Claude Opus 4.7 — a new model from Anthropic: what has changed and why it is needed

What is the model and when was it released

Claude Opus 4.7 is Anthropic’s flagship model, released on April 16, 2026. It is a direct update to Opus 4.6: the same pricing—$5 per million input tokens and $25 per million output tokens—but with significantly enhanced capabilities.

The model is available through all Claude products, the Anthropic API, as well as Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Developers integrate it using the model ID claude-opus-4-7.

The main improvement is code

The primary focus of Opus 4.7 is software development. The model handles complex and lengthy tasks significantly better: it follows instructions more accurately, sustains multi-step scenarios, and can verify its own results before providing an answer.

Users from early tests note that they can now hand over tasks that previously required constant oversight—and trust the outcome.

Here is what statistics from companies with early access show:

— Cursor tested the model on 93 tasks, and Opus 4.7 solved 13% more tasks than Opus 4.6, including four tasks that neither Opus 4.6 nor Sonnet 4.6 could handle
— Rakuten reported that Opus 4.7 resolves three times as many production tasks as the previous version, with double-digit improvements in code and test quality
— Replit recorded the same level of quality at lower costs: the model analyzes logs more accurately, finds bugs, and suggests fixes.

The model now sees better

Opus 4.7 processes images with a resolution of up to 2,576 pixels on the long edge—more than three times higher than previous Claude models.

In practice, this means the model can now confidently read dense screenshots, technical diagrams, complex charts, and crowded tables. Previously, it would often misinterpret such images or miss details entirely.

XBOW, a company specializing in autonomous penetration testing, reported that visual content recognition accuracy increased from 54.5% with Opus 4.6 to 98.5% with Opus 4.7.

Long tasks without failures

One of the key changes is the ability to work consistently over extended periods. Previous models often got stuck in loops or stopped midway when tackling complex, multi-step tasks.

The Devin team described it this way: the model now works coherently for hours, continues solving complex problems instead of giving up, and unlocks a new class of deep research tasks that could not be reliably executed before.

Notion noted a 14% improvement over Opus 4.6 with fewer errors. Opus 4.7 became the first model to pass their tests for understanding implicit user needs.

Improved memory between sessions

Opus 4.7 makes better use of file memory. The model remembers important data during long work sessions and applies it in subsequent tasks—without needing to provide the entire context again each time. This is especially important for agent scenarios where work spans multiple sessions.

New effort level: xhigh

Alongside Opus 4.7, a new parameter called xhigh—“extra high”—has been introduced. It sits between the high and max levels, giving developers more precise control over how deeply the model “thinks” about a task.

In Claude Code, the default effort level for all plans is now raised to xhigh. When testing Opus 4.7 for coding and agent scenarios, Anthropic recommends starting with high or xhigh.

Security: new protections for cyber tasks

Anthropic views Opus 4.7 as a testbed for new protection mechanisms in cybersecurity. The model automatically detects and blocks requests showing signs of prohibited or high-risk usage.

At the same time, for cybersecurity professionals—vulnerability researchers, pentesters, and red team members—Anthropic has launched the Cyber Verification Program. Those who pass verification gain access to the model’s full capabilities for legitimate work.

What to consider when migrating from Opus 4.6

Two factors affect token usage. First: Opus 4.7 uses an updated tokenizer—the same text may consume 1 to 1.35 times more tokens depending on the content. Second: at higher effort levels, the model reasons more extensively, especially in later stages of agent tasks—this improves accuracy but increases the number of output tokens.

Anthropic has published a separate migration guide from Opus 4.6 to Opus 4.7.