What Changed in Claude 4

Claude 4 (Opus 4, Sonnet 4, then 4.5, 4.6, 4.7 across 2025-2026) reshaped Anthropic’s lineup. The headline: better agentic reliability, 1M context across the lineup, prompt caching that actually pays off, and Claude Code as a flagship product.

What’s new vs Claude 3.7

• 1M context across Sonnet + Opus. Haiku stays at 200k.
• Prompt caching: default 5-min, optional 1-hour TTL, 90% off cached input. Industry-leading.
• Top-tier SWE-bench Verified, Aider, Terminal-Bench scores. Best agentic coder.
• Claude Code: terminal-first agent, bundled with Pro/Max plans.
• Skills + Hooks: reusable behavior packages and event listeners.
• Computer Use API: agents that operate a desktop UI.

The model lineup in 2026

• Opus 4.7 — top reliability, 5x the cost of Sonnet.
• Sonnet 4.6 — default daily driver, 95% of Opus quality.
• Haiku 4.5 — budget tier, fast, full tool surface.

What got better in practice

• Long agentic loops (30+ steps) are far more reliable.
• Output style is cleaner, less marketing-flavored.
• Web search is competent (still behind Perplexity for grounded research).
• Projects (persistent context across conversations) make ongoing work much smoother.

Where Claude still loses

• No native image gen, no Sora-style video, no voice mode equivalent (yet).
• API price slightly higher than GPT-5 ($3/$15 vs $2.50/$10) — but caching narrows the gap.
• Pro plan ($20) caps tighter than ChatGPT Plus on heavy days.

Compare to alternatives at Claude vs ChatGPT and Claude vs Gemini. Pricing breakdown at Claude pricing explained.