GPT-5.4: Three Versions, One Question
OpenAI drops GPT-5.4.
Official positioning: “our most capable and efficient frontier model for professional work.”
Three Versions, Same Day
First time OpenAI shipped the same base model in three versions simultaneously:
GPT-5.4 standard: Core version. Incorporates GPT-5.3-Codex coding capabilities, with improvements to how the model works “across tools, software environments, and professional tasks involving spreadsheets, presentations, and documents.”
GPT-5.4 Thinking: Visible reasoning process, interruptible and redirectable mid-thought.
GPT-5.4 Pro: For complex tasks, priced at 12x standard.
Same model foundation. Difference is reasoning depth and use case fit.
Coding Capability Is the Core Upgrade
GPT-5.4 inherits GPT-5.3-Codex directly—OpenAI’s previous flagship coding model. Combined with improved tool calling and software environment operation, code generation and execution is the most concrete improvement this cycle.
Computer Use Deserves Its Own Section
First time OpenAI adds native Computer Use to a general-purpose model.
Not telling you which buttons to click—actually executing operations in browser and desktop: clicks, typing, navigation.
Anthropic’s Claude had this first. Now OpenAI has it too. Computer Use is no longer a differentiator—it’s becoming standard for flagship models.
Million-Token Context
Codex version supports 1M token context. One shot for entire codebases—no chunking.
Practical value for large-scale code review, cross-file refactoring, multi-document analysis.
Pricing
| Version | Input | Output |
|---|---|---|
| Standard | $2.50/M | $15/M |
| Pro | $30/M | $180/M |
Pro at 12x standard. OpenAI has not published data on how much better Pro actually is on real tasks. Only way to know is to test.
Computer Use 安全机制、版本边界、差异化定价逻辑
1. What are the Computer Use safety mechanisms
Model that operates computers—if abused, harm is significant. What guardrails exist? No public technical details.
2. What is the actual gap between Thinking and Pro
12x price difference. How much better is Pro on real tasks? No reference data from OpenAI.
3. Why ship three versions simultaneously
本质上是差异化定价。不同支付意愿的用户买不同档位。But the boundaries between the three tiers—OpenAI hasn’t clarified.