Anthropic Releases Claude Opus 4.7, Its Most Capable Generally Available Model

Anthropic released Claude Opus 4.7 on April 16, 2026, its most capable model available for general use, two months after Claude Opus 4.6 shipped in February, continuing what the company describes as a bimonthly flagship upgrade pattern.

According to Anthropic, the model is a notable improvement on Opus 4.6 in advanced software engineering, with gains on the most difficult tasks.

It also introduces new infrastructure controls, including task budgets and a new effort level, giving developers finer management over how the model allocates reasoning across long-running agentic workflows.

The model is available across all of Anthropic’s Claude products, its API, and cloud partners Microsoft, Google, and Amazon.

The Coding and Agentic Benchmark Story

The clearest benchmark gains are in software engineering. 

Anthropic confirmed Claude Opus 4.7 scores 64.3% on SWE-bench Pro, a rigorous multi-language software engineering benchmark, up from 53.4% for Opus 4.6 and ahead of GPT-5.4’s published score of 57.7%. 

On SWE-bench Verified, the AI model reaches 87.6%, up from 80.8%. On the MCP-Atlas scaled tool use, it scores 77.3% compared to GPT-5.4’s 68.1%.

For agentic computer use, Opus 4.7 scores 78.0% on OSWorld-Verified, a 5.3-point improvement over Opus 4.6 and ahead of GPT-5.4’s 75.0%.

Early-access partners confirmed similar real-world results. CodeRabbit reported that code review recall improved by over 10% on complex pull requests while precision stayed stable. 

Rakuten reported Opus 4.7 resolves three times more production tasks than Opus 4.6 on its SWE-Bench variant. 

Genspark’s engineering team said the model showed meaningfully improved loop resistance, a critical quality for autonomous agents that must not stall indefinitely on edge cases.

Three Technical Features That Change How Developers Use the Model

Three product changes ship alongside the model, according to Anthropic.

First is the xhigh effort level, a new setting between high and max that gives more control over reasoning depth and response speed. Anthropic also raised the default effort level in Claude Code to xhigh and recommends high or xhigh for coding and agentic use.

As confirmed by Axios, the company is introducing task budgets in public beta, allowing developers to set a token limit across an agentic loop so the model can prioritize and finish tasks instead of stopping mid-task.

The third update is vision. The model now processes images up to 2,576 pixels on the long edge, up from 1,568, applied automatically without any API changes.

One migration consideration developers should plan for: Anthropic Opus 4.7 uses a new tokenizer, so the same input can use up to 1.0 to 1.35 times more tokens, depending on content type. With the unchanged token pricing, this can increase the cost per request, especially for token-heavy workloads. 

The model also follows instructions more literally than the previous model of Claude, improving reliability on complex engineering tasks, but may cause older prompts to behave unexpectedly.

Why Opus 4.7 Ships With Intentionally Reduced Cyber Capabilities

The Claude Opus 4.7 launch is directly tied to the safety architecture Anthropic began building after it revealed the Claude Mythos Preview on April 7. 

Anthropic confirmed it reduced Opus 4.7’s cybersecurity capabilities during training, not due to a specific risk in the model but to test safeguards it plans to use for future Mythos-class models before wider release.

The company said in its blog post that Opus 4.7 ships with safeguards that automatically detect and block requests linked to prohibited or high-risk cybersecurity uses. It added that lessons from real-world use of these safeguards will also inform future model development at scale. 

Opus 4.7 is available immediately on Claude Pro, Max, Team, and Enterprise plans, as well as through the API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. Pricing is unchanged from Opus 4.6.

Source: Introducing Claude Opus 4.7