Anthropic quietly released Claude Opus 4.7 on April 16, and after spending the past ten days putting it through its paces on production workloads, I can confirm this is the most capable Claude model to date. The improvements are not incremental. This release pushes the frontier in agentic coding, high-resolution vision, and enterprise document processing in ways that meaningfully change what you can delegate to an AI system.
Coding Performance That Actually Matters
The headline number is 87.6% on SWE-bench Verified, up from 80.8% on Opus 4.6. That seven-point jump represents real-world capability: the model can now resolve significantly more actual GitHub issues without human intervention. More impressively, SWE-bench Pro (which tests harder, multi-file problems) jumped from 53.4% to 64.3%. This is the largest coding improvement in a single Claude generation.
What does this mean practically? I have been using Claude Opus 4.7 through Claude Code for complex refactoring tasks that previously required multiple rounds of supervision. The model now handles long-running tasks with consistency, maintaining context across extended coding sessions without the drift that plagued earlier versions. The new xhigh effort level (available in the API) gives the model more compute budget for difficult problems, and the difference is noticeable on architectural decisions and complex debugging.
CursorBench performance also improved from 58% to 70%, which matters if you use Cursor as your primary IDE. The model better understands the context of your entire codebase and generates more appropriate completions.
Vision Gets a 3x Resolution Upgrade
Perhaps the most underappreciated improvement is vision. Claude Opus 4.7 now supports images up to 2,576 pixels on the long edge (approximately 3.75 megapixels), more than triple the previous 1.15 megapixel limit. This sounds like a spec sheet detail until you try processing technical diagrams, architectural blueprints, or dense spreadsheet screenshots.
The practical difference is substantial. I recently tested the model on chemical structure diagrams and circuit schematics that would have been illegible to Opus 4.6. The model now extracts accurate information from high-resolution technical documents without requiring manual cropping or pre-processing.
OSWorld-Verified, which measures computer use capabilities, improved from 72.7% to 78.0%. For teams building agents that interact with desktop applications or browser interfaces, this means fewer failures on visual element detection and interaction.
Enterprise Document Processing
The finance and legal capabilities deserve attention. Opus 4.7 achieves state-of-the-art performance on finance agent evaluations and scores 90.9% accuracy on BigLaw legal tasks. Document reasoning errors dropped by 21% compared to the previous version.
These improvements matter for organizations processing contracts, financial reports, or regulatory filings. The model better follows complex multi-step instructions and maintains accuracy across long documents. Combined with the vision improvements, you can now feed it dense PDF tables and trust the extracted data.
What This Means for AI Practitioners
The pricing remains unchanged at $5 per million input tokens and $25 per million output tokens. Anthropic held the line on costs while delivering meaningful capability improvements across the board.
For those of us building with Claude, the practical implications are clear. Tasks that required supervision can now run autonomously. Vision-based workflows that needed preprocessing can work directly on source documents. Long-horizon agentic tasks fail less often.
The model is available across all Claude products, the API (model ID: claude-opus-4-7), Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry. If you have existing Opus 4.6 workloads, the migration is straightforward since the API surface is unchanged.
Looking Ahead
The gap between Claude Opus 4.7 and GPT-5.5 (which scored 82.7% on Terminal-Bench 2.0 compared to Claude's 69.4%) suggests we are in a period of rapid capability leapfrogging. OpenAI leads on some benchmarks, Anthropic on others. The practical implication is that model selection now depends heavily on your specific use case rather than overall "best model" rankings.
For coding-heavy workloads, particularly those involving complex multi-file changes and long-horizon tasks, Claude Opus 4.7 is currently the strongest option. For agentic workflows that require extensive tool use and autonomous operation, GPT-5.5 may have the edge. The right choice depends on what you are building.
What is clear is that April 2026 represents a step function in AI coding capabilities. The models we have today can genuinely handle work that required human engineers a year ago. The question is no longer whether AI can help with coding, but how to restructure engineering workflows to take full advantage of what these systems can do.
Sources: