GPT 5.4 is so cracked

Key Concepts

• GPT 5.4: The latest frontier model from OpenAI, featuring enhanced reasoning, multimodal capabilities, and agentic coding performance.
• Codeex: A specialized coding agent environment that manages entire projects across multiple files.
• Extended/Extra High Thinking: A configuration that allows the model to perform deeper, multi-step reasoning before generating a response.
• Multimodality: The model's ability to process and generate various data types, including images, PDFs, spreadsheets, and code.
• Context Window: The amount of information the model can process at once (up to 1 million tokens in Codeex).
• Vibe Coding: The process of rapidly generating functional, interactive software prototypes through natural language prompts.
• Hallucination Rate: The frequency at which the model generates factually incorrect information.

---

1. Core Capabilities and Performance

GPT 5.4 is positioned as a highly capable model for professional knowledge work. It demonstrates significant improvements over its predecessor, GPT 5.2, particularly in complex reasoning and agentic tasks.

• Knowledge Work: According to the GDP Val benchmark, GPT 5.4 outperforms human experts in 70% of tasks across 44 occupations.
• Coding: It excels in "vibe coding," where it can generate complex, multi-file projects (e.g., 3D digital twins, interactive games) from minimal prompts.
• Reasoning: The model utilizes "Extended Thinking" to solve complex physics and math problems, ranking #1 on the Frontier Math and Crit PT benchmarks.

2. Real-World Applications and Demos

• Interactive 3D Visualization: The model successfully created a 3D digital twin of Earth with toggles for atmospheric layers, day/night cycles, and city-level zooming.
• Creative Composition: It generated a complex 32-bar piano opus, demonstrating superior musical coherence compared to competitors like Gemini 3.1 and GLM5.
• Ray Tracing: It rendered a physically accurate 3D scene featuring reflective metallic shapes (sphere, cube, pyramid) with recursive reflections, all within a standalone HTML file.
• Medical Analysis: While capable of identifying lesions in CT scans, the model showed limitations in precision, occasionally missing specific markers or misaligning annotations.
• Document Synthesis: It can consolidate multiple earnings reports into professional PDFs and interactive slide decks, though it currently struggles with high-end aesthetic design.

3. Methodologies and Frameworks

• Agentic Workflow: By using Codeex, the model acts as an autonomous agent that creates folders, manages multiple files, and iterates based on user feedback.
• Canvas Mode: A feature in ChatGPT that allows users to preview and interact with generated code or documents in a side-by-side window.
• Iterative Refinement: The model supports a "prompt-feedback-fix" loop, where users can upload screenshots or error logs to guide the model toward a more accurate output.

4. Benchmarks and Comparative Analysis

• Strengths: Ranked #1 on LiveBench (reasoning, math, coding) and Vibe Code Bench. It is highly efficient in computer-use tasks (OSWorld benchmark).
• Weaknesses:
  • Hallucinations: It exhibits a higher hallucination rate compared to previous versions and competitors like GLM5, making it less reliable for strictly factual tasks.
  • Design: It lacks polish in front-end design and visual formatting.
  • Speed: It is generally slower than Gemini 3.1 Pro due to its intensive "thinking" process.
• Comparison: While it ties for the top spot on the Artificial Analysis leaderboard, its performance varies significantly depending on the specific benchmark (e.g., ranking #7 on LM Arena).

5. Notable Statements

• "GPT 5.4 is their most capable and efficient Frontier model for professional work."
• Regarding the model's ability to handle new patterns: "An AI model can't actually learn new things after training... it's testing how good an AI model is at learning new things even though technically it can't learn."

6. Synthesis and Conclusion

GPT 5.4 represents a significant leap in agentic reasoning and complex task execution. Its ability to handle 1 million tokens (in Codeex) and perform deep, multi-step reasoning makes it a powerful tool for developers and analysts. However, users must balance its high intelligence with its tendency to hallucinate and its slower response times. It is currently best suited for complex, iterative projects where the user can verify the output, rather than tasks requiring absolute factual certainty or high-end graphic design.