What Changed in GPT-5

GPT-5 (released August 2025) is the biggest practical leap from GPT-4o. Beyond the headline benchmarks, the real changes that affect day-to-day use are the reasoning router, voice + multimodal, and the API restructure.

The headline changes

• Reasoning router: GPT-5 picks fast vs slow thinking automatically. You don’t need to flip a model.
• 400k context: 4x GPT-4o, smaller than Claude (1M) and Gemini (2M) but plenty for most.
• Pricing drop: $2.50/$10 per 1M tokens vs GPT-4o’s same price — but with reasoning routed automatically.
• Mini + nano tiers: $0.25/$2 and $0.05/$0.40 dramatically expand cheap-tier usage.
• Atlas + Operator: agentic browsing built in.
• Sora 2: video generation rolled into ChatGPT for Plus / Pro.

What got better in practice

• Fewer obvious hallucinations on factual queries.
• More consistent instruction-following on long prompts.
• Voice mode (Advanced Voice) now feels conversational, not turn-based.
• Vision: better at reading dense text (PDFs, screenshots).

What didn’t change as much as marketing implied

• Coding: still trails Claude Sonnet/Opus on SWE-bench. Solid for autocomplete, second to Claude on agents.
• Memory: still has cross-session leakage if not pruned.
• Long context: 400k is generous but not always reliable past ~200k.

Run the cost math at the Gemini vs ChatGPT cost calculator. Compare to Claude at Claude vs ChatGPT.