On March 5, OpenAI released GPT-5.4, and the biggest change is this: there is no longer a dedicated programming model.
Previously, GPT-5.3-Codex was a standalone model specialized for coding, released separately from the main GPT-5.x line. Now, GPT-5.4 directly unifies frontier-level coding, reasoning, and Computer Use capabilities into a single model architecture.
Triple capability
| Dimension | GPT-5.4 | vs GPT-5.3-Codex |
|---|---|---|
| Coding (SWE-bench Pro) | 57.7% | 55.6% |
| Computer control (OSWorld) | 75% | 64% |
| Knowledge work (GDPval) | 83% | N/A |
Coding is slightly stronger, computer control pulls ahead significantly, and knowledge work capability was entirely absent from Codex. It is the first single model to reach frontier-level performance across all three dimensions.
Five versions
- Standard: Regular version
- Thinking: Deep reasoning version
- Pro: High-end version
- Mini: Lightweight version
- Nano: Ultra-lightweight version
API pricing covers the full range from a few dollars to tens of dollars. In ChatGPT it is called "GPT-5.4 Thinking", and in the API it is gpt-5.4.
Experimental support for 1 million token context, configurable via the model_context_window and model_auto_compact_token_limit parameters.
Strategic signal
GPT-5.4 represents a clear shift in OpenAI's product strategy: from "different models for different scenarios" to "one model does it all."
This converges with DeepSeek V3.1's hybrid model approach and Qwen3's dual-mode design. It seems "unified model" is becoming an industry consensus — users should not have to worry about "which model to use for this problem."
Sources: CocoLoop, NxCode Technical Guide