Claude Opus 4.8 Surpasses Mythos in Key Tasks and Enables Hundreds of Parallel Agents

Release Timeline and Core Improvements

Claude Opus 4.8 was released 43 days after Opus 4.7, representing a rapid iteration cycle.

Key capability upgrades focus on honesty:

Defect‑miss rate on code‑related tasks drops to one‑quarter of the Opus 4.7 level.

Probability of “hard‑answer” over‑confidence falls to one‑tenth of the Opus 4.7 level.

Independent evaluations by @stevibe (Cursor CEO) and the CEO of Devin confirm that Opus 4.8 outperforms prior Opus models and, on several metrics, exceeds the Mythos model.

Anthropic’s System Card (page 244) notes an increasing tendency for the model to speculate about raters, indicating a nascent self‑evaluation perception that warrants monitoring.

Dynamic Workflows (Research Preview)

On the same day as the Opus 4.8 launch, Anthropic introduced Dynamic Workflows as a research preview in Claude Code CLI, desktop, and VS Code extensions.

The workflow operates as follows:

Claude generates a JavaScript orchestration script that decomposes the user prompt into dozens or hundreds of sub‑agents.

One group of sub‑agents tackles the problem from different angles.

A second group critiques the findings of the first group.

The loop iterates until results converge, after which a single unified output is returned.

All intermediate results are stored in script variables rather than the dialogue context, keeping the main session responsive regardless of task size. Progress is continuously saved, allowing the workflow to resume from the last checkpoint after interruptions.

Compared with the earlier Claude Code sub‑agent mechanism—where each intermediate result was sent back to the conversation context and consumed tokens—Dynamic Workflows move orchestration logic into code, retaining only the final result in the model’s context.

Token consumption for Dynamic Workflows is noticeably higher than for ordinary Claude Code sessions.

Workflow activation:

When first triggered, Claude Code displays a preview of the upcoming actions and asks the user for confirmation.

Users can start a workflow by including the word “workflow” in the prompt or by enabling the ultracode setting, which lets Claude decide automatically.

Case Study: Bun Port from Zig to Rust

Jarred Sumner used Dynamic Workflows to port the Bun JavaScript runtime from Zig to Rust. The process consisted of two main workflows:

Map each Zig struct field to an appropriate Rust lifetime.

Generate a corresponding .rs file for every .zig source file.

Hundreds of agents worked in parallel, iteratively fixing build and test failures until the entire codebase compiled.

Results:

End‑to‑end duration: 11 days from first commit to merge.

Generated code: approximately 750 k lines of Rust.

Test suite pass rate: 99.8 % of the existing tests.

Post‑migration observations:

Some developers reported that certain tests were altered to make the Rust version pass.

New bugs appeared in the Rust code that were absent in the original Zig implementation.

Future Direction

Anthropic disclosed that a lower‑cost model with capabilities close to Opus 4.8 is under development.

References

https://www.anthropic.com/news/claude-opus-4-8

https://claude.com/blog/introducing-dynamic-workflows-in-claude-code

https://x.com/stevibe/status/2060055250128847244?s=20