Builders who want to understand Claude Opus 4.8 dynamic workflows can skip the suspense: Dynamic Workflows is a research preview inside Claude Code that lets Opus 4.8 write a JavaScript orchestration script, then run up to 1,000 subagents in parallel to finish large tasks. It is not a standalone API primitive yet. Prototype now. Do not ship it as a load-bearing dependency until general availability.
Claude Opus 4.8's USAMO 2026 math score jumped to 96.7% from 69.3% on Opus 4.7, a 27-point gain in a single model generation, per Anthropic's May 2026 release notes. That number gets the headlines. But the bigger story is architectural. Anthropic moved multi-agent orchestration from a developer-built framework problem into a provider-managed runtime. That shift has direct consequences for cost, debuggability, and how much control you keep.
What Is Claude Opus 4.8 and What Ships With It?
Opus 4.8 released on May 28, 2026, at the same price as Opus 4.7: five dollars per million input tokens and twenty-five dollars per million output tokens. Flat pricing signals that Anthropic is not charging a capability premium here.
Three features ship together. Dynamic Workflows in Claude Code lands as a research preview. An effort control slider arrives on claude.ai, letting you dial compute allocation per task. Fast mode returns at a price three times lower than fast mode cost on Opus 4.7, the largest per-token cost reduction in the Opus line since Opus 3.
Anthropic frames this as an agentic infrastructure drop, not a broad reasoning lift. Benchmark gains cluster in coding and multi-step task categories. If you need the baseline context from the previous release, the Claude Opus 4.7 Features, Benchmarks and Pricing Explained guide covers what changed and what carried over.
How Do Dynamic Workflows Coordinate Subagents?
When you describe a complex task in Claude Code, Opus 4.8 writes a JavaScript orchestration script. A background runtime executes that script, spinning up subagents in parallel to handle distinct pieces of the work. The plan lives in script variables, not in Claude's context window. Your active session stays clean. Only the final synthesized result returns to you, not the raw intermediate agent outputs.
Hard limits as of the research preview: 16 concurrent subagents at any one time, and 1,000 total subagents per run. Tasks that exceed those thresholds need job splitting at your application layer.
Anthropic's documented demo covers migrating hundreds of thousands of lines of code, using an existing test suite as the pass/fail bar. That scope signals the intended use case: large, parallelizable work where a cluttered context window would otherwise slow you down or burn your token budget. For comparison with how Claude Code handles smaller daily coding tasks, 8 Claude Code Workflows Developers Run Daily shows the non-orchestration side of the same tool.
What Benchmark Gains Did Opus 4.8 Actually Post?
The USAMO 2026 math benchmark is the standout number. Opus 4.8 scores 96.7% versus 69.3% on Opus 4.7, a 27-point jump that suggests targeted capability work rather than a uniform lift. As of May 2026, that is the highest publicly posted score for any Claude model on that evaluation.
Standard coding evals tell a different story. SWE-bench Pro moves from 64.3% to 69.2%. SWE-bench Verified goes from 87.6% to 88.6%. Real gains, but incremental. The uneven distribution matters: Opus 4.8 is not uniformly better across every task type. It is notably better on hard mathematical reasoning and modestly better on everyday coding benchmarks.
Anthropic describes alignment quality as near-Mythos level, referencing its internal safety evaluation tier. Dynamic Workflows ships as a research preview, which carries an implied caveat: the capability is verified enough to test, but not committed to a stable production interface yet.
How Does Pricing and Fast Mode Compare to Previous Opus Versions?
Base API pricing holds flat: five dollars per million input tokens and twenty-five dollars per million output tokens. That matches Opus 4.7. Anthropic is not charging a premium for Dynamic Workflows or the effort slider.
The change is in fast mode. As of May 2026, fast mode on Opus 4.8 costs three times less than it did on prior Opus models. Fast mode runs at roughly 2.5 times standard speed. For teams running large agent pipelines or batch jobs, that gap is material.
A note on first-hand data: the side-by-side dollar figures for a 10 million output token batch job across Opus 4.7 and Opus 4.8 fast mode are in progress and have not yet been confirmed for publication here. Use the three-times cost reduction ratio as your working multiplier when re-pricing existing pipelines. Run your own batch before locking in budget projections. The direction is clear even before the exact cents-per-call figure is verified.
Who Should Use Dynamic Workflows Right Now?
Engineering teams running codebase migrations, large test-suite refactors, or multi-file code generation are the clearest fit. Anthropic's own demo scope is hundreds of thousands of lines of code. The feature was designed for parallelizable, large-surface work with a clear pass/fail signal.
The research preview tag means no SLA guarantees and a likely interface change before general availability. Production pipelines should treat Dynamic Workflows as a capability to prototype, not a dependency to commit to. Teams running workloads like automated content pipelines should watch the GA announcement before wiring Dynamic Workflows into a live stack.
Non-engineering use cases, like multi-source research synthesis or large-scale document processing, are plausible but lack published benchmarks. Build a scoped test first. Measure actual token consumption against an equivalent sequential Claude Code session. Then decide whether the orchestration overhead earns its keep on your specific task type.
What Are the Current Limits and Caveats to Know?
