Claude and GPT models reach near-identical performance levels

Closing the performance gap

Recent data from Epoch AI reveals that the performance gap between the leading AI models from OpenAI and Anthropic has nearly vanished. According to calculations based on Epoch AI's public benchmarking data, OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 are now separated by only 0.7 points on the Capabilities Index.

This narrow margin follows a period of rapid development for both companies since early 2024. Anthropic's Claude moved from a score of 125.6 with Claude 3 Opus to 168.4 with Opus 5.5. During the same timeframe, OpenAI's GPT rose from a 128.9 snapshot of GPT-4o to 169.1 with GPT-6 Astra.

The data was compiled by Alex Michael on X using independent evaluations from Epoch AI. These figures are notable because they are not self-reported by the labs, providing a more objective look at how these models compare against one another in a standardized environment.

Significant gains in reasoning

The most striking aspect of the new data is the sheer scale of improvement both models have achieved in specific reasoning tasks. In competition math, both labs saw their models jump from single-digit percentages to a perfect 100.0% score. This suggests that both systems have mastered the logic required for these types of problems.

Other areas of growth include factual recall and graduate-level science. For example, Claude's factual recall rose from 12.6% to 72.2%, while GPT's rose from 26.0% to 75.6%. In graduate-level science, GPT improved by 46.6 points and Claude improved by 43.4 points, showing consistent progress across different types of complex knowledge.

However, as models improve, they begin to hit

The shift toward model fungibility

The data suggests a broader trend where the underlying model weights are becoming increasingly fungible. This means that the core intelligence provided by the base models is becoming more similar across different providers. Instead of one model being vastly superior in raw logic, both labs are reaching a point where their primary outputs are nearly indistinguishable on standard tests.

Because the models are clustering at the top of these benchmarks, the distinction between them is shifting. The focus is moving away from which model is "smarter" in a vacuum and toward how these models are implemented. This shift is a key component of the Subsidized AGI Economy, where the primary value is no longer just the model itself but how it is utilized.

The importance of the agentic harness

As the base models become more similar, the competitive advantage for companies like OpenAI and Anthropic is moving toward the "harness." This refers to the agentic scaffold or the specific software environment built around the model. The harness determines how the model interacts with tools, handles multi-step tasks, and delivers specific results to the end user.

For businesses, this means that the "intelligence" of the model is just one part of the equation. The way a lab designs the surrounding system to manage complex workflows becomes the primary differentiator. This is why the 0.7-point gap on the index is less significant than the architecture that allows a model to perform work in a real-world setting.

Context for business strategy

It is important to note that these benchmark scores reflect models at their best tested settings, which may involve using more compute per answer. This does not necessarily represent the performance of a model in a standard consumer product where resources might be more limited. Readers should view these numbers as a ceiling of capability rather than a guaranteed experience in every use case.

This analysis aligns with frameworks used by over 95,000 executives who follow the Business Engineer for AI strategy. These leaders look at how models like Claude, GPT, and Perplexity can be integrated into corporate structures. For these decision-makers, the goal is finding the right balance between a model's raw capabilities and the practical harness that makes it useful for a company.