What’s New in Claude Opus 4.7? Deep Dive into Capabilities and Migration Tips

Anthropic has officially released Claude Opus 4.7, making the model available to all users and via API at the same pricing as Opus 4.6 (input $5 / M tokens, output $25 / M tokens).

Core upgrade focus : Opus 4.7 is designed for high‑difficulty, long‑cycle tasks. It executes complex work more rigorously, follows instructions with greater precision, and proactively verifies its output before responding, allowing users to hand over hard‑core programming tasks without close supervision.

Specific improvements include a three‑fold increase in visual resolution (supporting up to 2576 px on the long side, ~3.75 MP) for high‑resolution screenshot analysis and complex chart extraction, stricter literal instruction compliance (which may change the behavior of prompts written for older versions), and benchmark gains across finance agents, code generation, document reasoning, and long‑context scenarios. File‑system memory is also enhanced, retaining key notes across multi‑turn long sessions.

New features released : a new effort level xhigh sits between high and max, giving developers finer control over reasoning latency; Claude Code now defaults to xhigh. The API introduces a public beta for Task Budgets to help manage token consumption on long tasks, and a new /ultrareview command offers specialized code review with three free uses for Pro/Max users.

Security updates : Opus 4.7 is among the first models deployed under Anthropic’s Project Glasswing with new network‑security safeguards. Compared with Claude Mythos Preview, its offensive capabilities are deliberately throttled, and it includes built‑in automatic detection and blocking mechanisms. Security professionals conducting legitimate penetration testing can apply to the Cyber Verification Program.

Migration considerations : The model uses an updated tokenizer, causing token consumption for identical inputs to rise to roughly 1.0–1.35 × that of Opus 4.6. High‑effort, multi‑turn agent scenarios may also see higher output token counts. Users should benchmark real traffic before a full rollout. The API model identifier is claude-opus-4-7.