OpenAI rolls out GPT-5 to all ChatGPT users

OpenAI officially unveiled GPT-5 on Thursday, introducing what CEO Sam Altman called "PhD-level experts in your pocket" and marking the company's most ambitious attempt yet to unify artificial intelligence capabilities into a single, adaptive system. The new model, now the default for all ChatGPT users including those on free plans, promises fewer hallucinations, enhanced coding abilities, and automatic routing between fast responses and deeper reasoning modes.
The launch represents a shift in OpenAI's approach to AI deployment. Unlike previous releases that required users to choose between different models for various tasks, GPT-5 features a unified architecture that automatically determines whether to provide quick answers or engage deeper reasoning capabilities based on query complexity. According to OpenAI, GPT-5 with web search enabled produces 45% fewer factual errors than GPT-4o, while its reasoning mode reduces errors by 80% compared to the company's previous o3 model

Enhanced Performance Across Key Domains

GPT-5 demonstrates state-of-the-art performance across multiple benchmarks, achieving 74.9% on SWE-bench Verified for coding tasks and 88% on Aider Polyglot. The model particularly excels at what developers call "vibe coding" - generating complete software applications from single prompts with minimal user guidance. Early testing partners including Cursor, Windsurf, and Vercel reported improved code quality and reduced error rates compared to previous models.
The model also shows improvements in health-related queries, scoring 46.2% on HealthBench Hard compared to 31.6% for o3. OpenAI emphasizes that while GPT-5 can help users understand medical information and formulate questions for healthcare providers, it remains a complement to, not replacement for, professional medical advice.
In mathematical reasoning, GPT-5 achieved 94.6% on AIME 2025, while GPT-5 Pro reached 88.4% on GPQA, a benchmark for challenging science questions. The model's reasoning capabilities extend beyond traditional benchmarks, with OpenAI claiming it's the first to exceed human performance on SimpleBench, scoring 90% against an average human score of 83%.
openai.com favicontomsguide.com favicongithub.blog favicon8 sourcesUnified System Architecture and AccessibilityThe new model operates through a real-time router that selects between a fast, efficient model for routine queries and a deeper reasoning model for complex problems. This system continuously improves through user feedback signals, including preference rates and measured correctness. Users can explicitly trigger reasoning mode by including phrases like "think hard about this" in their prompts.
OpenAI is making GPT-5 available across all user tiers, though with varying usage limits. Free users get access until reaching a cap, after which they're switched to GPT-5 mini. Plus subscribers receive higher usage limits, while Pro users enjoy unlimited access plus GPT-5 Pro with extended reasoning capabilities. Enterprise and education customers will gain access next week.
The model introduces new safety measures called "Safe Completions," which aim to provide helpful responses within safety constraints rather than refusing queries outright. According to OpenAI's safety research lead Alex Beutel, this approach resulted in both safer and more helpful interactions during testing..
Post By Jay Nirmal