Anthropic has unveiled Opus 4.8, an updated version of its most powerful artificial intelligence model. The developers have focused on the accuracy and «honesty» of the model, which is now four times less likely to make errors in its own code.
Anthropic has unveiled Opus 4.8, an updated version of its most powerful artificial intelligence model. The developers have focused on the accuracy and «honesty» of the model, which is now four times less likely to make errors in its own code.
The company announced the release on its official blog. The model is already available globally at the standard rates of the previous version ($5 per million incoming tokens and $25 per outgoing), and also received a fast mode at a price of $10/$50 respectively. Developers can connect the model via the API using the identifier claude-opus-4-8.
The main technological improvement of Opus 4.8 was the fight against hallucinations and hasty conclusions, which are often the fault of modern AI models. Early testers note that the model has become proactive in pointing out doubtful points and making false assumptions less often. This is confirmed by the investment company Bridgewater Associates, where partners noted the model’s ability to independently detect problems during the analysis of input data.
In addition to improving accuracy, Anthropic has introduced several new features to streamline operations:
Dynamic Workflows. This feature helps Opus manage hundreds of parallel subagents to solve large-scale tasks in Claude Code. The tool is now capable of performing codebase-level migrations of hundreds of thousands of lines of code from launch to merge, guided by existing tests. The feature is available in preview for Enterprise, Team, and Max plans.
Effort Control. claude.ai and Cowork now have a toggle that adjusts the depth of analysis. At high settings, the model «thinks» longer and more carefully for complex tasks, while at low settings, it responds faster and spends the user’s limits more economically.
Updates to the Notification API. Developers can now modify system instructions directly while an AI agent is executing a task (for example, updating permissions or token limits) without flushing the prompt cache or restarting the queue.
Analysts attribute the rapid release of Opus 4.8 (only 41 days after the release of version 4.7) to increased competition from OpenAI products and Google’s Gemini Flash model.
Meanwhile, the company is holding back the release of its most advanced model, the Mythos, for now. It is currently undergoing closed cybersecurity testing by a limited number of organizations as part of the Glasswing project. Anthropic emphasizes that security is a priority due to the high power of the model, but promises to open access to the Mythos line to all customers in the coming weeks.
Recall that Anthropic raised $65 billion in its latest funding round. After the closing of this deal, the company’s valuation reached $965 billion.
