image.png

AIBase | August 8, 2025 report

On August 7, 2025, OpenAI officially released its latest generation of artificial intelligence model, GPT-5, marking another milestone in the field of generative AI. As the most powerful and intelligent model from OpenAI to date, GPT-5 not only achieved significant breakthroughs in technical capabilities but also further promoted the popularization of AI through multi-level pricing and access strategies. This article will comprehensively analyze the model characteristics, pricing structure, version differences, evaluation scores, user accessibility, model capabilities, and industry opinions on GPT-5.

I. Overview of the GPT-5 Model

GPT-5 is OpenAI's latest cutting-edge model, integrating advanced reasoning capabilities, multimodal processing, and task execution abilities, aiming to provide users with a smarter and more efficient AI experience. Compared to previous models, GPT-5 automatically selects the appropriate processing method based on task complexity and user needs by combining a unified system architecture with fast response lightweight models and deep reasoning models. OpenAI CEO Sam Altman called it "the best model in the world" and stated that it is an important step toward general artificial intelligence (AGI), although he also pointed out that GPT-5 still lacks continuous learning ability, which is considered one of the key features of AGI.

Main Features:

  1. Unified System Architecture: GPT-5 dynamically selects between fast response models or deep reasoning models (such as GPT-5Thinking) based on task type, complexity, and user intent through a real-time router, simplifying user operations and eliminating the complexity of manually selecting models.
  2. Multimodal Capabilities: Supports text and image processing and plans to expand to video and audio processing, further enhancing its application capabilities in various scenarios.
  3. Reducing Hallucinations: GPT-5 has made significant progress in reducing generated false information (hallucinations), with a hallucination rate 26% lower than GPT-4o.
  4. Improved Security: Through the "safe completions" mechanism, GPT-5 provides high-level responses when handling potentially risky questions, avoiding the generation of harmful content. OpenAI conducted 5,000 hours of safety testing during development to ensure the model's robustness and reliability.

II. Pricing and Version Differences

GPT-5 offers multiple versions and pricing strategies to meet the needs of different user groups. Below are detailed explanations of the main versions and their pricing structures:

1. User-side Pricing (ChatGPT Platform)

  • Free Tier: All ChatGPT free users can use GPT-5, which is the first time OpenAI has opened up the reasoning capability model to the public for free. However, free users have undisclosed prompt limits, and after reaching the limit, they will switch to a lighter yet high-performance GPT-5Mini.
  • Plus Tier ($20/month): Plus users enjoy higher GPT-5 usage quotas, suitable for daily tasks, with better performance than the free tier.
  • Pro Tier ($200/month): Pro users have unlimited access to GPT-5 and can use the enhanced GPT-5Pro, which provides more accurate answers through additional computing resources, suitable for high-intensity and specialized task requirements.
  • Team, Enterprise, and Education Users: These users will receive GPT-5 as the default model one week after launch, with relatively generous quotas suitable for organizational applications.

2. Developer API Pricing

GPT-5 offers three versions through OpenAI's API, targeting different cost and performance needs:

  • GPT-5: $1.25 per million input tokens, $10 per million output tokens.
  • GPT-5Mini: $0.25 per million input tokens, $2 per million output tokens, suitable for high-cost-effectiveness scenarios.
  • GPT-5Nano: $0.05 per million input tokens, $0.40 per million output tokens, available only for API use, the most economical choice, competitive with Google's Gemini 2.5 Flash and Flash-Lite.
  • GPT-5Chat: Designed specifically for Pro-tier users, automatically adapting to conversation needs in ChatGPT.

Additionally, developers can control the length of the model's response using the new "verbosity" parameter to optimize costs and output efficiency.

