Anthropic · 2025 · 05 · 22 · Model · ~1 min read
Anthropic launched Claude 4 (Sonnet 4 + Opus 4)
What's actually new
- Opus 4 is back. The top tier had been missing in 2024. Now it's the strongest general-purpose model on most tests.
- Long-running agent runs. Opus 4 stayed on a coding task for hours in the demo, then handed off properly.
- Better at sticking to instructions in long sessions — drift over multi-hour conversations dropped substantially.
If you want more
Worth knowing
- Opus 4 is expensive. $15 in / $75 out per million tokens makes long agent runs cost real money.
- Pinned 'best model on the planet' claims faded fast. OpenAI shipped GPT-5 within ~3 months.
Who should care
Developers shipping AI agents with Anthropic. Anyone evaluating coding assistants seriously. Researchers tracking long-context agent reliability. Teams that paused on agent products waiting for something more reliable.
What to do about it
If you've been holding off building an agent because models lose context after an hour, try Opus 4 on your task. Set a budget cap before you start.
Honest take
Claude 4 was the moment AI agents felt genuinely usable for multi-hour work, not just 5-minute demos. Opus 4's return ended the awkward 'Anthropic's top model is missing' year. The 'agent' wave that everyone had been promising for two years finally had a workhorse model behind it. The question turned from 'can the AI run for hours?' to 'how do we trust what it did?' — a much better problem to have.
Sources
Last verified · 2026 · 05 · 05 · Found a fact wrong? corrections@aguidetocloud.com