Model ReleaseAug 2025
You can start trusting what AI says
The shift
Assume AI is wrong until proven right→ Verify key claims, trust the rest
How it happened
Jan
DeepSeek R1 shows reasoning reduces errors
Chain-of-thought reasoning dramatically cut fabricated answers
Mar
Claude adds citations
Anthropic ships source attribution — model links claims to documents
Jun
GPT-4o gets web search built in
Real-time fact-checking during generation, not just after
Aug
GPT-5 launches with 80% fewer hallucinations
Internal benchmarks show step-change in factual accuracy
The players
OpenAI
GPT-5 made the biggest single accuracy jump
Anthropic
Led on source citations and honest uncertainty
DeepSeek
Proved reasoning chains reduce hallucinations
What you used to think
- "AI just makes stuff up — I can't rely on it"
- "I have to fact-check every single sentence"
- "I'd never use AI for anything important"
What's actually true now
- ✓GPT-5 cut hallucinations by 80% — most output is now accurate
- ✓Models can search the web and cite sources on their own
- ✓People use AI for legal briefs, medical summaries, financial analysis
What to do about it
Use AI for first drafts of important work. Spot-check key facts and numbers, but stop rewriting everything from scratch. The baseline quality has changed.