Claude Opus 4.7 Arrives: 13% Coding Boost with Quiet Confidence

Anthropic has a distinctive trait: their product updates are remarkably low-key.

Unlike certain companies (who shall remain unnamed), where every model upgrade demands an online launch event with dozens of media outlets and a CEO delivering passionate speeches for 30 minutes, Anthropic just quietly updates the model, posts a technical blog, and moves on.

Claude Opus 4.7 is no exception. It silently launched on April 16—I almost missed it despite checking their API documentation daily.

But this update packs quite a bit worth discussing.

13% Coding Improvement: How’d They Do It?

Let’s start with the most direct change: Opus 4.7 shows an average 13% improvement on HumanEval and MBPP coding benchmarks.

Thirteen percent might not sound like much, but remember—Opus 4.6 was already in the top tier. Moving from 90% to 93% is far harder than jumping from 60% to 70%.

After digging through their technical blog, several key improvements emerge:

Training data now includes code execution feedback. Previously, model training only looked at “static correctness” (can it compile?). This time, they added an execution environment where generated code actually runs, with training weights adjusted based on execution results (bugs, performance metrics).

This matters because a lot of code “looks right but breaks when run.” A sorting algorithm might be syntactically perfect but have O(n²) complexity, exploding on large datasets. With execution feedback, the model learns to “not just write correctly, but write well.”

Introduction of “verification mode.” Opus 4.7 adds a new capability: after generating code, it automatically creates test cases to verify its own output. If tests fail, the model attempts fixes.

This implements the “Chain of Verification” approach—letting the model check itself. In practice, enabling verification mode boosts code accuracy by another 5-8 percentage points.

Refined reasoning effort levels. Previously, Claude had three reasoning effort levels: low, medium, high. Opus 4.7 adds “xhigh,” specifically for ultra-complex tasks.

I tested it—in xhigh mode, generating a complex algorithm takes 8 seconds instead of 3, but accuracy is notably higher. Yes, latency increases, but if you’re doing code reviews or architectural design where “one mistake costs dearly,” this trade-off is worthwhile.

High-Resolution Image Support: More Than Just “Seeing”

Another significant update: Opus 4.7 supports image inputs up to 3.75 megapixels.

To put this in context: an iPhone 15 Pro photo is about 12 megapixels. 3.75 megapixels means shrinking the original to 1/3 size before feeding it to the model.

What’s this good for?

I tested a scenario: screenshotting an entire page of code and asking the model to analyze the logic. With Opus 4.6, if the image was slightly blurry, the model couldn’t recognize variable names. With 4.7, even a full 4K monitor screenshot is accurately read—every character recognized.

Another interesting application: technical documentation. Many docs are scanned PDFs with charts and code snippets. Previously, you’d need OCR to extract text before processing. Now you can directly screenshot entire document pages to Opus 4.7—it understands text, charts, and code simultaneously, handling complex layouts.

Self-Verification: The Model Starts “Second-Guessing” Itself

Opus 4.7 introduces another new capability: “self-verification.”

Simply put, before giving an answer, the model asks itself several questions:

• Did I miss any important information?
• Are there logical holes in my reasoning?
• Could there be alternative explanations?

If the model detects potential issues, it re-reasons or explicitly tells you “I’m uncertain.”

This capability is particularly useful for complex problems. Previously, when I asked the model to analyze a complex distributed system architecture, Opus 4.6 would give a “seemingly reasonable” answer that didn’t hold up under scrutiny.

Opus 4.7 instead marks: “Here’s an assumption—I’m not sure if it’s correct, need more information.” This ability to “know what you don’t know” is a crucial reliability improvement.

Real-World Experience: Details Worth Noting

After using Opus 4.7 for a week, several details stand out:

Better understanding of Chinese code comments. Previously, Claude would often “miss” Chinese comments in code. Opus 4.7 shows marked improvement, accurately understanding Chinese comment semantics.

More stable handling of long code. Tested a 5000-line Python file—Opus 4.7 maintains consistent context understanding throughout, no “forgetting” mid-way.

Price unchanged. This is commendable. Many model companies raise prices after upgrades, but Opus 4.7 API pricing matches 4.6. Better performance at the same cost.

What This Means for Developers

Honestly, Opus 4.7 isn’t a “disruptive” upgrade—it’s a series of subtle, incremental improvements.

But precisely these small improvements make the actual user experience significantly better.

If you’re already using Claude for AI-assisted coding, switch to Opus 4.7. If your use cases are mainly simple tasks (writing scripts, generating docs), Sonnet series offers better value—no need for Opus.

What I’m most excited about is further development of “self-verification” capabilities. If models can truly “know when they don’t know,” AI-assisted decision-making reliability will reach a new level.

After all, a model that “knows what it doesn’t know” is far more reliable than one that “dares to say anything.”