3. Version Differences

  • GPT-5: The core model, balancing speed and performance, suitable for most complex tasks.
  • GPT-5Mini: A lightweight version, faster and cheaper, suitable for high-traffic applications such as chatbots or mobile voice assistants.
  • GPT-5Nano: An ultra-lightweight version, available only via API, optimized for low-cost and low-latency scenarios.
  • GPT-5Pro: An enhanced version, available only to Pro users, offering stronger reasoning capabilities and higher-quality outputs, suitable for professional development and enterprise-level applications.
  • GPT-5Thinking: A deep reasoning model designed for complex tasks, with longer response times but higher accuracy; Pro users can choose to use it by setting it.

III. Evaluation Scores and Performance

GPT-5 demonstrated excellent performance across multiple benchmark tests, particularly in programming, mathematics, and health-related tasks. Here are its performances in key evaluations:

  • SWE-bench Verified (Programming Tasks): GPT-5 scored 74.9%, surpassing Anthropic's Claude Opus 4.1 (74.5%) and Google DeepMind's Gemini 2.5 Pro (59.6%). Compared to the previous model o3, GPT-5 showed significant improvements in efficiency, with a 22% reduction in output tokens and a 45% decrease in tool calls.
  • Aider Polyglot (Code Editing): GPT-5 set a record of 88%, with a third fewer errors compared to o3, demonstrating its strong capabilities in software engineering tasks.
  • AIME2025 (Mathematics): Without using tools, GPT-5 scored 94.6%, setting a new benchmark in mathematics.
  • GPQA Diamond (PhD-Level Science Questions): GPT-5Pro scored 89.4%, outperforming Claude Opus 4.1 (80.9%) and xAI's Grok4Heavy (88.9%).
  • Humanity's Last Exam (Comprehensive Test): GPT-5Pro scored 42% when using tools, slightly below xAI's Grok4Heavy (44.4%).
  • HealthBench Hard (Health-Related): GPT-5 scored 46.2%, showing more accurate and reliable performance in answering health-related questions.

However, GPT-5 did not fully surpass its competitors in certain areas. For example, in the ARC-AGI test, GPT-5 did not meet expectations, indicating limitations in certain abstract reasoning tasks. Additionally, its knowledge cutoff date is 2024, limiting its ability to handle the latest information.

IV. User Access Barriers

The release of GPT-5 reflects OpenAI's commitment to AI democratization, lowering user access barriers through multi-layered access strategies:

  • Free Users: Can access GPT-5 through ChatGPT without additional fees. Free-tier users get their first experience with a reasoning model, though there are usage quota limits. Full functionality may be gradually unlocked over several days.
  • Paid Users: Plus and Pro users can obtain higher quotas or unlimited access through subscriptions. Pro users can also use advanced models like GPT-5Pro and GPT-5Thinking.
  • Developers: Can easily integrate the GPT-5 series models through OpenAI's API platform (platform.openai.com) or Python SDK. API supports multiple version choices to meet different budget and performance needs.
  • Enterprise and Education Users: Organizations can obtain customized access rights through Microsoft Azure AI Foundry or OpenAI's Team/Edu/Enterprise programs, suitable for large-scale deployment.

Additionally, GPT-5 supports personalized settings, such as choosing four preset personalities (Cynic, Robot, Listener, Nerd) and chat colors, further enhancing the user experience.

V. Model Capabilities

GPT-5 has shown significant performance improvements in multiple areas, especially in the following aspects:

  1. Programming and Software Engineering:

    • GPT-5 is hailed as the "best programming model in the world," capable of handling complex codebases, generating front-end UI, debugging code, and completing end-to-end tasks.
    • For example, OpenAI demonstrated GPT-5 designing a single-page game application called "Jumping Ball Runner," featuring a colorful UI, parallax scrolling background, and cartoon characters, showcasing its potential in quickly generating complex applications.
    • Developer feedback indicates that GPT-5 outperforms Claude Sonnet 4 in handling outdated codebases and real-time task execution.

