LLM Wiki

Appearance

Built 26/04/16 17:00commit 95b5faf

Introducing Claude Opus 4.7 ​

中文 | English

Summary ​

This source is Anthropic's launch announcement for Claude Opus 4.7. For this vault, its main value is not just the model release itself, but the operator-facing changes bundled with it: stronger instruction following, better long-running coding performance, improved file-system memory use, a new xhigh effort level, the /ultrareview Claude Code review surface, and a staged cyber-safeguard rollout positioned below Mythos Preview.

Source ​

Key Contributions ​

  • Records Anthropic's claim that Opus 4.7 improves difficult coding and long-running agentic work relative to Opus 4.6.
  • Makes instruction-following drift operationally visible by warning that prompts tuned for older models may need retuning because the new model obeys instructions more literally.
  • Adds a concrete Claude Code control-surface change: default effort now shifts to xhigh, and /ultrareview becomes a dedicated review-session command.
  • Highlights improved file-system memory and higher-resolution vision as practical upgrades for long, multimodal, tool-using work.
  • Connects the model release to staged cyber-governance: Opus 4.7 is the first public deployment target for new cyber safeguards developed in response to Mythos Preview.
  • Documents migration-relevant token-usage risk from tokenizer changes and heavier later-turn reasoning.

Practical Implications ​

  • Model upgrades can invalidate prompt and harness assumptions even when the migration is framed as a direct successor release.
  • Claude Code operating practice should track model-release notes because meaningful workflow changes now land as part of model announcements, not only Claude Code docs.
  • Effort controls, review commands, and file-system memory behavior belong in the same operating picture as benchmark deltas.
  • Anthropic's staged release pattern matters: less capable public models may be used to field-test safeguards before broader release of stronger capability tiers.