J'ai fait coder la même app à Claude sur chaque niveau d'effort

Summary

The presenter compares six Claude Code effort settings—Low, Medium, High, Extra, Max, and Ultra Code—by giving each the same task and assessing the output, time taken, and estimated cost. The aim is to see when extra reasoning improves results enough to justify its price. Costs are API-equivalent estimates, not direct subscription charges; subscription usage is calculated differently.

1. Recreating the Raycast website

Claude was given four screenshots and a prompt to recreate the site, including its animations.

• Low reproduced the visible pages simply and quickly. It cost $0.57 and took about a minute.
• Medium made a more interactive result, apparently using the website link to infer details not shown in the screenshots. It took about six minutes; the presenter considered whether the improvement justified its higher cost.
• High added more features and detail, including a keyboard and redesigned logos, but still had some imperfect elements. It cost $5.37 and took about 20 minutes.
• Extra improved some animations and details further, at $8.29, but the difference from High was relatively modest.
• Max produced stronger animations, a favicon, and a more complete page. It cost about $23.
• Ultra Code delivered the most polished recreation, with particularly convincing animations and an SVG logo, but cost about $134 and took just over three hours.

Takeaway: High appeared to offer the best balance for a strong result at a much lower cost than Ultra Code. Ultra Code may be worthwhile when the result must be exceptionally precise or client-ready.

2. Building a 3D conference world

The next prompt asked Claude to create a navigable 3D environment for an AI and automation conference, with rooms, stages, videos, and elements of the presenter’s identity.

• Low already produced a functional, game-like world, with movement controls, rooms, and interactive elements. It cost $4.53.
• Medium added features such as better-designed rooms, a networking area, and collision handling, for $6.61.
• High provided more detailed characters and scenery, shadows, a fuller program, and a conference logo. It took about two hours, with 112 checks; the exact cost is unclear in the subtitles.
• Extra offered some improvements, such as lighting and trees, but looked only modestly different from High and cost more.
• Max was a substantial visual step up: the presenter described it as feeling more like a polished video game, with richer details and interactions such as greeting people and ordering coffee.
• Ultra Code also produced a detailed world, including more people, rooms, and interactions. It took about two hours and cost $27. The presenter speculated that its sub-agent orchestration may have helped it work more efficiently.

Takeaway: Effort levels could make a noticeable difference on an ambitious creative task. The presenter found Max and Ultra Code more compelling than the lower settings, while noting that some interactions remained imperfect.

3. Reviewing a contract

Claude was asked to annotate a Word document containing ten potentially problematic clauses. The presenter built a small app to compare the comments from the six runs.

Most effort settings identified and explained the issues. The Low setting partially misunderstood two points; one example concerned a client’s delay not moving the project’s final deadline. The presenter had expected the highest settings to perform much better, but found little practical difference among most of them. Ultra Code performed many checks without a clear improvement.

Takeaway: For this contract-review example, the presenter judged Ultra Code unnecessary; the other settings were broadly comparable. The test illustrates that more reasoning does not guarantee a better result on every task.

4. Creating a quotes-and-invoices app

The final build was a small application for preparing quotes and invoices.

• Low produced a functional app with basic actions such as adding quote lines, duplicating a quote, and opening an email draft. The presenter considered the design generic but said it worked. It took 12 minutes and cost $3.41.
• Medium added a PDF preview, invoice conversion with deposits, customer management, history, and a better dashboard. It took 34 minutes and cost $13.
• High improved the live document preview and settings. It took 44 minutes and cost $16.
• Extra added some refinements, including keyboard shortcuts, but the presenter saw limited improvement for $35.
• Max had a better visual design and notifications, but cost about $79 for what the presenter considered a relatively small improvement.
• Ultra Code was a disappointment in this run: its design was worse than High’s, despite costing $23.

The presenter explains that Ultra Code acts as an orchestrator: it may deploy sub-agents and additional checks when it judges them useful, but it does not necessarily do so for every task. That helps explain why Ultra Code’s cost and results varied across these examples.

Overall conclusions

• More effort is not automatically better. Simple tasks may already be well served by a low setting, and extra checks can add cost without improving the result.
• Medium or High should be sufficient for many everyday tasks, according to the presenter.
• Max or Ultra Code are better reserved for demanding work, such as polished, client-facing deliverables or projects with a high bar for precision.
• The best setting depends on both the task and the user’s expectations: a functional personal tool does not need the same level of refinement as a production-ready deliverable.
• These are practical demonstrations rather than controlled benchmark results. The presenter emphasizes comparing actual outputs, time, and cost—not just relying on abstract benchmarks.

Guides and demonstrations featured

• A side-by-side website recreation using screenshots.
• A 3D-world creation demonstration.
• A contract-review comparison using annotated Word files and a custom comparison app.
• A quotes-and-invoices app build.
• A practical guide to choosing between effort levels, including how to think about Ultra Code’s use of sub-agents.

Main speaker and sources

• Shubham Sharma — presenter and narrator, describing the prompts, outputs, timings, costs, and conclusions.
• Claude Code — the coding system tested across the six effort settings.
• The presenter also refers to testing and analysis carried out with his team.