Anthropic ships Claude Opus 4.8 with Dynamic Workflows

The release cadence for frontier AI models keeps tightening. On May 28, 2026 — just 41 days after Opus 4.7 — Anthropic shipped Claude Opus 4.8, which it calls its most capable generally available model, and the headline feature is aimed squarely at people who build software with AI.

Dynamic Workflows: one prompt, many subagents

The marquee addition is Dynamic Workflows, a research-preview capability in Claude Code. Rather than grinding through a task one step at a time, Claude can now "plan the work and then run hundreds of parallel subagents in a single session," verifying their outputs before reporting back. It's available in preview to Enterprise, Team, and Max plans.

For anyone learning to build with AI, this is a glimpse of where agentic coding is heading: you describe an outcome, and the model fans the work out across many workers instead of holding everything in one linear conversation. Big refactors and multi-file features are the obvious early use cases.

Quieter, more honest, less buggy

Beyond the swarm, 4.8 targets the rough edges of its predecessor. Anthropic says the model is "around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked," and it specifically fixes the comment-verbosity and over-eager tool-calling behavior that frustrated some Opus 4.7 users. On benchmarks, Anthropic reports 84% on Online-Mind2Web (computer-use and browser tasks), gains over 4.7 on its Finance Agent v2 suite, and stronger results on CursorBench across effort levels.

Pricing stays put

Importantly for budgets, pricing is unchanged from 4.7: $5 per million input tokens and $25 per million output tokens, with a fast mode at $10/$50, as TechCrunch noted in its launch coverage. A more capable model at the same price is a quiet but real win for hobbyists and small teams.

Why it matters

The practical takeaway: if you use Claude Code, you get a sharper, more self-checking model at no extra cost — and a preview of swarm-style automation for larger jobs. As coverage of the launch emphasized, the honesty and self-review improvements matter as much as raw capability when you're trusting an agent to touch your codebase. Start small: let it plan a multi-step change, watch how it verifies its own work, and scale up from there.