OpenAI GPT-5 Release Notes

OpenAI has introduced GPT-5, a unified AI system that represents a significant leap in intelligence over previous models. By integrating a real-time router that dynamically selects between a fast, efficient model and a deeper reasoning model (GPT-5 thinking), the system provides state-of-the-art performance across coding, mathematics, writing, and health.

Unified System Architecture

GPT-5 operates as a unified system comprising three core components: a smart, efficient model for standard queries, a deeper reasoning model known as "GPT-5 thinking" for complex problems, and a real-time router. The router determines the appropriate model based on conversation complexity, tool requirements, user intent (e.g., prompts like "think hard about this"), and conversation type.

To ensure continuous availability, a "mini" version of each model handles queries once primary usage limits are reached. OpenAI intends to integrate these separate capabilities into a single model in the future.

Technical Capabilities and Domain Performance

GPT-5 achieves new state-of-the-art (SOTA) results across several critical benchmarks, demonstrating improved intelligence in specialized fields:

  • Mathematics: 94.6% on AIME 2025 (without tools).
  • Coding: 74.9% on SWE-bench Verified and 88% on Aider Polyglot.
  • Multimodal Understanding: 84.2% on MMMU.
  • Health: 46.2% on HealthBench Hard.
  • General Science: GPT-5 pro achieves 88.4% on GPQA without tools.

Coding and Development

GPT-5 is OpenAI's strongest coding model to date, with specific advancements in debugging large repositories and complex front-end generation. The model demonstrates an intuitive understanding of aesthetic design, including typography and spacing, allowing it to generate responsive websites and apps from single prompts.

Health and Medical Information

GPT-5 scores significantly higher on HealthBench than previous models. It functions as an active thought partner by proactively flagging concerns and asking clarifying questions. The model adapts responses based on the user's geography, knowledge level, and context to provide safer, more reliable information.

Creative Writing and Expression

The model shows improved capability in handling structural ambiguity and literary depth, such as sustaining free verse or unrhymed iambic pentameter, making it more effective for drafting professional reports, emails, and memos.

Reliability, Factuality, and Honesty

GPT-5 introduces significant reductions in hallucinations and deceptive behavior compared to its predecessors:

  • Factuality: Responses are approximately 45% less likely to contain factual errors than GPT-4o. When using the "thinking" mode, GPT-5 is approximately 80% less likely to contain factual errors than OpenAI o3.
  • Open-Ended Factuality: On LongFact and FActScore benchmarks, "GPT-5 thinking" produced six times fewer hallucinations than o3.
  • Honesty: In tests where images were removed from multimodal prompts (CharXiv), GPT-5 gave confident answers about non-existent images only 9% of the time, compared to 86.7% for o3. Deception rates in production traffic dropped from 4.8% (o3) to 2.1% (GPT-5).

Safety and Behavioral Refinements

Safe Completions Paradigm

OpenAI has moved beyond simple refusal-based safety training to a "safe completions" approach. This method teaches the model to provide the most helpful answer possible within safety boundaries, which may involve partial answers or high-level overviews. If a refusal is necessary, the model is trained to explain the reason and offer safe alternatives.

Reduction of Sycophancy

GPT-5 is designed to be less effusively agreeable and use fewer unnecessary emojis. In targeted evaluations, sycophantic replies were reduced from 14.5% to less than 6%, aiming for a tone that resembles a "helpful friend with PhD-level intelligence."

Biological Risk Safeguards

Because the "GPT-5 thinking" model is classified as "High capability" in the Biological and Chemical domain, OpenAI implemented a multilayered defense system. This includes 5,000 hours of red-teaming with partners like CAISI and UK AISI, always-on classifiers, and reasoning monitors to minimize biological risks.

GPT-5 Pro and Access Tiers

GPT-5 pro replaces OpenAI o3-pro. It utilizes scaled parallel test-time compute to provide extended reasoning for the most complex tasks. In evaluations of 1,000 economically valuable reasoning prompts, external experts preferred GPT-5 pro over "GPT-5 thinking" 67.8% of the time, noting 22% fewer major errors.

Availability:

  • Free Users: Access to GPT-5, transitioning to GPT-5 mini after usage limits are reached.
  • Plus, Team, and Pro Users: Higher usage limits. Pro subscribers receive unlimited access to GPT-5 and exclusive access to GPT-5 pro.
  • Enterprise and Edu: Access rolling out the following week.
  • Developer Access: Available via the Codex CLI for Pro, Plus, and Team users.

Sources