Opus 4.7 Is GREAT (except the token usage)

Highlights

• Opus 4.7 provides a 10% performance increase on the SWE Bench Pro benchmark and a 7% leap on verified coding tasks compared to version 4.6.

• Input prompts can cost up to 35% more tokens due to an updated tokenizer, while the new default 'extra high' effort level in Claude Code uses roughly the same tokens as the previous version's maximum setting.

• Visual processing capabilities allow for image resolutions three times higher than previous models, improving data extraction and computer-use tasks.

• Instruction following is strictly literal in Opus 4.7, requiring users to adjust existing prompts that relied on the loose interpretation of previous versions.

• Long-context performance in needle-in-a-haystack tests shows a significant decline compared to version 4.6, despite improvements in multi-session file system memory.

• In a single-prompt 20-minute test for a financial dashboard, Opus 4.7 generated a functional React 18/TypeScript application with higher UI quality than GPT 5.4 or Gemini 3.1.

Timeline

1. 00:00 - 01:06

   Core Upgrades and Token Economics

   The release focuses on refining coding and vision performance while introducing a higher cost per interaction. The model processes text differently, meaning the 'token burn' is higher even before accounting for increased internal reasoning. These changes serve as a bridge toward Mythos, a model restricted from public release until cybersecurity risks are mitigated.

   • Opus 4.7 introduces improved vision, self-verification, and more creative UI generation capabilities.
   • The updated tokenizer increases token usage by up to 35% for identical input prompts compared to older models.
   • Anthropic is using Opus 4.7 to test cyber safeguards required for the future release of the more powerful Mythos model class.
2. 01:06 - 02:12

   Benchmark Performance and Instruction Following

   Benchmark data shows consistent gains in software engineering tasks but highlights a specific weakness in long-context reliability. The shift toward literal instruction following means prompts must be more precise. Safety protocols are intentionally suppressing scores in specific high-risk categories like cybersecurity to prevent misuse.

   • SWE Bench Pro scores increased by 10%, while cybersecurity scores slightly decreased due to artificial safety safeguards.
   • Needle-in-a-haystack tests reveal a performance regression in long-context retrieval compared to Opus 4.6.
   • The model interprets instructions literally, which may break prompts designed for the more flexible interpretations of previous versions.
3. 02:12 - 03:30

   Vision, Memory, and Effort Levels

   Higher resolution image support targets improvements in automated computer use and complex data extraction from visuals. Memory management now relies more on persistent file notes, potentially offsetting some token costs over long projects. Users can optimize costs by downgrading from the 'Extra High' default to 'High' effort without losing performance relative to the previous 4.6 model.

   • Multimodal support now handles images with 3x the resolution of previous versions.
   • File system memory improvements allow the model to retain notes across multiple sessions to reduce upfront context needs.
   • Opus 4.7 High effort level outperforms Opus 4.6 Max effort level while using fewer tokens than the new Extra High default.
4. 03:30 - 05:12

   UI Design Comparison and Competition

   Testing across a simple cafe website prompt shows that while Opus 4.7 is a step above 4.6 in font choice and layout, Gemini 3.1 currently leads in visual composition for simple landing pages. GPT 5.4 shows the least progress in design creativity, maintaining a highly recognizable and generic style. Opus 4.7 remains a top contender for UI tasks due to its ability to handle responsive layouts and asset integration.

   • Opus 4.7 creates responsive websites with integrated assets from Unsplash in a single prompt.
   • Gemini 3.1 produced the most aesthetically pleasing cafe website result in head-to-head testing.
   • GPT 5.4 remains behind in aesthetic variety, often defaulting to a repetitive card-based design language.
5. 05:12 - 10:49

   Full-Stack Development Stress Test

   In a complex task involving a financial dashboard with a backend API, Opus 4.7 provided the best balance of UI quality and code structure. Although it used in-memory storage instead of a persistent SQLite database like 4.6, its use of modern TypeScript made it the most professional output. GPT 5.4 failed to generate a full project structure, providing only basic HTML and CSS files, while Gemini 3.1's more advanced Next.js approach resulted in a broken user interface.

   • Opus 4.7 built a functional financial dashboard with React, Vite, and TypeScript in 20 minutes.
   • The model correctly utilized TypeScript for the entire project, whereas GPT 5.4 and Opus 4.6 defaulted to JavaScript or older React versions.
   • Gemini 3.1 attempted a Next.js architecture but failed to produce clickable tabs or working features in a single prompt.