Anthropic has released Claude Opus 4.8 and dynamic workflows for Claude Code, enhancing AI capabilities for complex coding tasks.
Anthropic has launched Claude Opus 4.8 and introduced a new feature called dynamic workflows for Claude Code, enhancing its capabilities.
Claude Opus 4.8 improves upon its predecessor with better judgment, increased honesty about its performance, and extended independent operation time. This update is now available at the same price as before.
Introducing Claude Opus 4.8: it builds on Opus 4.7 with sharper judgment, more honesty about its own progress, and the ability to work independently for longer than its predecessors.
Available today at the same price.
— Claude (@claudeai) May 28, 2026
The dynamic workflows feature allows the AI to autonomously create orchestration scripts that can initiate numerous parallel sub-agents, verify their outputs, and present the results to users. This is particularly useful for complex projects involving security audits, bug detection, and code migration.
This feature is currently in beta and can be accessed via the command-line interface, the desktop version of Claude Code, the VS Code extension, as well as through APIs and platforms like Amazon Bedrock and Microsoft Foundry.
Users can start a workflow using direct commands or through ultracode, enabling the model to optimize computational resources and determine when to use a multi-step approach.
It is important to note that dynamic workflows require significantly more tokens than regular sessions with Claude Code.
The model segments tasks into subtasks, assigns them to various agents, and consolidates the results after ensuring mutual verification. For instance, a developer named Jared Samner utilized this feature to migrate the Bun project from Zig to Rust, generating around 750,000 lines of Rust code with a remarkable 99.8% success rate on existing tests, completing the task in just 11 days.
Performance highlights for Opus 4.8 include:
- 69.2% on SWE-Bench Pro;
- 49.8% on Humanity’s Last Exam without tools, rising to 57.9% with tools;
- 83.4% on OSWorld-Verified;
- 1890 points on GDPval-AA;
- 53.9% on Finance Agent v2.
However, in Terminal-Bench 2.1, it scored 74.6%, trailing behind GPT-5.5's 78.2%.
Anthropic emphasized that Opus 4.8 shows improved honesty in task execution, often expressing uncertainty and better identifying issues before presenting results.
This launch follows Anthropic's recent report on Project Glasswing, a program aimed at vulnerability detection using the Claude Mythos model.
