OpenAI Releases GPT-5.2: Continued Capability Gains Amid Diminishing Marginal Returns
In late November, Google released Gemini 3.0, which surpassed GPT-5.1 across a wide range of benchmarks, posing a serious challenge to OpenAI. In response, CEO Sam Altman reportedly declared a “red alert” internally, mobilizing company-wide resources to accelerate improvements to GPT’s performance.
The result was the release of GPT-5.2 in the first half of December. Similar to GPT-5.1, this was not a disruptive, paradigm-shifting launch. However, in professional and specialized scenarios aimed at replacing or augmenting knowledge workers, GPT-5.2 demonstrates meaningful performance improvements. At the same time, the model’s overall trajectory suggests that it has entered a phase of diminishing marginal returns, where incremental gains require increasingly concentrated effort and resources.
Benchmark Performance: GDPval Emerges as a Highlight
Across multiple public and semi-public benchmarks, GPT-5.2’s overall improvements are not dramatic; however, it performs exceptionally well on GDPval (Gross Domestic Product value modeling), a class of evaluations focused on macroeconomic analysis and strategic reasoning.
The GDPval benchmark measures an AI system’s ability to replace knowledge workers across 44 professions spanning nine GDP-relevant sectors. These sectors represent components of overall economic output and include areas such as real estate, government, manufacturing, and legal services, among others. The dataset comprises 1,320 real-world case descriptions and corresponding solutions written by domain experts, with 30 cases for each profession. AI models generate their own solutions based on the same case descriptions. Both the human- and AI-generated solutions are then evaluated by independent human experts, who are unaware of the source of each response, and are rated as “better,” “equal,” or “worse.”
In the latest evaluation round, GPT-5.2 achieved a double-digit percentage improvement over GPT-5.1 on GDPval and continued to maintain a leading position among comparable models. As shown in the figure below.
According to the evaluation, GPT-5.2 Pro outperformed human experts in 60% of cases, matched human performance in 14.1% of cases, and underperformed in only 25.1% of cases.
On closer reflection, these results are deeply unsettling. AI systems have already surpassed humans in a range of professional, knowledge-intensive domains. As a result, waves of layoffs may soon emerge across multiple GDP-critical sectors that employ large numbers of knowledge workers.
NVIDIA H200 Approved for Export to China: Policy Easing, but a Muted Market Response
Another major development in December was the U.S. government’s decision to allow NVIDIA to sell its H200 AI chips to China. On paper, this represents a clear easing of policy and briefly pushed NVIDIA’s stock price higher.
What is more noteworthy, however, is the reaction from the Chinese market.
China’s Response: Neither Excitement nor Gratitude
Unlike the scramble for H100 chips seen several years ago, the response to H200 in China has been notably restrained:
· Large cloud providers and research institutions remain cautious: Concerns over potential policy reversals have led them to avoid creating new dependencies in critical infrastructure.
· Regulators have kept their distance: Approval and deployment processes have progressed slowly, sending no particularly positive signals.
· Industry sentiment is conservative: Many voices characterize H200 as “usable computing power, but not strategic computing power.”
A commonly cited phrase in Chinese tech circles is: “Being able to buy does not mean one should buy; being able to use does not mean one dares to use.”
At its core, this reaction is not driven purely by performance or pricing considerations, but by broader concerns over long-term availability, supply chain security, and technological self-reliance. Compared with the high level of dependence on NVIDIA GPUs in previous years, China’s AI industry is now more explicitly prioritizing domestic computing resources and system-level optimization as medium- to long-term strategic directions.
As a result, approval of the H200 appears more like a symbolic gesture of détente than a substantive reopening of the market.
The United States Pushes for Nationwide AI Regulation: Clearing the Way for Innovation, or Establishing Control?
In the first half of December, the U.S. government sent a clear signal that it intends to promote a nationwide, unified AI regulatory framework in order to avoid the compliance chaos created by the current patchwork of state-level legislation.
The background to this move is the growing loss of control at the state level. Over the past year, multiple U.S. states—particularly California—have rapidly enacted legislation addressing issues such as AI safety, copyright, and liability, creating significant unease among large AI companies. For businesses, complying with 50 different sets of rules translates into extremely high compliance costs.
For example, in late September, California passed the Frontier Artificial Intelligence Transparency Act, which requires AI companies to provide detailed disclosures related to catastrophic risk assessments, safety measures, and societal harms. The law also establishes internal whistleblower protection mechanisms within AI laboratories, encouraging employees to report major safety risks or incidents.
The core objectives of unified federal regulation are clear:
· Prevent a fragmented “regulatory patchwork” from hindering the development of the AI industry
· Establish AI as a strategic priority at the national level
Divisions across stakeholders are equally pronounced
· Technology companies largely support the initiative: unified and predictable rules reduce uncertainty and compliance burdens.
· Some state governments and civil rights organizations oppose it: they worry that federal standards may “race to the bottom,” weakening protections for consumers and the public.
At its core, this contest is not about whether AI should be regulated, but about who gets to define the boundaries of AI—and where those boundaries should be drawn.
From a global perspective, the U.S. move also carries a clear element of international competition. Positioned between the European Union’s high-intensity regulatory regime and China’s state-led model, the United States appears to be attempting to chart a third path: one that is nationally unified yet relatively permissive.
Summary: Models Are Evolving, Compute Is Diverging, and Rules Are Consolidating
Taken together, the three major threads of the first half of December point to a single underlying reality:
· Model capabilities continue to improve, but they are no longer the sole determining factor
· Computing power is being repriced and redistributed through the lens of geopolitics
· Regulation is moving from a phase of discussion to one of framework-setting
The AI industry is transitioning from a period of unchecked expansion to a stage defined by constraints, selection, and strategic management. In the competition ahead, the decisive factor may no longer be whose model is 5% better, but who can achieve the most stable balance across technology, supply chains, and institutional frameworks.
