Let's start with the number that stands out most. In an internal benchmark with 93 software engineering tasks, Opus 4.7 solved 13% more problems than its predecessor, Opus 4.6. That doesn't sound like much until you understand what it means in practice: four tasks that no previous Anthropic model — not Opus 4.6, not Sonnet 4.6 — could solve are now handled by 4.7.
Think of it like a mechanic who used to fix 87 out of every 100 car problems and now fixes 100. The 13 that were missing weren't the easy ones.
The second change is less obvious but equally important. Opus 4.7 processes images up to 2,576 pixels on the long side. That's more than three times the resolution that previous Anthropic models could work with.
Why does this matter? Because many real-world tasks involve images with fine details: technical diagrams, interface screenshots, photographed spreadsheets, code captures. When a model can't see clearly, it guesses. When it sees in high resolution, it works with what's actually there.
A technical detail that will interest API users: Anthropic introduced a new effort level called xhigh. It sits between the existing "high" and "max" levels.
In practice, it's like a finer-grained speed control. Before, you had gear 3 and gear 5. Now you have gear 4. Depending on the task, you can request more reasoning without paying for the maximum cost. For projects running hundreds of calls per day, that granularity represents real savings.
The third pillar of this update is behavioral: Opus 4.7 was trained to maintain consistency and rigor in tasks that take a long time to complete. We're not talking about a single question and answer. We're talking about sessions where the model needs to plan, execute, verify its own results, and report back — without losing the thread halfway through.
This is especially relevant for agentic workflows, where the model acts as an autonomous agent that uses tools, accesses files, makes calls to external APIs, and makes chained decisions. Anthropic states that 4.7 is more careful about verifying its own outputs before considering a task complete.
Worth noting: Opus 4.7 comes with automatic protections against high-risk uses in cybersecurity. Any request that signals malicious intent is blocked before being processed.
The model is available on the Claude platform, the Anthropic API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Pricing has not changed from Opus 4.6: $5 per million input tokens and $25 per million output tokens.
The company continues to bet on safety and reliability as differentiators. Opus 4.7 is not the most secretive or most powerful model Anthropic has — internally, there is Claude Mythos, still in restricted access. But what was delivered to the public this time is substantial: higher visual resolution, more precision in code, more control for developers, and more reliable behavior on complex tasks.
For those working with automation, software development, or document analysis, Opus 4.7 represents an update worth testing.
