Anthropic Launches Claude Opus 4.7: Enhanced Coding Capabilities Meet Strict New Security Safeguards

2026-04-17 companies

San Francisco, Thursday, 16 April 2026.
Anthropic’s Claude Opus 4.7 significantly boosts coding performance while introducing stringent automated safeguards against high-risk cybersecurity requests, paving the way for safer enterprise AI adoption.

The Capabilities Leap and Engineering Focus

Anthropic released Claude Opus 4.7 on April 15, 2026, marking a direct upgrade to its predecessor, Claude Opus 4.6, which launched just two months prior in February 2026 [2][4][7]. Positioned as a major leap for advanced software engineering and knowledge work, the model requires less supervision for complex, multi-step tasks [1][4]. It extends agentic capabilities with improved long-horizon autonomy and systems engineering [6]. Early enterprise testing reveals notable efficiency gains; cloud content management company Box reported a 56% reduction in model calls and a 50% drop in tool calls, alongside a 24% increase in response speed and a 30% reduction in AI Units used [4].

Token Economics and the Cost of High Effort

While Anthropic has maintained the base pricing structure of Opus 4.6—charging $5 per million input tokens and $25 per million output tokens—the underlying token economics have shifted [7][8]. Opus 4.7 introduces an updated tokenizer, meaning the same text input can map to between 1.0 and 1.35 times more tokens depending on the content type [4][7]. This structural change requires corporate IT departments to carefully monitor their AI budgets, as the actual cost per query may rise despite the static base price [alert! ‘inference based on token increase and flat base pricing’].

The ‘Mythos’ Strategy and Unprecedented Security Guardrails

Perhaps the most intriguing aspect of the Opus 4.7 release is what the model is deliberately restricted from doing. Opus 4.7 arrives shortly after Anthropic announced Project Glasswing in early April 2026, which included the limited rollout of its most powerful model, Claude Mythos Preview, to select partners like JPMorgan Chase, Apple, and Google [2][7][8]. Anthropic acknowledges that Opus 4.7 does not advance its overall “capability frontier,” as Mythos Preview outperformed it on every relevant evaluation [8]. Instead, Anthropic intentionally experimented with “differentially reducing” the cyber capabilities of Opus 4.7 during its training process [2][8].

Multimodal Advancements and Broad Enterprise Distribution

Beyond text and code, Opus 4.7 delivers substantial upgrades to its multimodal vision capabilities. The model can process images at more than three times the resolution of previous iterations, supporting files up to 2,576 pixels on the longest edge, or approximately 3.75 megapixels [1][7]. This enhanced visual acuity translates to a massive performance jump in specific tasks; for instance, in autonomous penetration testing visual-acuity benchmarks, Opus 4.7 scored 98.5%, up from 54.5% achieved by Opus 4.6 [7]. This represents an 80.734 percent relative improvement in visual benchmark performance [7]. This allows the model to produce higher-quality interfaces, slides, and documents by accurately interpreting charts and dense user interface screens [1][6].

Sources

Artificial intelligence Anthropic