🚀 OpenAI launches GPT-5 for all ChatGPT users with "PhD-level" capabilities 🚀
The most anticipated AI model of 2025 is finally here

OpenAI today released GPT-5, marking what CEO Sam Altman calls the biggest leap since the iPhone's transition to Retina display. The new model is available immediately to all ChatGPT users, including free tier users, representing the first time a reasoning model has been accessible without payment barriers. Altman described the experience as having "a team of Ph.D. level experts in your pocket," emphasizing the model's ability to provide expert-level responses across mathematics, science, finance, law, and healthcare domains.

The launch comes as OpenAI approaches 700 million weekly active users and enters discussions for a potential stock sale at a $500 billion valuation. GPT-5 underwent 5,000 hours of safety testing and demonstrates significant improvements in hallucination reduction, with a 26% lower error rate compared to GPT-4o and a 65% reduction in the thinking version compared to o3.

💰 GPT-5 pricing strategy undercuts competition by massive margins 💰
Claude Opus 4.1 costs 12x more than GPT-5 for comparable performance

OpenAI has positioned GPT-5 as an aggressively competitive offering with pricing that significantly undercuts major competitors. At $1.25 per million input tokens and $10 per million output tokens, GPT-5 costs exactly half the input price of GPT-4o while maintaining the same output pricing. The comparison with Anthropic's Claude Opus 4.1 is particularly striking, with Claude priced at $15 input and $75 output per million tokens—making it 12 times more expensive than GPT-5 for similar capabilities.

The pricing strategy extends across the entire GPT-5 family: GPT-5 Mini at $0.25/$2.00 and GPT-5 Nano at $0.05/$0.40 per million tokens. OpenAI also introduced a 90% discount on cached tokens used within the previous few minutes, making conversational applications significantly more cost-effective. This pricing approach appears designed to capture market share while the company scales toward its ambitious revenue targets.

🔬 GPT-5 sets new benchmarks in coding and health use cases 🔬
Model achieves 74.9% on SWE-Bench,

GPT-5 has established new performance standards across critical benchmarks, particularly in software engineering and healthcare applications. The model scored 74.9% on SWE-Bench Verified, slightly edging out Claude Opus 4.1's 74.5% while being significantly more cost-effective. Additional coding benchmarks show impressive results: 55% on SWE-Lancer for freelance-style coding tasks and 88% on Aider Polyglot for multi-language programming capabilities.

Healthcare performance represents a particular focus area, with GPT-5 achieving substantial improvements on HealthBench evaluations. The thinking version scored 46.2% on HealthBench Hard, up from o3's 31.6%, with all scores validated by multiple physicians. OpenAI specifically highlighted health-related questions as one of ChatGPT's three most common use cases, alongside writing and coding, explaining the concentrated effort in this domain.

TECHNICAL DEEP DIVE

🧠 GPT-5 architecture introduces smart model routing 🧠

GPT-5 represents a significant architectural evolution from previous OpenAI models, implementing what the company calls a "unified system" that intelligently routes queries between different specialized models. The system includes a smart and fast model for routine questions, a deeper reasoning model for complex problems, and a real-time router that analyzes conversation type, complexity, tool requirements, and explicit user intent to determine the optimal model selection.

This hybrid approach allows GPT-5 to dynamically adjust its computational resources based on query complexity. When users include phrases like "think hard about this" in their prompts, the system automatically engages the deeper reasoning model. The API implementation offers more granular control with three distinct models - regular, mini, and nano - each capable of operating at four reasoning levels: minimal, low, medium, and high.

🛡️ "Safe completions" replace binary refusal safety system 🛡️

OpenAI has introduced a groundbreaking safety approach called "safe completions" that fundamentally changes how AI models handle potentially risky queries. Instead of the traditional binary system of either providing full assistance or complete refusal, GPT-5 aims to maximize helpfulness while staying within safety constraints. This approach is particularly valuable for dual-use cases in fields like biology or cybersecurity, where high-level information can be provided safely without including specific details that could enable harmful applications.

The safe completions system represents a shift from intent-based safety (analyzing what the user wants) to output-centric safety (ensuring the response itself is safe). This allows GPT-5 to provide educational or informational responses to sensitive topics while avoiding the provision of actionable harmful content. The approach required extensive training and represents one of the most significant safety innovations in large language model development.