Most People Did Not Realise AI Had a Quality Dial Until Now
Most people assumed every AI tool gives one fixed level of effort, like a single-speed bicycle. Until 28 May 2026, that assumption was mostly right. On that day, Anthropic released Claude Opus 4.8 with a feature called Effort Control, a four-position slider that decides how hard the model thinks before answering. It is the first widely available AI quality dial.
The Surprising Fact for most business owners is this: the difference between the lowest setting and the highest is not 10 percent or 20 percent. On complex reasoning tasks, the gap exceeds 70 percent on Anthropic's published benchmarks. The same model, the same prompt, the same five seconds of waiting versus 45 seconds, produces fundamentally different output.
This article walks through what Effort Control does, the four settings, when to use each one, and how a Hong Kong business owner should decide which level fits which task.
What Is Effort Control?
Effort Control is a new four-level setting in Claude Opus 4.8 that lets the user choose how much computational thinking the model performs before producing an answer. The four levels are Low, High (default), Extra, and Max. Higher levels cause Claude to spend more time, generate more internal reasoning steps, and check its own work more thoroughly before responding.
According to Anthropic's launch documentation on 28 May 2026, Effort Control sits next to the model selector inside Claude.ai. The setting is a slider, not a hidden parameter. Each user can switch levels per conversation, and the chosen level is visible on every response.
The feature only applies to Claude Opus 4.8. Earlier models, including Opus 4.6 and 4.7, do not expose the control. Users on Claude.ai, the Claude API, and Claude Code all see the same four settings, though the underlying compute budget and pricing differ by surface.
How Does Effort Control Actually Work?
Effort Control adjusts Claude's internal reasoning budget. At Low effort, Claude generates fewer hidden reasoning tokens and returns its first reasonable answer. At Max effort, Claude can spawn parallel reasoning paths, verify each against the prompt, and only commit when its self-check passes.
A practical way to picture this: imagine asking a senior employee for an answer. At Low, the employee gives the first answer that pops into their head. At High, they think for 10 seconds and check their assumptions. At Extra, they sketch the answer, review it, and revise once. At Max, they sketch, review, revise, ask themselves what could be wrong, and only then reply.
Anthropic's published benchmarks for Opus 4.8 show this is not marketing language. On the SWE-Bench Verified coding benchmark, Max setting outperforms Low by more than 70 percent on resolution rate. On long-form analytical tasks, the gap is smaller but still significant.
What Each Effort Level Does and When to Use It
The four settings map onto distinct task types. Choosing the right one is not about always picking the highest. It is about matching effort to job.
Low effort: use this for quick lookups, simple summaries, single-paragraph drafts, and any task where speed matters more than depth. Typical response time is under five seconds. Good fit for: rewriting a sentence, summarising a single short email, generating a quick caption.
High (default): Anthropic's recommended setting for most business work. Response time is 8 to 15 seconds. Good fit for: drafting a customer reply, preparing a one-page brief, writing a product description, doing routine research with one or two sources.
Extra: use when output quality genuinely matters. Response time runs 25 to 45 seconds. Good fit for: drafting a board paper, analysing a contract, building a multi-step plan, comparing several options with trade-offs.
Max: reserve for the highest-stakes work. Response time can exceed one minute. Good fit for: writing a legal-sensitive policy, debugging a complex bug, reasoning over a multi-page strategy document, building an audit response.
What Are Dynamic Workflows? The Companion Feature
Dynamic Workflows is a companion feature launched with Opus 4.8 that lets Claude Code take on very large problems by spawning many parallel subagents in a single session. According to Anthropic, the model can plan the work, run hundreds of subagents concurrently, verify each result, and only then report back.
For a Hong Kong business, the practical use is automation that previously required a developer. A user can ask Claude Code to audit 200 customer service emails for compliance issues, generate individual reply drafts, and flag the cases that need a human. The agent does all 200 in parallel, then synthesises the findings.
Dynamic Workflows is currently exclusive to Claude Code, not the consumer Claude.ai surface. It also runs longer at higher effort settings, which is why Effort Control and Dynamic Workflows are paired in Anthropic's launch announcement.
What Does It Cost? And Why Fast Mode Got Cheaper
Claude Opus 4.8 API pricing did not change versus Opus 4.7: US$5 per million input tokens and US$25 per million output tokens. The cost change is on Fast Mode, which runs the model at 2.5 times the normal speed.
