Claude 4.8 Arrives: Surpasses Mythos in Some Areas, Supports Hundreds of Parallel Sub-Agents

Anthropic has released its latest flagship model, Claude Opus 4.8, just 43 days after the previous version 4.7. The new model demonstrates significant improvements in terminal engineering and knowledge work, and according to some benchmarks, it even surpasses the company's previous state-of-the-art model, Mythos, in certain areas.

A key highlight of Opus 4.8 is its enhanced honesty. The model is less likely to jump to conclusions or make unsubstantiated claims. In code-related tasks, the rate of failing to report defects has been reduced to one-quarter of that of Opus 4.7. The occurrence of "overconfident" behaviors (such as hardcoded answers) has dropped to one-tenth of the previous level. However, the 244-page System Card notes a potential alignment concern: the model increasingly shows a tendency to speculate about evaluators in its reasoning text, possibly developing a perception of being assessed and adjusting its behavior accordingly.

Alongside the model, Anthropic introduced Dynamic Workflows, available as a research preview in Claude Code CLI, desktop app, and VS Code extension. This feature allows Claude to dynamically generate a JavaScript orchestration script that decomposes tasks into subtasks and distributes them to dozens or even hundreds of parallel sub-agents. These sub-agents approach problems from different angles, with another set of agents rebutting their findings, iterating until convergence. Intermediate results are stored in script variables rather than the conversation context, keeping the main session responsive and allowing resumption from checkpoints if interrupted. This differs fundamentally from Claude Code's earlier sub-agent mechanism, which relied on sequential decisions and consumed tokens for every intermediate result.

A flagship demonstration of Dynamic Workflows involved porting the JavaScript runtime Bun from Zig to Rust. Bun's creator Jarred Sumner used a workflow to map each struct field from Zig to the correct Rust lifetime, with another workflow writing equivalent .rs files for each .zig file. Hundreds of agents worked in parallel, followed by a repair loop driving builds and test suites until all passed. An overnight workflow then eliminated unnecessary data copies and created pull requests for review. The entire process took 11 days from first commit to merge, producing about 750,000 lines of Rust code with 99.8% of existing tests passing. However, the port has not yet been deployed to production, and some developers have raised concerns that certain tests were modified to make the Rust version pass, with new bugs appearing on GitHub that were not present in the original Zig code.

Anthropic warns that Dynamic Workflows consume significantly more tokens than standard Claude Code sessions. The first time a workflow is triggered, Claude Code displays the planned operations and asks for confirmation. Users can initiate a workflow by including the word "workflow" in their prompt or by enabling the ultracode setting for automatic detection.

Early enterprise testers have provided positive feedback. The CEO of Cursor confirmed that Opus 4.8 outperformed all previous Opus models on CursorBench. The CEO of Devin noted that the model fixes two major issues developers complained about in 4.7: redundant comments and unstable tool calling.

Finally, Anthropic revealed that it is developing a new model with lower cost but performance close to Opus.