Anthropic Lanunches Opus 5 With New Coding, Agent Features

Anthropic introduced a new AI model called Claude Opus 5 on Friday, July 24, 2026. The company built it to handle coding, business work, and long-running AI agents. Opus 5 performs close to Anthropic’s most advanced model, Fable 5, but costs much less to run. Anthropic frames the release as a sign that the AI industry is shifting focus, moving away from pure power and toward cost and everyday usefulness.

The new model is now live across all of Anthropic’s platforms, including its chatbot and developer tools. Pricing stays the same as the previous Opus model, at $5 per million input tokens and $25 per million output tokens. Opus 5 has also become the default model for Claude Max subscribers. It is now the strongest model included with the Claude Pro plan.

Anthropic reported strong gains in software development, business automation, and scientific research tasks. The company said Opus 5 more than doubled its coding performance compared to the older Opus 4.8 model on internal tests, while also reducing the cost of each task. This combination means businesses can complete more work without paying more. Anthropic pointed to this efficiency as one of the model’s biggest selling points.

Companies that tested the model early gave positive feedback. The CEO of Cognition, the company behind the Devin coding tool, said the new model nearly matches Fable 5’s coding ability at half the price, and performs especially well at finding and fixing bugs. The head of Zapier said Opus 5 led his company’s automation testing without using extra computing power compared to older Claude models. These reactions suggest the model works well in real business situations, not just in lab tests.

Opus 5 also comes with updated safety measures. Anthropic said it avoided training the model on tasks related to cyberattacks. Instead, the company focused its safety systems on supporting helpful work in areas like cybersecurity and biology research. The model can identify weaknesses in code, but it cannot scan compiled programs or generate attack tools. This approach aims to prevent misuse while still allowing useful security work.

The model also shows stronger performance handling complex agent tasks. On tests that measure how well AI can control a computer and complete multi-step jobs, Opus 5 outperformed earlier models. Anthropic said the model got better at reviewing its own work, catching errors, and finishing large projects with less human oversight. This makes it a stronger option for teams that want AI to manage bigger tasks independently.

Overall, Anthropic’s launch of Claude Opus 5 offers a lower-cost model built for coding and business agents, without raising prices from the previous version. Early feedback from companies like Cognition and Zapier points to strong real-world performance. As more businesses begin using the model, its full impact on coding and automated workflows should become clearer in the months ahead.