What Are Effort Control and Dynamic Workflows in Claude Opus 4.8?
On May 28, 2026, Anthropic released Claude Opus 4.8 with two features that quietly change how you actually use Claude day to day. Most people heard the headline, scrolled past, and kept prompting the way they always have. That is a mistake.
Effort Control is a new slider next to the model picker in Claude.ai and Cowork. It lets you choose how much thinking Claude does before answering, on a scale from default to "max." Dynamic Workflows is a research-preview feature in Claude Code that lets Claude plan a multi-step task and orchestrate up to 1,000 parallel sub-agents in one session.
Together, the two features answer a question every intermediate Claude user has hit: "Why are my results inconsistent on hard tasks, and what do I do about it?" The short answer is that you have been using one effort level for every job, and Claude has been guessing how hard to think.
How Does Effort Control Actually Change Output Quality?
Effort Control gives Claude a budget for reasoning tokens before it produces a final answer. On higher effort settings, Claude spends more time decomposing the problem, considering alternatives, and checking its own work. On the default setting, it answers faster and uses fewer tokens.
According to Anthropic's release notes, Opus 4.8 defaults to "high" effort, which produces the same quality as Opus 4.7's default with similar token usage. The two new levels are "extra" (xhigh in Claude Code) and "max." Extra is the recommended setting for difficult tasks like long technical writing, complex data analysis, or multi-step planning.
Where you will notice the difference: structured reasoning where one wrong assumption breaks the whole answer. Think strategic memos, financial models, legal-style argument, or any task where you would normally have to do three rounds of "actually, let me reconsider that." On those tasks, max effort cuts your follow-up prompts roughly in half.
Where it does not matter: short writing, fact lookup, simple summaries. On these, default effort produces the same answer in a fraction of the time. Cranking the slider to max for "rewrite this email" is just wasting your wait.
When Should You Use Each Effort Level?
Most users default to one setting forever and never touch it. That is a missed opportunity. Here is a working framework based on task type.
Default (high) — for the 80% of daily tasks: drafting messages, summarising documents, brainstorming, rewriting, simple research questions, answering technical questions you mostly already understand.
Extra (xhigh) — for tasks where a wrong direction wastes hours: outlining a long report, designing a workflow, writing a strategy document, comparing options where the trade-offs are subtle, debugging logic problems, drafting anything that has to be exactly right the first time.
Max — for irreversible, high-stakes outputs: legal-style argument, financial scenario modelling, anything you will hand off without re-reading carefully, deep code reviews where missing an issue is expensive. Max effort makes Claude visibly slower, but the quality jump on hard reasoning is real and measurable.
A practical habit: pick your effort level the same way you pick the model. Match it to the difficulty of the task in front of you, not to your default mood.
What Are Dynamic Workflows and How Do They Work?
Dynamic Workflows is Claude Opus 4.8's headline feature in Claude Code. According to Anthropic, when triggered, Claude plans a complex task, then runs hundreds of parallel sub-agents in the background, with each agent working on a slice of the larger problem. Claude verifies their outputs and reports back as one consolidated result.
You trigger it in one of two ways: include the word "workflow" anywhere in your prompt, or turn on a setting called "ultracode" in Claude Code. Ultracode combines xhigh reasoning effort with automatic workflow orchestration on every prompt.
The use case Anthropic showcased is codebase-scale migration: refactoring hundreds of thousands of lines across a repo, with the existing test suite as the bar for success. But the same orchestration helps non-developer practitioners too, anywhere you have a large task that breaks down into many similar smaller tasks.
Practical applications outside of code: processing 500 customer reviews into structured insights, generating personalised outreach for 200 prospects, auditing every page of a documentation site, generating product descriptions across an entire SKU catalogue. Each of those is a workflow, not a single prompt.
How Do You Write a Prompt That Actually Triggers a Workflow?
Just typing the word "workflow" is necessary but not sufficient. The prompt also has to describe a task that genuinely benefits from parallelism. If you ask for a single output, Claude will not spawn sub-agents even with the magic word.
