GPT-5.4 Released: OpenAI's First Unified Model

According to QbitAI (量子位), GPT-5.4 is OpenAI's first model that integrates reasoning, coding, computer-native interaction, deep web search, and million-token context into a single model.

Core Capability Improvements

1. Deep knowledge work

• GDPval benchmark: 83.0%
• Investment banking modeling test: 87.3% (GPT-5.2: 68.4%)

2. Native computer use — OpenAI's first general-purpose model with native computer operation support

• WebArena browser tasks: 67.3% success rate
• Online-Mind2Web screenshot-based operation: 92.8% success rate
• OSWorld-Verified desktop operation: 75.0% (above the human average of 72.4%)

3. Advanced coding and debugging — Absorbs GPT-5.3-Codex capabilities, supports a /fast mode (1.5x token output speed), and adds Playwright (Interactive) visual debugging

• SWE-Bench Pro: 57.7%

Efficiency Optimizations

• Significantly reduced reasoning token consumption compared to GPT-5.2
• Tool search mechanism lowers token usage by 47%

API Pricing

| Model | Input (per 1M tokens) | Output (per 1M tokens) | |---|---|---| | GPT-5.4 | $2.5 | $15 | | GPT-5.4 Pro | $30 | — |

GPT-5.4 Pro is aimed at high-end use cases.

Timeline

• GPT-5.4 Thinking replaces GPT-5.2 Thinking
• GPT-5.2 will be retired in three months
• The GPT-5.1 series will be removed from ChatGPT on March 11