Anthropic has introduced Claude Opus 4.8 and paired the release with a research preview of Dynamic Workflows inside Claude Code. The update centers on two changes: the model keeps the same pricing while improving reasoning, self-assessment of progress, and long-duration autonomous work; at the same time, a new Fast mode raises execution speed to 2.5x while cutting cost to one-third of the standard level.
The company announced the release on X and in its official blog on May 28, 2026. For developers, Fast mode is enabled directly from the Claude Code terminal with the command /fast. Anthropic presented it as a way to run the same model with a different cost-speed profile rather than as a separate model tier.
Opus 4.8 focuses on longer independent execution
According to Anthropic, Claude Opus 4.8 builds on the previous generation with sharper logical judgment, more honest tracking of its own progress, and stronger ability to work independently over extended periods. That matters in software tasks where completion is not a single response but a sequence of actions, checks, and revisions.
The company also framed the release around durability. Instead of handling only short coding sessions, the model is designed to stay on task for longer stretches, which fits large engineering jobs such as framework replacement across hundreds of files, API deprecation work, and low-level code migration.
Dynamic Workflows adds orchestration across many subagents
The more dramatic change is Dynamic Workflows, now available as a research preview in Claude Code. Anthropic describes it as a system that lets Claude take on complex engineering work end to end by planning, splitting, assigning, and validating tasks through parallel AI agents.
In this setup, Claude writes coordination scripts on its own, breaks a large assignment into smaller units, and launches dozens to hundreds of subagents in parallel. Those agents approach the job from different angles, while adversarial agents attempt to break or challenge the current output. The process continues through repeated iteration until results converge. This turns the tool into something closer to an automated project coordinator than a single coding assistant.
Anthropic said the workflow can run continuously for hours or even days, and progress can be preserved if execution is interrupted. That makes the feature aimed squarely at large, high-complexity software projects rather than ordinary prompt-response use.
Bun case study shows an 11-day migration of 750,000 lines
To demonstrate the system, Anthropic shared a case involving Bun creator Jarred Sumner. Using Dynamic Workflows, Bun’s core codebase was migrated from Zig to Rust. The company said Claude coordinated hundreds of agents to write .rs files in parallel, with two AI reviewers assigned to each file.
After generation, the system kept running fix loops that pushed compilation and testing forward until the build passed. Anthropic said the project covered about 750,000 lines of Rust code, reached a 99.8% test pass rate, and went from the first commit to final merge in just 11 days.
The company also cautioned that launching hundreds of subagents consumes far more tokens, which means higher compute cost. Its recommendation is to start with smaller tasks first. Developers who want to try the multi-agent workflow can access it through direct commands or by enabling ultracode in settings.