Claude Opus 4.7 Complete Guide — SWE-bench 87.6%, Vision 98.5% & New xhigh Effort Mode [April 16, 2026 Release]

Claude Opus 4.7 is Anthropic's commercially available top-tier model, officially released on April 16, 2026. Positioned as 'smarter, more agentic, and more precise' than Opus 4.6, it achieves SWE-bench Verified 87.6% — surpassing Gemini 3.1 Pro's 80.6%. Pricing remains unchanged from Opus 4.6 (input: $5 per million tokens, output: $25 per million tokens), and it is immediately available across all Claude products and APIs.

Here is a side-by-side comparison of key benchmarks between Opus 4.6 and Opus 4.7.

The 'xhigh' level newly introduced in Opus 4.7 sits between the existing 'high' and 'max' levels. It allows fine-grained tuning of the trade-off between reasoning depth and response speed, making it ideal for complex engineering problems and tasks requiring deep analysis. It offers deeper reasoning than 'high' while being faster than 'max' — a well-balanced option for demanding workloads.

All five Effort Control levels and their recommended use cases:

Opus 4.7's vision accuracy jumped 44 percentage points from 54.5% to 98.5%, reaching a truly practical level. It now supports images up to 2,576 pixels (3.75 megapixels) — more than three times the resolution of the previous version. This enables highly accurate processing of screenshots, UI mockup interpretation, chart and diagram reading, and complex technical drawing analysis.

Opus 4.7 strengthens agentic capabilities across three axes. (1) Long-horizon autonomy: resolves complex multi-step tasks without human intervention three times more often. (2) Systems engineering: improved ability to understand, design, and implement large codebases. (3) Complex code reasoning: solved 4 tasks in the internal 93-task benchmark that previous versions could not. Multi-turn tool-calling improved by +13 points without tools and +6 points with tools.

Opus 4.7 significantly improves its ability to interpret instructions literally. Key improvements include: (1) better handling of ambiguous instructions, (2) more thorough problem-solving, (3) avoiding unsolicited generalization of instructions, and (4) not inferring and executing things the user did not ask for. This greatly improves predictability in 'vibe coding' (intuitive natural-language code generation), making it easier to get exactly the intended output.

Opus 4.7 introduces a new ability to automatically double-check its own outputs. In the internal 93-task benchmark, Opus 4.7 was the only model to solve 4 tasks that both Opus 4.6 and Sonnet 4.6 failed to resolve. This self-verification capability improves reliability in long-running autonomous tasks, giving developers greater confidence when delegating code review and automated testing.

Claude Mythos is Anthropic's frontier system developed under the codename Project Glasswing. It is currently available only to approximately 40 carefully selected enterprise and government partners and is not publicly accessible. Opus 4.7 sits below Mythos in Anthropic's model hierarchy but serves as the highest-performing model commercially available — in practice, the best model for the vast majority of users.

Opus 4.7 pricing is unchanged from Opus 4.6: $5 per million input tokens and $25 per million output tokens. Prompt caching provides up to 90% discount and the Batch API offers a 50% discount. With significantly improved performance at the same price point, existing users effectively get a substantial improvement in cost-efficiency.

Comparison of SWE-bench, Vision, and pricing across major competing models:

In SWE-bench Verified, Opus 4.7 records the highest score among publicly available commercial models.

According to Anthropic staff, getting the most out of Opus 4.7 requires a few days of learning. Here are the top five techniques: 1. Use xhigh mode: Specify xhigh for complex reasoning and deep analysis to get high-quality output without consuming max-level resources. 2. Leverage literal interpretation: Write concrete, explicit instructions rather than vague ones to get exactly the intended output. 3. Delegate to self-verification: Assign long-horizon tasks to Opus 4.7 and adopt a workflow of reviewing its self-verified results. 4. Exploit Vision 98.5%: Paste screenshots and UI mockups directly to instruct analysis and code generation. 5. Pair with Claude Code: Take advantage of improved vibe coding predictability for large-scale autonomous refactoring.

Migrating from Opus 4.6 to Opus 4.7 is essentially just a model name change in your API calls. However, due to improved literal interpretation, instructions that were previously interpreted loosely may now be processed more literally. Key adjustment points: (1) Rewrite prompts to be more explicit (e.g., instead of 'do something reasonable,' specify 'generate Y in format X'), (2) Explicitly specify xhigh when needed, (3) Take advantage of expanded image support up to 3.75MP for vision tasks. Full API compatibility with Opus 4.6 is maintained.

Five use cases where Opus 4.7 delivers exceptional results: 1. Large-scale refactoring: Safely restructure codebases of tens of thousands of lines with 87.6% SWE-bench accuracy and 70% CursorBench reliability. 2. UI mockup analysis: Generate code directly from Figma screenshots using Vision 98.5%, supporting images up to 3.75MP. 3. Complex agentic workflows: Use xhigh + long-horizon autonomy to execute CI/CD pipeline builds and multi-step deployment automation without human intervention. 4. Automated code review: Leverage self-verification and thorough problem-solving to integrate bug detection, improvement suggestions, and test generation. 5. Technical documentation: Use literal interpretation and thoroughness to generate API specs, architecture design docs, and user guides that precisely match specifications.

Following the release, reactions on X (formerly Twitter) and Reddit AI communities have been enthusiastic. Common praise includes 'clearly smarter than 4.6 as a coding agent,' 'Vision has finally reached a practical level,' and 'xhigh mode improved quality of complex design work.' Some users note that the shift to literal interpretation requires adjustment, and that getting the most out of Opus 4.7 takes a few days of learning — something Anthropic has acknowledged officially. Many threads sharing Opus 4.7 usage tips have emerged across the community.

Q1. Is there a price increase from Opus 4.6? No. Pricing is identical: $5 per million input tokens and $25 per million output tokens. No changes. Q2. What is Claude Mythos? Claude Mythos is Anthropic's frontier system developed under the codename Project Glasswing. It is available only to approximately 40 selected enterprise and government partners, with no announced date for general availability. Q3. Does xhigh mode cost more? The deeper reasoning may consume more tokens, but the per-token pricing (model rate) is the same as other Effort Control levels. Q4. Is it available in Claude Code immediately? Yes. As of the April 16, 2026 release date, Opus 4.7 is available in Claude Code, Claude Chat, Claude Cowork, Anthropic API, AWS Bedrock, Google Vertex AI, and Azure AI Foundry. Q5. What specifically improved in Vision? Accuracy jumped from 54.5% to 98.5% (+44 points), and supported image size expanded to up to 2,576 pixels (3.75 megapixels) — more than three times the previous limit. Screenshot analysis, diagram reading, and UI mockup understanding are now at a genuinely practical level. Q6. Is the switch to Opus 4.7 automatic? No. You need to explicitly specify the model name in the API or Claude Code, using a format such as `claude-opus-4-7-20260416`. Q7. When will Claude Mythos be publicly available? Anthropic has not announced a date for general availability. It currently operates as a frontier system for limited partners only.

If you want to integrate Claude Opus 4.7's coding agent, Vision, and xhigh reasoning capabilities into your systems, Oflight's AI consulting service is here to help. From API integration design and agentic workflow construction to internal document automation, our experienced engineers support every step. Contact us at /services/ai-consulting for details.