer moves performance on hard reasoning problems. Reasoning models add computational time during inference, a trade-off that reshapes how companies think about cost and latency in production systems.
Oracle pushed back completion dates for OpenAI data centers from 2027 to 2028, citing labor and material shortages. The delay affects a $300 billion contract signed this summer. Oracle stated that no delays apply to contractually required sites and all milestones remain on track, suggesting the pushed-back facilities are expansion projects beyond minimum obligations. The delays fit supply chain realities, though Oracle made no mention on an earnings call two days before the announcement, raising questions about how the company manages investor expectations around its AI infrastructure bet.