Anthropic Claude Opus 4.7: Stability Over Performance

Claude Opus 4.7: A Strategic Release for Anthropic

On April 16, 2026, Anthropic unveiled its latest model, Claude Opus 4.7, which directly succeeds Opus 4.6. This model is immediately available across several major platforms such as Claude, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. Despite the enhancements made, the usage cost remains unchanged, set at $5 per million tokens for input and $25 for output. This announcement comes nine days after the launch of Claude Mythos Preview and the Project Glasswing program, highlighting a strategy that goes beyond mere technical improvements.

An Enhanced Vision for More Autonomous Agents

One of the standout features of Opus 4.7 is its improved visual capability, allowing for high-resolution image perception. This enhancement is not merely aesthetic; it is crucial for the autonomy of agents in complex tasks requiring interaction with graphical interfaces. Claude AI now supports file uploads of up to 100,000 tokens, or approximately 75,000 words, while offering advanced customization and ethical design.

Anthropic's partners attest to these improvements. For instance, Cursor saw its score on the CursorBench benchmark rise from 58% to 70%. GitHub reports a 13% improvement across 93 coding tasks, while Notion observes a 14% increase in multi-step workflows, with reduced token consumption and a third fewer tool errors. Devin from Cognition highlights the model's ability to work consistently for hours without supervision.

These visual capabilities enhance the generation and correction of interfaces in code, manipulation of PowerPoint presentations, and autonomous operation of third-party applications. These features are essential for the agentic products that Anthropic has been promoting since January, such as Claude in Chrome and Claude in Excel.

Cyber Capability Throttling: A Thoughtful Strategy

Anthropic has intentionally limited the cyber capabilities of Opus 4.7 during its training, integrating automatic safeguards to block risky cybersecurity requests. For legitimate professionals, a Cyber Verification Program has been established, making access to full cyber capabilities conditional on prior validation.

This throttling reinforces the decision to keep Claude Mythos Preview locked down. If Opus 4.7 requires a deliberate reduction of its cyber capabilities, it justifies maintaining strict control over Mythos. Anthropic thus demonstrates that its most advanced model must remain confined for security reasons.

A Response to Infrastructure Saturation

Several technical indicators suggest that Opus 4.7 is designed to be lighter and more efficient. Hex indicates that the model in "low effort" mode delivers results comparable to Opus 4.6 in "medium effort" mode. Notion and Replit confirm superior performance at a reduced cost. Despite these gains, the price remains the same as that of Opus 4.6.

This strategy addresses infrastructure saturation, highlighted by a widespread outage on March 2, 2026, and a public ticket from AMD in early April, signed by Stella Laurenzo, indicating a degradation in Claude Code's performance since February across 6,852 analyzed sessions. The abrupt closure of the ecosystem to third-party tools on April 4, with OpenClaw explicitly mentioned, also underscores this pressure.

Thus, Opus 4.7 is a technical response aimed at stabilizing service quality while reducing inference costs. Additional efforts, such as the design of dedicated chips, are also underway.

The Infrastructure Challenge Against Competition

For 2026, the crucial question is whether Claude can remain the top consumer model as competition strengthens on the infrastructure front. Models like Gemini 3.1 Pro and GPT-5.4 offer competitive pricing and features, putting intense pricing pressure on Anthropic.

Anthropic has recognized the need to adapt its infrastructure to meet growing demand. It remains to be seen whether these efforts will be sufficient to maintain its market position.