OpenAI unveils GPT-5 with intelligent routing and enhanced reasoning


OpenAI has introduced GPT-5, its most advanced AI system to date, designed to bring expert-level intelligence to everyone. This new model marks a significant leap over previous generations with enhanced performance across coding, math, writing, health, visual perception, and more.

GPT-5: Unified System with Intelligent Routing

GPT-5 operates as a unified system combining:

  • A smart, efficient base model that answers most queries quickly.
  • A deeper reasoning model (“GPT-5 thinking”) for complex problems.
  • A real-time router that automatically selects the best model based on conversation type, complexity, tools needed, and explicit user intent (e.g., when prompted with “think hard about this”).

The router continuously learns from real user data such as model switches, response preferences, and accuracy, improving its decisions over time.

Enhanced Performance and Real-World Usefulness

GPT-5 outperforms all earlier models on key benchmarks and provides faster, more useful responses in everyday scenarios. The model shows significant progress in reducing hallucinations, improving how it follows instructions, and minimizing sycophantic behavior. Its strengths are especially notable in writing, coding, and health-related tasks.

Advanced Coding Capabilities

GPT-5 is OpenAI’s strongest coding model to date, excelling at:

  • Complex front-end code generation.
  • Debugging large repositories.
  • Creating responsive websites, apps, and games with a refined sense of spacing, typography, and white space.

Early users highlighted its intuitive design choices that help translate ideas into polished digital products from just one prompt.

Superior Creative Writing Support

As a writing collaborator, GPT-5 helps shape and polish ideas into compelling, rhythmically rich text, handling challenging forms like unrhymed iambic pentameter or free verse. It’s also improved in everyday writing assistance such as drafting and editing reports, emails, and memos.

Best-in-Class Health Guidance

GPT-5 is the best OpenAI model yet for health-related queries. It scored significantly higher on HealthBench, a physician-defined evaluation, acting as an active thought partner by flagging concerns and asking clarifying questions.

GPT-5 modifies its health responses based on the user’s context, knowledge, and geography to provide safer, more accurate information. The company warned it is meant to assist, not take the place of, medical professionals.

Benchmark Excellence

GPT-5 sets new standards with scores like:

  • 94.6% on the AIME 2025 math benchmark without tools.
  • 74.9% on SWE-bench Verified and 88% on Aider Polyglot coding tests.
  • 84.2% on MMMU for multimodal reasoning.
  • 46.2% on HealthBench Hard.
  • 88.4% on GPQA by GPT-5 pro for complex science questions.

Improved Instruction Following and Tool Use

The model shows strong gains in:

  • Handling multi-step instructions.
  • Coordinating across multiple tools.
  • Adapting to changing contexts for complex, evolving tasks.

This means GPT-5 follows user instructions more faithfully and completes more work end-to-end using its tools.

Advanced Multimodal Understanding

GPT-5 excels at interpreting and reasoning over images, video, spatial data, and scientific information. This enables accurate understanding of charts, photos, diagrams, and more.

Performance on Economically Important Tasks

On an internal benchmark of complex knowledge work across 40+ occupations (law, logistics, sales, engineering), GPT-5 matches or exceeds expert performance in roughly half the cases, outperforming previous models like OpenAI o3 and ChatGPT Agent.

Efficiency and Speed

Trained on Microsoft Azure AI supercomputers, GPT-5 achieves more with less thinking time, requiring 50-80% fewer output tokens than OpenAI o3 across tasks like visual reasoning, coding, and scientific problem solving.

Hallucination Reduction and Enhanced Factuality

With web search enabled on anonymized real-world prompts:

  • Compared to GPT-4o, GPT-5 is roughly 45% less prone to factual mistakes.
  • When “thinking,” they are about 80% less likely to have errors than OpenAI o3.

On public factuality benchmarks (LongFact, FActScore), GPT-5 thinking shows about six times fewer hallucinations than o3, a major advance in long-form content accuracy.

Honest and Transparent Responses

GPT-5 is better at:

  • Recognizing impossible or underspecified tasks.
  • Clearly communicating its limitations and refusals.

In tests, it confidently guessed about non-existent images only 9% of the time, compared to 86.7% for OpenAI o3. Overall, GPT-5 reduced deception rates from 4.8% (o3) to 2.1%, improving trustworthiness.

Safer, More Nuanced Safety Training

Moving beyond refusal-only safety, GPT-5 introduces “safe completions” — aiming to give the most helpful, safe answers possible, sometimes partially or at a high level. When refusal is necessary, it transparently explains why and offers safe alternatives. This approach better handles ambiguous intent and sensitive dual-use topics like virology.

Reduced Sycophancy and Refined Style

Compared to GPT-4o, GPT-5 is:

  • Less effusively agreeable.
  • Uses fewer unnecessary emojis.
  • More thoughtful and subtle in responses.

Sycophantic replies dropped from 14.5% to under 6%, balancing honesty and user satisfaction.

Comprehensive Biological and Chemical Safety

GPT-5 thinking is classified as High capability in biological and chemical domains, protected by:

  • Over 5,000 hours of red-teaming with partners.
  • A multilayered safety stack including threat modeling, safe completions training, continuous classifiers, and enforcement pipelines.

Precautionary safeguards are active to minimize risks.

New Ways to Customize ChatGPT

GPT-5 supports improved instruction-following and introduces a research preview of four preset personalities—Cynic, Robot, Listener, and Nerd—allowing users to tailor interactions without custom prompts. Users can opt-in, adjust anytime, and they’re built to reduce sycophantic replies.

GPT-5 Pro Variant for Complex Tasks

GPT-5 pro, replacing OpenAI o3-pro, uses longer, efficient parallel compute to provide the highest quality answers. It outperforms GPT-5 thinking on hard benchmarks and is preferred by experts 67.8% of the time, with 22% fewer major errors in health, science, math, and coding.

How to Use GPT-5

GPT-5 is now the default model for signed-in ChatGPT users, replacing GPT-4o, OpenAI o3, o4-mini, GPT-4.1, and GPT-4.5. It automatically applies reasoning where beneficial. Paid users can select “GPT-5 Thinking” or prompt for deep reasoning. Free users get limited usage and may switch to GPT-5 mini, which is smaller and faster.

Availability and Access

Rollout began August 7, 2025, for Plus, Pro, Team, and Free users; Enterprise and Education users gain access shortly after.

  • Coding with GPT-5 is available via Codex CLI.
  • Pro subscribers get unlimited access and GPT-5 Pro; Plus and Team users have higher usage limits than free users.

OpenAI plans to integrate GPT-5’s capabilities into a single, unified model in the near future.

2
0
i
Rate This