Fast Mode for Opus 4.8 is now three times cheaper than the previous generation, at US$10 per million input tokens and US$50 per million output tokens. For comparison, that is a meaningful step down: on previous Opus generations, Fast Mode commanded a much larger premium for the same speed gain.
Inside Claude.ai, Effort Control does not change the subscription price directly. Higher effort uses more compute behind the scenes, which Anthropic absorbs within the existing Pro and Team plan limits. Heavy users on Max effort may hit rate limits faster, but the bill does not change per query.
How Honest Is Opus 4.8? The Underrated Benchmark Most People Missed
Buried in Anthropic's launch announcement is the most underrated number in the release: Opus 4.8 fails to raise important events to the user only 3.7 percent of the time, scores 0 percent on uncritically reporting flawed results (a first for Claude), and shows a more than tenfold reduction in overconfidence versus Opus 4.7.
What this means in plain language: if Claude does something risky or makes a flawed inference, it now tells the user more reliably. Previous generations would sometimes deliver wrong answers in confident language. Opus 4.8 is, on Anthropic's measurement, about four times less likely than its predecessor to let code-level flaws slip through unremarked.
For a business deploying AI in customer-facing or decision-support workflows, this honesty improvement is more important than any speed gain. It reduces the supervision burden.
Common Misconceptions About Effort Control
Three misconceptions about Effort Control are already common. Each will cost you efficiency if you act on them.
Misconception one: higher effort is always better. Not true. On simple tasks, Max effort delivers the same answer as High, just slower. The cost is your time, not the bill. Use the level that fits the task.
Misconception two: Effort Control replaces good prompting. Not true. A poorly written prompt at Max effort still produces a mediocre answer. Effort Control raises the ceiling of the answer, but prompt quality still sets the floor.
Misconception three: Low effort means low quality. Not true. Low effort means less internal reasoning. For routine tasks, the answer is fully sufficient and significantly faster. Anthropic positions Low as a productive default for high-volume, simple work, not as a degraded mode.
How Should a Hong Kong Business Use Effort Control?
A practical playbook for an SME is to map effort levels to your team's existing task types.
For customer service and front-office work, default to Low or High. Speed matters more than depth, and most replies follow templates. Tens of replies per hour is the right cadence.
For marketing and content writing, default to High, switch to Extra for hero pieces such as a homepage or annual report passage.
For data analysis and reporting, default to Extra. Analysis benefits from Claude's longer reasoning, especially when the prompt asks for multiple comparisons or scenarios.
For strategic, legal, or financial documents, use Max. The slower response is worth it when the cost of an error is high.
The rule of thumb: use the lowest effort level that gives an answer you would not need to revise. Going higher beyond that is wasted time.
Frequently Asked Questions About Effort Control
Is Effort Control available on Claude Free?
No. As of June 2026, Effort Control is available on Claude Pro, Team, Enterprise, and the Claude API. Claude Free users only see the default High setting on a different model.
Does Effort Control work in Cantonese or Traditional Chinese?
Yes. The setting is language-agnostic. The underlying reasoning quality differences hold across English, Traditional Chinese, Simplified Chinese, and Cantonese prompts.
What is the difference between Effort Control and Extended Thinking on earlier Claude models?
Extended Thinking on Opus 4.5 and 4.6 was a binary on or off toggle. Effort Control on Opus 4.8 is a four-position slider with cleaner gradation and better cost-quality calibration.
Will higher effort always produce a better answer?
Higher effort produces a more thoroughly reasoned answer. On a task that does not benefit from more reasoning, the outputs at High, Extra, and Max may be indistinguishable.
The Bottom Line for Hong Kong Business Owners
Effort Control is the first time a major AI vendor has handed the quality dial to the user instead of hiding it behind product tiers. The four-level slider is a small interface change, but it makes the cost-quality trade-off explicit and controllable for the first time.
For Hong Kong businesses already using Claude, the immediate action is to retrain your team on when to use each level. The wrong default wastes time at Max or shortcuts quality at Low. For businesses still evaluating AI vendors, Opus 4.8 with Effort Control is the cleanest example yet of an AI tool that respects the user's judgment about when depth matters and when speed wins.
UD stands with you, making AI human.
Ready to Pick the Right AI Tools for Your Team?
Claude, ChatGPT, Gemini, Microsoft Copilot — every vendor now ships a slightly different control surface. Choosing the right combination for your business needs more than a feature comparison. UD has spent 28 years helping Hong Kong SMEs deploy AI tools the right way. We will walk you through every step, from comparing vendors to training your team.
