English

← Back to list

Claude Opus 4.7: +13% Coding Gains, Same Price

itadol5j

Claude Opus 4.7 official release

🔥 Claude Opus 4.7 Is Out: A Full Leap in Coding Ability

Anthropic has officially released Claude Opus 4.7, the latest in the Opus lineup:

  • A big coding leap: +13% on a 93-task benchmark
  • Fewer tokens: low-effort Opus 4.7 ≈ medium-effort Opus 4.6
  • Better instruction-following: sticks to instructions strictly, with fewer tool errors

Opus 4.7 shows major improvements in advanced software engineering, and it especially shines on the hardest tasks.

📊 Key Numbers

Metric Opus 4.7 vs Opus 4.6
93-task coding benchmark +13% improved
Tool error rate reduced 1/3
CursorBench 70% vs 58%
Visual precision benchmark 98.5% vs 54.5%

🧠 Core Improvements

1. More Reliable on Complex Tasks

Opus 4.7 performs better on complex, long-running tasks:

  • Rigor and consistency: stays rigorous and consistent when handling complex tasks
  • Precise instruction-following: follows user instructions to the letter
  • Self-verification: designs its own methods to verify its output

2. Higher-Resolution Vision

Vision got a major upgrade:

  • Higher resolution: can inspect images at higher resolution
  • Better taste: more taste and creativity on professional work
  • Higher quality: produces better UIs, slides, and documents

3. Fewer Tool Errors

Opus 4.7's tool calls are more reliable:

  • 1/3 fewer tool errors: steadier tool calling
  • Error recovery: keeps going when a tool fails instead of stopping
  • Higher quality-per-tool-call: more output quality from every tool call

💰 Pricing Unchanged

Opus 4.7 costs exactly the same as Opus 4.6:

Type Price
Input tokens $5 per million
Output tokens $25 per million

Fewer tokens: low-effort Opus 4.7 ≈ medium-effort Opus 4.6, which means lower cost.

🎯 How It Relates to Mythos Preview

Anthropic was explicit:

  • Mythos Preview is more powerful than Opus 4.7, but its release is restricted
  • Opus 4.7 is the first model to be tested under the new safeguards
  • Cyber capabilities: Opus 4.7 falls short of Mythos Preview
  • Safeguards: automatically detect and block high-risk cyber requests

Cyber verification program

Security professionals can apply for the cyber verification program:

  • Vulnerability research
  • Penetration testing
  • Red teaming

💡 What Users Are Saying

Vercel:

  • "Outstanding on single coding tasks — more correct and more complete than Opus 4.6"
  • "More honest about its own limitations"
  • "Even proves properties of the system's code before starting work"

Cursor:

  • "A clear CursorBench jump: 70% vs 58%"
  • "More autonomy and more creative reasoning"

Devin:

  • "Can work for hours on end"
  • "Pushes through hard problems instead of giving up"
  • "Unlocked deep investigation work that previously couldn't run reliably"

Notion:

  • "+14% success rate with fewer tokens"
  • "First model to pass our implicit-requirements test"
  • "Keeps going after tool failures instead of stopping"

Rakuten:

  • "Rakuten-SWE-Bench: solved 3x more production tasks"
  • "Double-digit gains in both code quality and test quality"

📝 Summary

Claude Opus 4.7 is a major step up in coding ability:

  • +13% coding benchmark, 70% vs 58% on CursorBench
  • Fewer tokens: high quality at lower cost
  • Better instruction-following: strict adherence, fewer tool errors
  • Self-verification: it checks its own output
  • Vision gains: 98.5% vs 54.5% visual precision
  • Same price: $5/1M input, $25/1M output

A full leap in coding ability at the same price. That's Opus 4.7.

Official announcement:https://www.anthropic.com/news/claude-opus-4-7

Claude API:https://platform.claude.com/docs/en/about-claude/models/overview