gpt-5-2025-08-07

Basic Information

• Release Date: The GPT-5 model was officially launched at 1:00 a.m. (Beijing Time) on August 8, 2025, and OpenAI rolled out API access permissions to developers worldwide.
• Model Portfolio: Initially released in August 2025, the model lineup includes GPT-5, mini, nano, and Chat versions. The core architecture features a dual-track system—gpt-5-main (fast response) and gpt-5-thinking (deep reasoning)—with a Pro version available for subscribers. On September 16, an independent programming model, GPT-5-Codex, was added, specifically optimized for coding tasks.
• Basic Parameters: The total context window length is 400K tokens (including 272K tokens for input and 128K tokens for output; the output includes invisible reasoning tokens). GPT-5’s input cost is $1.25 per million tokens, while the output cost is $10 per million tokens. Input costs are 50% lower than those of GPT-4o.

Core Features

• Leading Performance Across Multiple Domains: In the SWE-bench Verified programming test, GPT-5 achieved an accuracy rate of 74.9% for the first time. GPT-5 Pro scored 89.4% on the GPQA Diamond-level doctoral test, attained an accuracy rate of 94.6% without tool assistance in the 2025 AIME math competition, and scored 46.2% on the HealthBench Hard test.
• Extremely Low Hallucination Rate: The fact-checking error rate when connected to the internet is 45% lower than that of GPT-4o. The gpt-5-thinking version has a self-reasoning error rate 65% lower than that of the o3 model. Major response-level errors have been reduced by 78%, and the model openly acknowledges task limitations.
• Outstanding Agent Capabilities: Supports reusing inference contexts via the Responses API, significantly improving the efficiency of complex tool chains. Its ability to fix defects in coding scenarios surpasses competitors, with a success rate of 96.7% in the τ²-bench telecom tool chain test.

Technical Highlights

• Unified Architecture: Integrates GPT series generation capabilities with the o series reasoning components, adopting a dual-track design of “fast model + deep reasoning model.” Performance and efficiency are balanced through adjustable reasoning levels, reducing input costs by 50% compared to previous generations.
• Real-Time Routing System: Automatically analyzes task complexity and dynamically switches response modes, eliminating the need for users to manually switch models.
• Security Mechanisms: Employs a “safe completion” strategy instead of hard rejection. After more than 9,000 hours of red-team testing—including 400 external testers—the model’s security has improved in scenarios such as violent attack planning, and the rate of flattery responses has dropped below 6%.

Market Impact

• Brand Positioning: At the time of its August 2025 release, the model was valued at $300 billion. Concurrently, negotiations were initiated for employee stock sales, aiming for a valuation of $500 billion. It ranked first in all categories in LMArena’s tests.
• Ecosystem Integration: Simultaneously integrated with platforms such as Microsoft Copilot, Microsoft 365, and Azure AI Foundry.
• Industry Transformation: Driving AI from a mere tool toward an “industry operating system,” marking the inaugural year of AI agent capability platformization.

Application Scenarios

• Programming Development: Leveraging the independently released GPT-5-Codex model on September 16, developers can efficiently create responsive websites, apps, and 3D games, supporting continuous complex project iterations for up to 7 hours. GitHub paid users can integrate it for code review, achieving an accuracy rate of 74.5% in SWE-bench Verified tests. A live demonstration showed that a French-language learning website could be developed in just 3 minutes.
• Professional Fields: Assists in medical result analysis, doctoral-level scientific research, financial report analysis, and PPT generation.
• Education and Interaction: The Chat version supports language learning and offers four preset personalized interaction modes—cynic, robot, listener, and top student—to meet individualized learning needs.

Related Review: “GPT-5 Review: Failed to Blow Up the Market, But Precisely Slapped Competitors in the Face: Cheap, Powerful, and No Bullshit”