Artificial intelligence models are evolving at a rapid pace, and OpenAI has just raised the bar again with the release of GPT-5.4. Designed for complex professional workloads, the new flagship model integrates advanced reasoning, software development capabilities, and agent-style workflows into a single system.

Available inside ChatGPT under the name GPT-5.4 Thinking, the model is also accessible through the API and Codex for developers. OpenAI is also introducing a more powerful version called GPT-5.4 Pro, targeting enterprise-level applications that demand maximum performance.

Beyond raw performance improvements, GPT-5.4 introduces a major innovation: native computer control, allowing the model to operate software interfaces directly by interpreting screen content and simulating user actions. This marks a significant step toward AI systems that can perform real digital tasks rather than simply generating text.

A Model Designed for Professional Workflows

With GPT-5.4, OpenAI clearly targets professional productivity use cases. The model delivers noticeable improvements in tasks commonly associated with office work, including:

• Data analysis
• Business document creation
• Spreadsheet processing
• Presentation generation

These improvements are reflected in benchmark results. On GDPval, a test that evaluates AI performance across 44 real-world professions—ranging from accounting and finance to engineering and legal work—GPT-5.4 achieves an 83% success rate, compared to 70.9% for GPT-5.2.

OpenAI also reports stronger performance in document-heavy workflows. In financial modeling exercises similar to tasks performed by junior banking analysts, GPT-5.4 reaches 87.3% accuracy, a substantial improvement over the 68.4% score achieved by GPT-5.2.

Another key area of progress is reliability. According to OpenAI’s internal evaluations:

• Individual claims generated by the model are 33% less likely to be incorrect.
• Entire responses are 18% less likely to contain factual errors.

Reducing hallucinations remains one of the most critical challenges in modern AI systems, and these improvements represent a meaningful step forward.

Native Computer Control Arrives for the First Time

One of the most significant innovations in GPT-5.4 is the introduction of built-in computer control capabilities, sometimes referred to as computer use.

Unlike earlier implementations that relied on external tools or specialized agents, GPT-5.4 integrates this functionality directly within the model itself. The system can interpret screenshots of a computer interface and perform actions such as:

• Clicking buttons
• Navigating menus
• Typing text
• Interacting with applications

In practical terms, this allows AI agents to operate software environments much like a human user would.

Benchmark results highlight the impact of this feature. On OSWorld-Verified, a benchmark measuring a model’s ability to interact with desktop environments, GPT-5.4 reaches a 75% success rate. That’s a major leap from 47.3% achieved by GPT-5.2, and it even surpasses the 72.4% average human performance measured in the test.

On WebArena-Verified, which evaluates how well AI models use web browsers to complete tasks, GPT-5.4 scores 67.3%, slightly higher than the 65.4% result of GPT-5.2.

Developers can already access these computer-interaction features through the OpenAI API, along with dedicated documentation explaining how to integrate them into automated workflows.

Stronger Coding Performance with Codex Integration

GPT-5.4 also incorporates the capabilities of GPT-5.3 Codex, OpenAI’s specialized model designed for software development tasks.

By combining Codex’s programming strengths with GPT-5.4’s enhanced reasoning abilities, OpenAI aims to create a more capable AI developer assistant.

On SWE-Bench Pro, a benchmark focused on fixing real bugs in open-source projects, GPT-5.4 achieves a 57.7% success rate, slightly outperforming the 56.8% score of GPT-5.3 Codex while also delivering faster response times.

In Codex, developers can also enable a /fast mode, which increases token generation speed by up to 1.5× without reducing output quality.

OpenAI reports particularly strong improvements in frontend development tasks, with generated interfaces appearing more polished and functional than those produced by previous models.

Another new experimental feature called Playwright (Interactive) allows Codex to visually debug web applications while they are being built. The model can run tests directly in a browser and inspect the results in real time, making debugging workflows significantly faster.

A Massive Context Window: Up to 1 Million Tokens

GPT-5.4 introduces one of the largest context windows ever offered in a widely accessible AI model: up to one million tokens.

To put that into perspective, this is roughly equivalent to:

• 750,000 words of text, or
• Multiple full-length novels, or
• An entire large codebase analyzed in a single request.

In Codex, the expanded context window is currently available as an experimental feature. Developers can activate it using the parameters:

• model_context_window
• model_auto_compact_token_limit

However, there is an important pricing consideration. Requests that exceed the standard 272,000-token window are billed at double the normal rate.

Despite this cost, the extended context window opens the door to powerful new applications, including AI agents capable of planning, executing, and validating complex tasks across long sequences of interactions without losing context.

Availability in ChatGPT and API Pricing

GPT-5.4 is already available inside ChatGPT under the name GPT-5.4 Thinking for subscribers to Plus, Team, and Pro plans.

The model replaces GPT-5.2, although the previous version will remain accessible for three months in the Legacy Models section before being permanently removed on June 5, 2026.

Enterprise and educational organizations using ChatGPT can enable early access through administrator settings.

Meanwhile, GPT-5.4 Pro is reserved for Pro and Enterprise users who require higher performance for demanding workloads.

API Pricing Comparison

Compared to GPT-5.2, the new model is slightly more expensive at the API level:

Although the new model costs more per token, OpenAI claims greater token efficiency, meaning many tasks may require fewer tokens overall—potentially reducing total operational costs.

Conclusion

The launch of GPT-5.4 marks a major milestone in the evolution of AI models. By combining advanced reasoning, powerful coding capabilities, agent-style workflows, and native computer control, OpenAI is pushing AI systems closer to becoming fully autonomous digital assistants capable of executing real tasks.

For businesses, developers, and researchers, the expanded context window and improved reliability could significantly enhance productivity and automation capabilities.

While GPT-5.4 comes at a higher price than its predecessor, its performance improvements and new features may make it one of the most powerful AI tools currently available—especially for complex professional workloads.