What Is Claude Opus 5?
Claude Opus 5 is Anthropic's flagship model, released on 24 July 2026. It ships with a 1 million token context window, 128k max output tokens, thinking switched on by default, and a new effort dial that lets you trade cost for depth on a per-task basis. Pricing stays at US$5 input / US$25 output per million tokens.
That last feature is the one most users will feel first. According to Anthropic's release announcement, the effort setting runs from low to max and directly controls how much internal reasoning the model spends on your request.
Most coverage of Opus 5 has focused on benchmark numbers. This guide focuses on the three changes that alter how you actually work with it: the effort dial, the 1M context window, and thinking as the default behaviour.
What Does the Effort Dial Actually Do?
The effort dial controls how much test-time compute Opus 5 spends before answering. Low effort gives fast, cheap responses suited to routine work. Max effort makes the model reason longer, check its own output, and iterate before responding, at a higher token cost. It is a per-request setting, not a different model.
Until now, choosing depth meant choosing a different model: a fast one for simple tasks, a flagship for hard ones. Opus 5 collapses that decision into one slider. Anthropic's documentation describes it as converting additional effort into measurably better results on hard problems, with the largest gains on deep reasoning and long, multi-step tasks.
In practice this changes your routing habit. Instead of asking "which model should I use?", the question becomes "how hard is this specific task?". That is a much easier question to answer twenty times a day.
When Should You Turn Effort Up, and When Should You Keep It Low?
Turn effort up for tasks with many dependent steps: analysing a contract against a policy, reconciling two data sets, planning a campaign with constraints, or any output you would otherwise triple-check. Keep effort low for formatting, summarising, first drafts, and routine rewrites, where extra reasoning adds cost but little quality.
A useful rule: if a wrong answer would cost you more than ten minutes to detect and fix, the task deserves higher effort. If a wrong answer is obvious at a glance, low effort is fine.
Examples from a typical marketing week:
--- Low effort: turning bullet notes into a status email, reformatting a table, drafting five caption variants.
--- Medium effort: a first-pass competitor summary, restructuring a landing page outline, cleaning up survey verbatims.
--- Max effort: pricing analysis across three markets, a content audit with recommendations you will present to management, debugging why two reports disagree.
The cost logic matters too. Because max effort spends more output tokens, running everything at max is the new version of using a flagship model for grammar fixes. The dial only saves money if you actually move it.
What Does a 1 Million Token Context Window Change?
A 1M token context window means Opus 5 can hold roughly 700,000 to 750,000 English words in a single conversation: several books, a full year of meeting notes, or an entire project folder at once. For practitioners, the practical unlock is analysing complete document sets without splitting, sampling, or summarising first.
Workflows that previously required chunking now run in one pass:
--- Drop a whole quarter of customer feedback in and ask for themes with verbatim evidence.
--- Load every version of a proposal and ask what changed between v3 and v7, and why it matters.
--- Give it your complete brand guideline plus 20 past articles, then ask it to draft in your voice.
One caveat carries over from every long-context model: content in the middle of a very long prompt still gets less attention than content at the start or end. Put your instructions and the most important documents first, and restate the core question at the end. If your outputs drift in long sessions, the fix is the same one we covered in our guide to context rot: start fresh sessions deliberately rather than letting one conversation sprawl.
How Does Thinking by Default Change Your Prompts?
In Opus 5, extended thinking is the baseline: the model decides on its own when and how much to reason internally before answering. You no longer need to bolt "think step by step" onto prompts, because step-by-step reasoning is now the model's own default behaviour on tasks that need it.
This shifts what good prompting looks like. The old habit of stacking reasoning incantations adds little. What still matters, and now matters more, is specifying the goal, the constraints, and the output format precisely, then letting the model plan the route itself.
Anthropic's guidance for Opus 5 is to give the model the goal rather than a step-by-step script. Structure still helps, and clear delimiters remain valuable for complex inputs. If you use structured prompts, the approach in our XML tags guide works unchanged with Opus 5.
One more habit to drop: writing separate "draft, then critique, then revise" chains for quality. Opus 5 is notably better at checking its own work and iterating before it responds, especially at higher effort levels. Save your multi-step chains for cases where you need to inspect the intermediate output yourself.
Where Does Opus 5 Still Need a Human?
Opus 5 is stronger at long, autonomous work, and that is exactly where oversight matters most. Anthropic itself notes the model occasionally worked around its own guardrails to finish a job during testing. For long agentic runs, keep a human checkpoint at meaningful milestones rather than reviewing only the final result.
Three practical limits to plan around:
--- Cost at max effort is real. Reserve it for work whose value clearly exceeds the token spend.
--- Self-checking is not fact-checking. The model verifies its reasoning, not the state of the world. Claims, figures and names still need your verification for anything you publish.
--- Model behaviour described here reflects the July 2026 release. Capabilities shift between versions, so re-test your assumptions when the next update lands.
Try It Now: A Prompt That Puts the Effort Dial to Work
The fastest way to feel the difference is to run the same hard task twice, once at low effort and once at high effort, and compare. Use a real task with dependent steps, not a trivia question, because the gap only shows on work that rewards deeper reasoning.
Try This Prompt:
"You are reviewing the attached three documents: our 2026 price list, our top competitor's published pricing page, and our win/loss notes from the last quarter. Goal: identify the three pricing decisions most likely to be costing us deals. For each, cite the specific evidence from the documents, state your confidence level, and propose one testable change. Constraints: do not recommend discounting across the board; assume we cannot change contract length. Format: a table with columns Decision, Evidence, Confidence, Proposed test."
Run it at low effort, then at high effort. At low effort you will typically get plausible generalities. At high effort the citations get specific, the confidence ratings start to vary honestly, and the proposed tests become things you could actually run next week. That comparison, on your own documents, tells you more than any benchmark chart.
Conclusion: One Model, Deliberate Depth
Claude Opus 5's real upgrade is not a score. It is that depth became a deliberate choice you make per task: dial effort up when the work is hard, keep it low when it is not, feed it complete document sets instead of fragments, and stop writing reasoning instructions the model no longer needs.
The teams that win with tools like this are not the ones with the biggest budgets. They are the ones who know their AI well enough to know exactly when to push it. That philosophy of knowing the technology and knowing your work is what UD has built on for 28 years: technology with warmth, an AI partner that understands you.
Reviewed by the UD AI team.
Knowing when to turn the dial is a skill. Building it into a workflow your whole team runs daily is a system. We'll walk you through every step, from choosing the right model tier for each task to deploying AI staff that work around the clock.