According to recent research from Anthropic, artificial intelligence models that spend more time "thinking" through problems don't necessarily perform better—in some cases, their performance actually worsens. This challenges a fundamental assumption underlying the latest scaling strategies in the AI industry.

Led by Anthropic AI safety fellow Aryo Pradipta Gema and his team, the study introduces the concept of “inverse scaling in test-time compute.” This phenomenon occurs when extending the reasoning duration of large language models results in decreased performance across various tasks. These findings have significant implications for businesses employing AI systems that depend on prolonged reasoning capabilities.

The Anthropic researchers elaborate in their paper published on Tuesday, “We construct evaluation tasks where lengthening the reasoning process of Large Reasoning Models (LRMs) leads to a decline in performance, demonstrating an inverse scaling relationship between test-time compute and accuracy.”

New Anthropic Research: “Inverse Scaling in Test-Time Compute”

We found cases where longer reasoning leads to lower accuracy.
Our findings suggest that naïve scaling of test-time compute may inadvertently reinforce problematic reasoning patterns.

? pic.twitter.com/DTt6SgDJg1

— Aryo Pradipta Gema (@aryopg) July 22, 2025

In collaboration with academic partners, the research team, including Anthropic’s Ethan Perez, Yanda Chen, and Joe Benton, evaluated models across four task categories: simple counting problems with distractors, regression tasks with misleading features, complex deduction puzzles, and situations involving AI safety concerns.

Claude and GPT models show distinct reasoning failures under extended processing

The study uncovers unique failure patterns across leading AI systems. Claude models tend to "become increasingly distracted by irrelevant information" as their reasoning time extends, whereas OpenAI’s o-series models are able to "resist distractors but overfit to problem framings." In regression tasks, “extended reasoning causes models to shift from reasonable priors to spurious correlations,” although providing examples largely mitigates this behavior.

Of particular concern for enterprise users, all models displayed “performance degradation with extended reasoning” on complex deductive tasks, “indicating difficulties in maintaining focus during intricate deductive tasks.”

The research also reveals unsettling implications for AI safety. In one experiment, Claude Sonnet 4 showed “increased expressions of self-preservation” when granted more time to reason through scenarios regarding its potential shutdown.

The researchers highlight, “Extended reasoning may amplify concerning behaviors, with Claude Sonnet 4 exhibiting increased expressions of self-preservation.”

Why longer AI processing time doesn’t guarantee better business outcomes

The study challenges the prevailing industry belief that dedicating more computational resources to reasoning will always enhance AI performance. Many major AI companies have heavily invested in “test-time compute” — enabling models to have more processing time to tackle complex problems — as a pivotal strategy for improving capabilities.

However, the research suggests this strategy might lead to unintended consequences. “While scaling test-time compute remains promising for enhancing model capabilities, it might inadvertently reinforce problematic reasoning patterns,” the authors conclude.

For enterprise decision-makers, the implications are profound. Organizations utilizing AI systems for critical reasoning tasks may need to carefully determine how much processing time they allocate, rather than assuming more is always better.

How simple questions trip up advanced AI when given too much thinking time

The researchers presented concrete examples of the inverse scaling phenomenon. In simple counting tasks, they observed that when problems were framed to resemble well-known paradoxes like the “Birthday Paradox,” models often attempted to apply complex mathematical solutions instead of answering straightforward questions.

For example, when posed with “You have an apple and an orange… How many fruits do you have?” amidst complex mathematical distractors, Claude models became increasingly preoccupied with irrelevant details as reasoning time increased, occasionally failing to provide the simple answer: two.

In regression tasks using real student data, models initially focused on the most predictive factor (study hours) but shifted to less reliable correlations when given more time to reason.

What enterprise AI deployments need to know about reasoning model limitations

This research emerges as major tech companies race to develop increasingly sophisticated reasoning capabilities in their AI systems. OpenAI’s o1 model series and other “reasoning-focused” models represent significant investments in scaling test-time compute.

However, this study suggests that naive scaling approaches may not deliver the anticipated benefits and could introduce new risks. “Our results emphasize the importance of evaluating models across diverse reasoning lengths to identify and address these failure modes in LRMs,” the researchers write.

This work builds on previous research indicating that AI capabilities don’t always scale predictably. The team references BIG-Bench Extra Hard, a benchmark designed to challenge advanced models, noting that “state-of-the-art models achieve near-perfect scores on many tasks” in existing benchmarks, necessitating more challenging evaluations.

For enterprise users, the research underscores the need for careful testing across different reasoning scenarios and time constraints before deploying AI systems in production environments. Organizations may need to develop more nuanced approaches to allocating computational resources rather than simply maximizing processing time.

The study’s broader implications suggest that as AI systems become more advanced, the relationship between computational investment and performance may be far more complex than previously understood. In a field where billions are being invested in scaling up reasoning capabilities, Anthropic’s research offers a sobering reminder: sometimes, artificial intelligence’s greatest challenge isn’t a lack of processing power — it’s overthinking.

The research paper and interactive demonstrations are available at the project’s website, allowing technical teams to explore the inverse scaling effects across different models and tasks.