Anthropic releases Claude Opus 4.8 with Dynamic Workflows — up to 1,000 parallel subagents in research preview

Anthropic released Claude Opus 4.8 on May 28, 2026, alongside two companion releases: a cheaper fast mode and Dynamic Workflows, a Claude Code feature in research preview that fans work across up to 1,000 parallel subagents per session. The model is available immediately via API (`claude-opus-4-8`), on claude.ai, and across all Claude plans. Standard pricing is unchanged from Opus 4.7.

What is new in Opus 4.8. Anthropic's announcement focuses on three areas: agentic accuracy, honesty, and tool efficiency. On coding specifically, the company says Opus 4.8 is approximately four times less likely than its predecessor to let code flaws pass unremarked — the model flags issues in its own outputs rather than silently moving on. On honesty, Anthropic says Opus 4.8 is more likely to surface uncertainty about its own work rather than presenting incomplete or shaky conclusions with false confidence. On benchmark measures, the official announcement highlights strong performance on Online-Mind2Web (84% — described as a meaningful improvement over Opus 4.7 and GPT-5.5) and a claim of being the first model to exceed 10% on the Legal Agent Benchmark all-pass standard. The exact Opus 4.7 baselines are available in the comparison table in the official announcement.

Pricing: standard unchanged, fast mode significantly cheaper. Standard pricing remains $5 per million input tokens and $25 per million output tokens — the same as Opus 4.7. Fast mode — which runs at approximately 2.5x speed — is now priced at $10 per million input tokens and $50 per million output tokens, which Anthropic describes as roughly three times cheaper than the previous Opus fast mode pricing. For agentic workloads where speed matters more than marginal cost, the fast mode price reduction is the most practically significant pricing change.

Dynamic Workflows — what it is. Dynamic Workflows is a research-preview feature in Claude Code that changes the architecture of how complex tasks run. In a standard Claude Code session, the model plans and executes tasks sequentially inside a single context window. Dynamic Workflows runs differently: Claude writes a JavaScript orchestration script for the task you describe. A runtime then executes that script in the background, fanning work across up to 16 concurrent subagents per run, with a cap of 1,000 subagents total per session. Intermediate results live in script variables rather than the model's context window — so Claude's context holds only the final answer, not the accumulated state of every subtask. This means the practical working set of a single Claude Code session can now span hundreds of thousands of lines of code in parallel.

What it can actually do today. Anthropic cites one concrete example from a customer: Jarred Sumner, creator of Bun, used Dynamic Workflows to rewrite the Bun runtime from Zig to Rust — approximately 750,000 lines of Rust produced, with 99.8% of the existing test suite passing, shipped from first commit to merge in 11 days. This is the only independently attributable production example in the announcement. How well Dynamic Workflows generalizes across different codebases, tech stacks, and task types is not yet established — the research-preview designation reflects that.

Availability and plan access. The Opus 4.8 model itself is available on all plans. Dynamic Workflows requires Claude Code v2.1.154 or later and is available on Max, Team, and Enterprise plans only — not on Pro or Free. On Max and Team, it is enabled by default. On Enterprise, an administrator must activate it. Fast mode on Opus 4.8 requires usage credits rather than the standard plan allocation, per Anthropic's documentation.

What else shipped. Alongside the model, Anthropic marked two smaller items as generally available: effort control on claude.ai and Cowork (all plans), and the ability to pass system entries within the messages array in the Messages API — a capability previously in preview. Dynamic Workflows and the cheaper fast mode ship as research previews.

Who it affects. For developers running Claude Code today, the most immediately meaningful change is not the model benchmark but the Dynamic Workflows architecture — if you are on a plan that includes it. A session that previously required manually chaining prompts to work across a large codebase can now be described as a single objective and fanned out automatically. The research-preview caveat is real — expect iteration on the API, pricing, and behavior — but the underlying capability (parallel subagent orchestration capped at 1,000 agents per run) is materially new and represents a different class of automation than what Routines or standard Claude Code sessions offered. See our Claude Code review for the full capability baseline, our Claude AI review for how Opus 4.8 fits into the model lineup, our Cursor vs Claude Code comparison for the agentic coding head-to-head, and our best AI coding tools roundup for the full landscape.