Claude Opus 4.7: The Behavior Changes We Actually Wanted
Anthropic just released Opus 4.7, and the behaviour changes of the new models are the ones we actually wanted. They address a range of long-standing issues many of us had. A TL;DR:
- Response length calibrated to task: Short answers to short questions, long answers where they earn it. No more fixed verbosity baseline.
- More literal instruction following: No silent generalization, no inferred asks. Better for structured pipelines and tuned prompts.
- More direct tone: Less validation-forward phrasing, fewer emoji. Opinionated where 4.6 was warm and more conversational.
- Built-in progress updates in agentic traces: You can delete “summarize every N tool calls” scaffolding since the model does it natively now.
- Fewer subagents by default: Steerable via prompt, but the default is leaner.
- Stricter effort calibration: low finally means low. Raise effort for complex tasks instead of prompt-hacking around an unpredictable dial.
- Fewer tool calls, more reasoning: Thinks before it reaches for grep. Less trash, better results in most cases.
- Real-time cybersecurity safeguards: May refuse high-risk topics. Legitimate security work routes through the Cyber Verification Program.
- High-resolution image support: Up to 2576px long edge (from 1568). Bounding-box coordinates are now 1:1 with image pixels — no scale conversion required.
The through-line: prompts get simpler, outputs get sharper, and agents get leaner. A lot of the CLAUDE.md directives people have been carrying around since 4.5 as quasi guardrails can probably be retired.
The literalism change also means a prompt that was slightly underspecified on 4.6 may now do exactly — and only — what it says.
The catch
But there is also a downside of the upgrade: Opus 4.7 ships a new tokenizer that uses roughly 1x–1.35x as many tokens per text as 4.6, depending on the task. Although pricing remains at $5/$25 per MTok, the effective cost per task is expected to rise on text-heavy workloads.