  2. Reasoning and Complex Tasks:

    • GPT-5 excels in mathematical, scientific, and logical tasks through its "thinking" mechanism (chain-of-thought), making it suitable for multi-step reasoning scenarios.
    • For instance, when handling legal document summaries or loan advisor tasks, GPT-5 can provide precise results based on specific instructions.

  3. Multimodal Processing:

    • Supports text and image inputs, with future expansion to video and audio, making it applicable to a wider range of scenarios such as education, customer service, and content creation.
    • In the health sector, GPT-5 can answer complex health-related questions with greater reliability.

  4. Tool Usage and Integration:

    • GPT-5 supports parallel tool calls, intelligently selecting and combining tools (such as web search, RAG, SQL queries), significantly improving task execution efficiency.
    • Deeply integrated with the Microsoft ecosystem, supporting platforms such as Microsoft 365 Copilot and GitHub Copilot.

  5. Writing and Content Generation:

    • GPT-5 performs well in writing, capable of generating texts with literary depth and rhythm, suitable for creating stories, speeches, and professional reports. However, some testers believe its writing ability is slightly inferior to GPT-4.5.

VI. Industry Perspectives

The industry's reaction to the release of GPT-5 has been enthusiastic, but there are differing opinions:

  • Positive Feedback:

    • OpenAI CEO Sam Altman called GPT-5 "a Ph.D.-level expert," performing best in programming, writing, and health fields, marking OpenAI's return to the industry's leading position.
    • Aaron Levie, CEO of Box, stated that GPT-5 has achieved "breakthrough progress" in handling complex documents and logical reasoning tasks, especially in enterprise-level data extraction tasks.
    • Microsoft expressed high recognition of GPT-5's integration, believing that its reasoning capabilities and efficiency will significantly enhance the user experience of Microsoft 365 Copilot and GitHub Copilot.

  • Criticism and Doubts:

    • Some observers believe that the advancement of GPT-5 compared to GPT-4 is not as significant as the leap from GPT-3 to GPT-4, failing to fully meet the expectations of a "quantum leap."
    • In abstract reasoning tests such as ARC-AGI, GPT-5's performance did not meet expectations, showing limitations in certain cutting-edge tasks.
    • MIT Technology Review pointed out that GPT-5's user experience improvements are significant, but "good experience alone" cannot achieve Altman's promise of an automated future, and it still has a long way to go to reach AGI.
    • Some users question its pricing strategy, believing that high API costs may aim to restrict competitors (such as DeepSeek) from distilling the model.

  • Industry Competition:

    • GPT-5 faces fierce competition from models such as Anthropic's Claude, Google's Gemini, and Meta's LLaMA. Especially in the programming assistant market, Claude's precision and local context processing capabilities are considered advantageous.
    • DeepSeek's R1 model, with lower costs and comparable reasoning capabilities, has become a strong competitor to OpenAI.

VII. Summary and Outlook

The release of GPT-5 marks a comprehensive advancement in OpenAI's AI technology, user experience, and popularization strategy. Its unified system architecture, strong reasoning capabilities, and multimodal support make it perform well in areas such as programming, mathematics, health, and content generation. Through the free tier and multi-tier pricing strategy, OpenAI has further lowered the barrier to AI access, allowing more users and developers to benefit from cutting-edge technology.

However, GPT-5 is not without limitations. Its performance in abstract reasoning tasks did not fully meet expectations, and the knowledge cutoff date may affect its ability to handle real-time information. Moreover, the industry still has doubts about whether it can significantly surpass its competitors, especially when competing with models like Claude and Gemini.

Looking ahead, the widespread application of GPT-5 will drive the deeper integration of AI in education, healthcare, enterprise management, and software development. OpenAI plans to further expand its multimodal capabilities (such as video processing) and continuous learning functions to get closer to the goal of AGI. At the same time, industry competition and ethical issues will remain challenges that GPT-5 must face during its development process.

This article is compiled based on publicly available information before August 8, 2025. Information may change over time.