Three ingredients make a prompt workflow-friendly: a clearly defined unit of work, a clear count or list of those units, and a clearly defined success criterion for each unit. Here is a copy-paste-ready template.
Try this prompt in Claude Code (with ultracode enabled):
"Run a workflow to audit every blog post in /content/blog/ against the following 5 quality criteria: (1) opening hook within first 2 sentences, (2) at least one specific number or named example in the first 200 words, (3) clear answer to the title's implied question by the third paragraph, (4) at least one CTA at the bottom, (5) no broken internal links. For each post, return a JSON object with the post slug, a 0-5 score per criterion, and a one-line fix suggestion for any criterion scored 2 or below. Save the consolidated results to audit-results.json."
Notice the structure: clear unit (one post), clear list (every file in a folder), clear success criteria (5 named checks per post), clear output format (JSON). That is what lets Claude split the work cleanly across sub-agents.
What Are the Real Limitations You Need to Know About?
Dynamic Workflows is in research preview as of May 2026. According to Anthropic, sub-agents are capped at 1,000 per workflow, and the feature is only available on Claude Code Enterprise, Team, and Max plans. If you are on Pro, you have Effort Control but not Dynamic Workflows.
The feature is also still rough at the edges. Sub-agents occasionally drift from the original task definition, especially when the unit of work is loosely defined. Real-world reports note that workflows with vague success criteria tend to produce inconsistent quality across the 100+ outputs, which then takes longer to clean up than doing the job sequentially would have.
The token cost is non-trivial too. Running max effort on a single complex prompt might use 30,000 tokens. Running a workflow with 200 sub-agents on the same task can use 100x that. For one-off tasks the math is fine. For repeated workflows you run weekly, consider whether a cheaper model running sequentially would do the same job at a fraction of the cost.
The honest summary: Effort Control is universally useful and you should start using it today. Dynamic Workflows is powerful but plan-dependent and best used when the task genuinely needs parallelism, not as a flex.
How Does This Change Your Daily Claude Workflow?
Three practical adjustments most intermediate users should make this week.
First, retrain your habit. Every time you start a new conversation, pause for one second and ask: "Is this a default task or a difficult task?" If difficult, bump the slider before you type. That single half-second pause meaningfully changes your output quality over a month.
Second, identify your repeatable workflows. Look at your last 50 Claude conversations. Are there 3 to 5 tasks you repeat weekly? Those are workflow candidates. Write a single workflow prompt for each and save it. The first run takes 20 minutes to set up. Every run after that takes 30 seconds.
Third, recalibrate your expectations on hard tasks. If you have been frustrated that Claude "gets simple tasks but struggles with complex ones," the answer was never to write a longer prompt. It was to give Claude more thinking budget. Try the same prompt at max effort once. The difference is usually obvious.
Where Does This Leave Claude vs ChatGPT and Gemini?
Effort Control is not unique to Claude. OpenAI offers reasoning effort settings in its o-series models, and Google's Gemini has thinking-time controls. What is unique to Claude Opus 4.8 is the combination of a per-prompt slider in the consumer UI plus genuine workflow orchestration in Claude Code at the 1,000-agent scale.
For most intermediate practitioners, the model choice still comes down to what you use Claude for. If you do a lot of long-form writing, strategic thinking, or code-adjacent work, Opus 4.8 with extra effort is currently the strongest option for that mix of tasks. If you do a lot of multimodal work or live web research, ChatGPT or Gemini still have an edge in their respective specialities.
The wider point: model picking matters, but effort picking matters almost as much, and almost nobody is doing it deliberately. The practitioners who do will get visibly better results from the same tool other people are using.
We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Make Your AI Knowledge a Competitive Edge
Knowing how to use new features like Effort Control is one piece. Building a repeatable AI workflow that gets consistent results across your whole team is the harder part. We'll walk you through every step, from picking the right effort level for each task type to designing workflows that actually save hours.
