SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

GPT‑5.4 - OpenAI Launches a Flagship AI Model for Professional Work

Author: AI小集 · Original source: AI工具集

What is GPT‑5.4

GPT-5.4 is a flagship AI model launched by OpenAI, positioned as "the strongest frontier model designed specifically for professional work." For the first time, the model integrates reasoning,programming, native computer operation, deep web search, and a million-token context into a single model, without sacrificing performance in any individual area. On the OSWorld computer operation test, it surpassed human level for the first time with a 75% success rate; on the GDPval knowledge work test, 83% of tasks reached or exceeded professional standards. GPT-5.4 supports understanding interfaces from screenshots and performing mouse and keyboard operations, and can independently complete complex cross-application tasks, marking a key shift for AI from "answering questions" to "completing tasks."

GPT‑5.4

Main Features of GPT‑5.4

  • Native computer operation: Understands interfaces through screenshots, performs mouse clicks and keyboard input, and completes complex tasks across applications (such as sending emails, filling out forms, and operating web pages), with a 75% success rate on the OSWorld test, surpassing human level for the first time.
  • Deep knowledge work: Supports handling real work tasks across 44 occupations, including creating PPTs, financial modeling, data analysis, legal document analysis, and more, with 83% on the GDPval test reaching or exceeding professional standards.
  • Advanced programming and debugging: Integrates GPT-5.3-Codex capabilities, supports code generation, review, and optimization; the newly added Playwright Interactive enables visual debugging while writing code and can automatically test web applications.
  • Intelligent tool calling: Introduces a tool search mechanism that queries tool definitions on demand, reducing token consumption by 47%; supports flexible invocation of external tools and APIs in multi-step tasks.
  • Deep web search: Supports multiple rounds of continuous searching, filtering information, and integrating results, reaching 82.7% on the BrowseComp test, and excels at handling complex "needle-in-a-haystack"-type queries.
  • Ultra-long context processing: The API supports up to a 1 million token context, can process complete project documents in one go; supports high-fidelity image input up to 10.24 million pixels.
  • Real-time task control: Shows the work plan before complex tasks, and the direction can be adjusted at any time during execution without needing to start over.
  • Multimodal visual understanding: Visual reasoning, document parsing, and interface recognition capabilities are significantly improved, supporting high-resolution images and complex interface operations.

Performance of GPT‑5.4

  • Knowledge work:
    • GDPval test: 83.0% of tasks reached or exceeded industry expert level (GPT-5.2 only 70.9%).
    • Investment-banking-level spreadsheet modeling: 87.3% (GPT-5.2 was 68.4%).
    • PPT generation: Human reviewers68% preferred GPT-5.4.
  • Computer operation:
    • OSWorld-Verified: 75.0% success rate, surpassing the human baseline of 72.4% for the first time, while GPT-5.2 was only 47.3%.
    • WebArena-Verified: 67.3%, leading in browser task handling capability.
    • Online-Mind2Web: 92.8%, able to operate web pages precisely using only screenshot observation.
  • Programming capability:
    • SWE-Bench Pro：57.7%, slightly exceeding GPT-5.3-Codex(56.8%), while also having lower latency and higher token efficiency.
    • Terminal-Bench 2.0: 75.1%, with robust terminal task handling performance.
  • Tools and search:
    • BrowseComp: 82.7% (up to 89.3% for the Pro version), a 17 percentage point improvement over GPT-5.2 (65.8%), setting a new industry high.
    • Toolathlon: 54.6% (compared to 45.7% for GPT-5.2), with more accurate multi-step tool calling.
    • Token consumption: reduced by 47%, and the tool search mechanism significantly lowers costs while maintaining the same level of accuracy.
  • Academics and reasoning:
    • GPQA Diamond: 92.8% (94.4% for the Pro version), with science Q&A approaching a perfect score.
    • Humanity’s Last Exam (with tools): 52.1% (58.7% for the Pro version), with outstanding performance on high-difficulty comprehensive tests.
    • ARC-AGI-2: 73.3% (83.3% for the Pro version), a major leap in abstract reasoning ability, while GPT-5.2 Pro is only 54.2%.
  • Reliability:
    • Probability of a single factual error: reduced by 33%。
    • Error rate for complete responses: reduced by 18%, making it OpenAI's most factual model to date.
GPT‑5.4

How to use GPT‑5.4

  • ChatGPT: Visit the ChatGPT official website or ChatGPT App, GPT-5.4 is already available to ChatGPT Plus, Team, and Pro users, replacing GPT-5.2 Thinking as the default thinking model..
  • OpenAI API: Call the gpt-5.4 or gpt-5.4-pro model endpoints via an API key, supporting up to a 1 million Token context and the tool search feature, with billing based on Token usage.
  • Codex: Visit the Codex official website and enter /fast to enable accelerated mode, or use the experimental 1M context window to handle large code projects, with support for Playwright Interactive visual debugging.

GPT‑5.4 product pricing

  • ChatGPT subscription
    • Plus/Business subscription: includes GPT-5.4 Thinking (3,000 times per week).
    • Pro subscription: includes GPT-5.4 Pro (unlimited).
  • API usage-based billing
    • GPT-5.4: input $2.50/million Tokens, cached input $0.25/million Tokens, output $15/million Tokens.
    • GPT-5.4 Pro: input $30/million Tokens, output $180/million Tokens.

GPT‑5.4 application scenarios

  • Office automation: replace manual work to complete knowledge work such as creating PPTs, financial modeling, data analysis, and document processing.
  • Intelligent Agent: autonomously operate a computer to complete cross-application tasks, such as automatically sending emails, filling out forms, scheduling appointments, and batch data entry.
  • Software development: full-stack development, code review, and Bug fixing, with support for Playwright interactive debugging while writing and testing, and the ability to independently build complex Web applications and games.
  • Enterprise processes: integrate with internal systems to achieve RPA automation, handling long-cycle, multi-step tasks such as tax filing, contract review, and customer service.
  • Deep research: multi-round web search integrates information and handles complex queries that require cross-source verification.