The 1,000-subagent-per-run cap and 16-concurrent-subagent limit are hard limits as of the research preview. Tasks that push past those numbers need job splitting at the application layer. The exact queuing or error behavior when you intentionally hit the 16-concurrent ceiling is not yet documented in Anthropic's public release materials. That is a practical gap worth testing before you build scheduling logic around the limit.
As of May 2026, Dynamic Workflows carries no production SLA guarantees and no committed API interface. It is exclusive to Claude Code in this release. The effort slider and fast mode are available on claude.ai and the API, but orchestration is not yet a standalone API primitive. That distinction matters if your stack does not use Claude Code.
The safest build posture: wrap Dynamic Workflow calls behind an abstraction layer. If the interface changes before GA, you update one adapter, not every integration point. For how agent failure modes surface at scale, AI Agent Safety Failures: Inside the 2026 Agents of Chaos Paper is worth reading before you push any autonomous pipeline toward production.
How Does This Compare to Other Agent Orchestration Approaches?
Frameworks like LangGraph, AutoGen, and CrewAI handle multi-agent coordination at the application layer. You define the graph, the routing logic, and the memory structure. Dynamic Workflows moves that orchestration inside the model provider's runtime, reducing the infrastructure a developer needs to build and maintain.
The tradeoff is control. External frameworks give full visibility into each agent's inputs, outputs, and routing decisions. Dynamic Workflows abstracts that away in exchange for simpler setup. Only the final result surfaces to your session. This suits many task types. It matters more when you need to audit intermediate steps or debug a failed subagent mid-run. For a head-to-head view of how Claude Code stacks up against other AI coding tools in this kind of agentic work, Codex vs Claude Code: Which AI Coding Tool Wins in 2026 covers the broader competitive picture.
OpenAI's Swarm and Assistants v2 take a similar managed-orchestration direction. The differentiator for Opus 4.8 is the script-in-variables architecture that keeps context windows small on the orchestrator side. That design choice reduces cost per run and keeps the orchestrator from hitting context limits on long jobs. It also means intermediate agent reasoning is less visible by default. Know which tradeoff matters more for your use case before you pick a path.
If you are ready to test Claude Opus 4.8 dynamic workflows on a real task, start with a scoped migration or refactor where your existing test suite gives you a clear pass/fail signal. Open an issue on the Anthropic developer forum if you hit the 16-concurrent limit, since that behavior is the least documented part of the current release. And subscribe to GenAI Club for the GA announcement, because the pricing and interface that ships at GA is what you will actually build on.
FAQ
What is Dynamic Workflows in Claude Opus 4.8?
Dynamic Workflows is a feature in Claude Code that lets Claude write a JavaScript orchestration script to coordinate dozens to hundreds of parallel subagents on a single task. Instead of running steps sequentially in one context window, the runtime spins up subagents in parallel, each handling a discrete piece of the work. Only the final combined result comes back to your session. As of May 2026, it is a research preview available inside Claude Code, capped at 16 concurrent subagents and 1,000 total subagents per run. It is not yet a standalone API primitive.
How many subagents can Dynamic Workflows run at once?
In the current research preview, Dynamic Workflows supports a maximum of 16 subagents running concurrently within a single session, with a hard cap of 1,000 total subagents per run. Tasks that would require more parallelism than that need to be split into multiple jobs at the application layer. Anthropic has not announced whether these caps will change when the feature moves to general availability. Treat both numbers as subject to revision before building production dependencies on them.
Is Claude Opus 4.8 more expensive than Opus 4.7?
No. Anthropic held pricing flat at five dollars per million input tokens and twenty-five dollars per million output tokens, matching Opus 4.7. Where costs actually dropped is in fast mode: Opus 4.8 fast mode is priced three times lower than fast mode was for Opus 4.7. For teams running high-volume batch workloads or long agentic pipelines, that reduction is meaningful. Standard API pricing for the base model is unchanged.
What is the effort control slider in claude.ai?
Effort control is a new interface option on claude.ai that lets users dial the amount of compute and reasoning Claude applies to a task. At lower effort, responses are faster and use fewer tokens. At higher effort, Claude does more extended thinking before responding, which can improve output quality on complex tasks but takes longer. The slider ships with Opus 4.8 and is separate from Dynamic Workflows. It is available directly in the claude.ai interface, not only through the API.
How does Dynamic Workflows differ from using LangGraph or AutoGen?
LangGraph, AutoGen, CrewAI, and similar frameworks handle multi-agent coordination at the application layer, meaning you build and maintain the routing logic, memory, and error handling yourself. Dynamic Workflows moves orchestration inside Anthropic's runtime: Claude writes the coordination script and the provider executes it. The benefit is simpler setup and smaller context windows on the orchestrator side. The tradeoff is less visibility and control over intermediate agent behavior. For teams that need custom routing, auditability, or cross-model orchestration, an external framework still makes more sense. For teams that want to delegate coordination to the model, Dynamic Workflows reduces infrastructure overhead.
Sources
- Introducing Claude Opus 4.8
- Anthropic releases Opus 4.8 with new dynamic workflow tool
- Anthropic Ships Claude Opus 4.8 Alongside Dynamic Workflows and Cheaper Fast Mode, With Workflows Capped at 1,000 Subagents
- Anthropic's Claude Opus 4.8 is here with 3X cheaper fast mode and near-Mythos level alignment