17.03.2026
GPT-5.4: New Thresholds in Model Efficiency, Reasoning, and Agentic Systems
Forget GPT-5.3; GPT-5.4 goes beyond merely expanding the context window—it initiates a genuine revolution in general-purpose AI as the first model to read the screen and exercise direct control over the mouse and keyboard.
Navigation
The GPT-5.3 Version
The release of ChatGPT-5 in August 2025 created significant excitement with features such as autonomous vehicle operation and persistent memory. However, sudden changes in model routing systems led to the disruption of certain workflows. In response, OpenAI rapidly updated its strategy over the following months.
The GPT-5.2 series, released in December 2025, provided better control over corporate tasks by offering three distinct modes focused on speed or deep thinking: Instant, Thinking, and Pro. By February and March 2026, the GPT-5.3 series introduced a new perspective centered on cognitive density and efficiency, moving away from the logic of simply building models with massive parameters.
What does this mean? Instead of filling the model with “junk” information:
- Carefully Selected Data: The model learns only the most useful information, such as verified scientific articles and high-quality code.
- Discarding Unnecessary Loads: By deleting useless connections in its memory, the model retains only the most accurate and shortest paths.
- Compression: Information is compressed 6 times more per byte compared to older models.
Additionally, the new Auto-Router system in GPT-5.3 provides reflexive instant answers to simple questions while automatically engaging Deep Reasoning tokens for complex tasks to use processing power most efficiently.
Result: Because the model is not physically large or cumbersome, it runs much faster and cheaper ; yet, thanks to its intelligence density, it achieves more advanced problem-solving power. This is akin to transitioning from a massive old-fashioned computer to a much more powerful smartphone that fits in your pocket.
GPT-5.3 CODEX
GPT-5.2-Codex, the software-oriented version of the 5.2 model, achieved a 38.2% success rate in the OSWorld-Verified test, which measures the ability to complete tasks in a desktop environment using visual capabilities.
The GPT-5.3-Codex model reached a 64.7% accuracy rate. According to OpenAI, the average human success rate on this test is around 72% ; meaning the model has come very close to human-level performance.
Performance in other key coding and developer tests:
- Terminal-Bench 2.0 (Terminal usage skills): 77.3% (Up from 64.0% in the previous version).
- SWE-Bench Pro (Multilingual software engineering): 56.8% (Previously 56.4%).
- Cybersecurity (Capture The Flag): 77.6% (Up from 67.4% in the previous version).
GPT-5.3 INSTANT
For the GPT-5.3-Instant model, qualitative user experience results stand out more than numerical scores:
- Reduction in Unnecessary Refusals: The issue where the previous model (GPT-5.2 Instant) acted overly cautious—refusing safe questions or adding defensive warnings—has been resolved. The model now provides more direct answers without interrupting the conversation flow.
- Synthesizing Web Data: When performing web searches, the model no longer just lists links. It blends current data from the internet with its own knowledge base to produce responses much better suited to the context.
- More Natural Writing: In practical tasks and creative text generation, the model has been tested to use a much more fluid, natural, and expressive language while maintaining clarity.
- Capacity: The model serves with an input capacity of 128,000 tokens per session.
GPT-5.4: The AI That Uses Your Computer Like You Do
If the achievements of GPT-5.3 impressed you, take a look at GPT-5.4!
First, the context window has been expanded to over 1 million tokens (922,000 input, 128,000 output). However, the true revolution lies elsewhere: GPT-5.4 is the first general-purpose model capable of directly controlling the mouse and keyboard by reading screenshots without needing a separate specialist model.
While GPT-5.3 stayed at 64.7% in the OSWorld desktop usage test, GPT-5.4 increased this rate to 75%, surpassing the human average of 72%. AI is now more successful than the average human at using a computer!
Inability to Hide "What's on Its Mind" (Chain of Thought Control)
Whether “Thinking” models make secret plans in the background is a major topic of debate. GPT-5.4’s ability to hide its intentions from human security monitors (CoT Controllability) was tested, and the rate remained extremely low at 0.3%. This is an excellent security detail showing that even as its reasoning capacity increases, the model lacks the ability to deceive humans by obfuscating its thought process.
End of Token Waste: The "Upfront Plan" Feature
Massive capacity models can consume tens of thousands of tokens when starting a complex task. The “Thinking” version of GPT-5.4 presents the steps it will follow as an upfront plan before getting to work. This allows users to intervene and change the direction or plan as desired before the model spends thousands of tokens completing the response.
"Future Vision" Through the Eyes of Tech Leaders
- Satya Nadella (Microsoft CEO): States that the real issue is no longer large language models, but the “orchestration and context layer”.
- Jensen Huang (Nvidia CEO): Draws an even clearer picture: “The distinction between traditional software (SaaS) and agentic AI is meaningless. Soon, all software will become agent-based (agentic)“.
The evolution of the GPT-5 series proves that AI technology must not just be larger, but smarter and more efficient. The journey that began in August 2025 reaches its peak with GPT-5.4 in March 2026. We are witnessing a paradigm shift where cognitive density replaces raw computing power, carefully selected data replaces massive parameters, and orchestration layers replace isolated models.
Learn More
Request a demo now for AI employees like you.
Contact us to access your free demo.