OpenAI launches GPT-5 with built-in reasoning

OpenAI just launched GPT-5, bringing a massive upgrade to ChatGPT that combines quick responses with deep reasoning under one unified hood.

Instead of making users manually pick between a speedy conversational model and a heavy reasoning engine, GPT-5 uses an internal real-time router. The system automatically inspects your prompt, complexity level, tool requirements, and explicit instructions like "think hard about this" to decide whether to answer instantly or engage GPT-5 thinking.

Smarter, faster, and less chatty

Under the hood, GPT-5 sets new state-of-the-art benchmarks across the board. It hit 94.6% on AIME 2025 math without tools, 74.9% on SWE-bench Verified coding, and 84.2% on MMMU multimodal reasoning. With extended reasoning enabled in GPT-5 pro, it reached 88.4% on GPQA.

Crucially, it gets those results while burning far fewer tokens. In OpenAI's evaluations, GPT-5 thinking outperformed OpenAI o3 while generating 50% to 80% fewer output tokens across scientific problem-solving and coding tasks.

OpenAI also tackled model personality and accuracy. Compared to GPT-4o with web search, factual errors dropped by roughly 45%. When thinking, errors dropped around 80% compared to o3. It uses noticeably fewer unnecessary emojis, strips away sycophantic praise, and handles delicate queries with a new safe completions approach rather than blanket refusals.

GPT-5 is rolling out to all ChatGPT users, with higher caps for Plus and access to GPT-5 pro for Pro subscribers. Per reporting from 9to5Mac, Apple also confirmed that Apple Intelligence will upgrade its ChatGPT integration to GPT-5 next month alongside iOS 26, iPadOS 26, and macOS Tahoe 26.

Why it matters

For everyday users and builders, GPT-5 takes the guesswork out of model selection. You no longer have to manage your quota by manually rationing reasoning tokens; the model handles routing invisibly, making routine coding, drafting, and complex debugging feel far smoother.

Across the industry, this marks a shift from brute-force scale to routing efficiency. Slashing output tokens by up to 80% while outperforming o3 directly attacks the inference cost bottleneck, raising the bar for Google and Anthropic.

We think wrapping frontier speed and reasoning into a single automated router is a brilliant product move that makes ChatGPT significantly cleaner to use. The obvious risk, however, is router error. If the system misjudges a subtle problem and skips deep thinking, users may still find themselves typing explicit workarounds to force the model to behave.