What's New in Claude Opus 4.7 (High-Level Tour)

The headline changes

Claude Opus 4.7 is Anthropic's most capable generally available model. Compared to Opus 4.6 the user-visible changes are:

1. Adaptive thinking replaces extended thinking budgets. You no longer set budget_tokens. The model decides how much to reason; you steer with the effort parameter.
2. Task budgets (beta) — give the model an advisory token budget for a full agentic loop. It sees a running countdown and paces itself.
3. High-resolution vision — maximum image resolution jumps from 1568px to 2576px (3.75MP). 1:1 pixel-to-coordinate mapping removes scale-factor math.
4. New xhigh effort level — recommended for coding and agentic work.
5. New tokenizer — same prompt can now use up to ~35% more tokens, driving better reasoning but needing max_tokens headroom.
6. 1M context at standard pricing — no long-context premium.
7. Sampling parameters removed — temperature, top_p, top_k now return 400 if set.

Behaviour changes worth memorising

• More literal instruction following at low effort.
• Response length calibrates to task complexity.
• Fewer tool calls and subagents by default; raise effort to get more.
• More direct, opinionated tone with less validation.
• Real-time cybersecurity safeguards on high-risk topics.

We spend the rest of the course unpacking each of these.