GPT-5.4 vs Claude Opus 4.6: The AI Showdown for Businesses

GPT-5.4 and Claude Opus 4.6: Two Titans of AI in Competition

In the world of enterprise-focused artificial intelligences, two models are vying for supremacy: GPT-5.4 from OpenAI and Claude Opus 4.6 from Anthropic. OpenAI, despite a turbulent period due to its controversial partnership with the Pentagon, has decided to launch a new offensive with GPT-5.4. Sam Altman, the CEO of OpenAI, has clarified that he has little influence over the decisions made by the military organization. This model, aimed at professionals, hopes to stem the exodus of users to Claude, which has established itself in this segment over the past few months. This initiative closely follows the launch of GPT-5.3 Instant, a model more focused on everyday tasks, deployed on March 3. GPT-5.4 was launched three days later, thereby strengthening OpenAI's offering.

Comparative Performance in Benchmarks

The performances of Claude Opus 4.6 and GPT-5.4 are often very close in comparative tests. On the BrowseComp benchmark, which evaluates the ability to find difficult information online, Claude Opus 4.6 has a slight edge with a score of 84% compared to 82.7% for GPT-5.4. Similarly, in the Humanity's Last Exam test, which measures multidisciplinary reasoning, Claude slightly outperforms GPT with 40% against 39.8%. However, for tool usage in the τ2-bench Telecom benchmark, both models are nearly equal, with 99.3% for Claude and 98.9% for GPT, and this benchmark is considered saturated.

In contrast, GPT-5.4 stands out significantly on the MCP Atlas benchmark, which tests large-scale tool usage. GPT-5.4 scores 67.2%, far surpassing Claude Opus 4.6, which reaches 59.5%. In the realm of vision and visual reasoning, GPT-5.4 also excels, achieving 81.2% on MMMU Pro, compared to 73.9% for Claude. In coding, both models are nearly equivalent on SWE-bench Verified, with 80.8% for Claude and 80% for GPT.

Performance Analysis

In terms of theoretical and practical capabilities, GPT-5.4 allows OpenAI to close the gap with Anthropic. This model is particularly effective in environments requiring numerous MCP connectors, typical of automated agents. On the other hand, Claude Opus 4.6 stands out for its ability to reason deeply and manage long contexts, especially in online research. In coding development, both models are on par, although GPT-5.4 is not yet optimized for this task. OpenAI plans to launch a codex version in the coming weeks, specifically dedicated to development.

Use Cases: Performance in Real Situations

To assess the practical capabilities of both models, we subjected them to three distinct scenarios:

• Summarizing a research document in 100 words
• Creating an Excel sheet from a company's financial reports
• Generating an SVG image of an iPhone

These exercises test a variety of skills: synthesis, data extraction and structuring, as well as visual code generation.

Document Summary: Advantage to GPT-5.4

The goal was to test the models' ability to summarize complex documents while adhering to a strict constraint of 100 words. The prompt required covering the methodology and main results without exceeding the imposed word count. Although neither strictly adhered to the limit, GPT-5.4 provided a clearer and more structured text, earning it the preference.

Creating an Excel Sheet: Claude Opus 4.6 Wins

For this test, the models had to transform raw financial data from Tesla for the year 2025 into a structured Excel file. Claude Opus 4.6 demonstrated a better ability to ingest and organize this data in a usable manner, surpassing GPT-5.4 in this exercise.

Generating an SVG Image: A Challenge for Both Models

Generating an SVG image of an iPhone highlighted the visual code creation capabilities of the models. The results showed that both models still have progress to make in this area, although each managed to produce functional results.