Claude Opus 4.7: More robust coding with less supervision.

Why this makes a difference for agentic workflows.

According to Anthropic, Opus 4.7 is substantially better at following instructions. That may sound like a small improvement, but the impact is significant in agentic workflows—situations where an AI model independently performs multiple steps in sequence, often using tools and CI pipelines. Previous models interpreted instructions loosely or skipped parts of them. Opus 4.7 takes instructions literally. Anthropic explicitly warns that existing prompts must be re-tuned, as old prompts may trigger unexpected behavior.

Clearer image, better memory.

The model now accepts images up to 2,576 pixels on the longest side, more than three times the size of previous Claude models. This opens up practical applications for computer-use agents that need to read dense screenshots and for data extraction from complex diagrams. Additionally, Opus 4.7 works better with file-system-based memory: it retains important notes across long, multi-session work sessions. For long-running GitLab pipelines or multi-step tickets in agent frameworks, this means less repeated context and more continuity.

New review: xhigh effort and /ultrareview.

Along with the model, Anthropic launched a new effort level called xhigh, which falls between high and max. This allows you to fine-tune the balance between reasoning depth and latency. In Claude Code, xhigh is now the default. Also new is the slash command /ultrareview, a dedicated review session that scans through changes and flags bugs or design issues that a careful reviewer would notice. For teams that rely on merge requests, this is a direct addition to existing code review practices.

What this means for development teams.

A model capable of executing long-running agentic tasks without crashing aligns directly with DevSecOps and CI/CD pipelines that can run for several hours. The reported improvements in tool error rates and the ability to continue processing despite tool errors make Opus 4.7 suitable for production agents rather than just experiments. For organizations already working with Claude, this means in practice: less manual intervention in automated workflows, more thorough code reviews via /ultrareview, and a significant step toward agents that independently push features from ticket to merge request. Combine that with the GitLab Duo Agent Platform and you get a chain where AI agents and pipelines work together rather than side by side.

Getting Started with Opus 4.7.

Are you already using Claude within your organization? If so, please re-tune your existing prompts, as the model interprets instructions more literally than its predecessors. Additionally, monitor token usage in real traffic, as Opus 4.7 uses a new tokenizer that can map input to 1.0 to 1.35 times as many tokens. Anthropic’s complete migration guide provides recommendations for each scenario.