Command Code: compare DeepSeek V4 and Claude Opus 4.8 on the same job

Set up a comparison that can be traced

The 16 June 2026 description compares Claude Opus 4.8 and DeepSeek V4 inside Command Code. Its four tests cover bug hunting, a landing-page redesign, a scoped button implementation and planning a large change. Use the same starting code and request to make a comparison interpretable.

1. Open https://commandcode.ai/?utm_source=youtube&utm_medium=firstvideo&utm_id=luke and follow the setup segment at 1:27.
2. Save the repository state before each model’s attempt and record which model and branch are active before changing or resetting anything.

Check bugs and redesigns against behaviour

A wide list of possible bugs and a precise diagnosis are different outputs. Check each claim against the affected code and a reproducible action. For the redesign, review the user’s task as well as the appearance; an attractive page can still make an important route harder to complete.

1. Record the bug location, expected behaviour and evidence before counting a finding as confirmed.
2. After the landing-page change, follow the main visitor flow and inspect the diff for unrelated changes.

Separate a scoped implementation from a proposal

The button task at 15:12 tests a bounded change; the larger planning task begins at 18:06. Review these against their different deliverables. A detailed enhancement proposal is not an implemented feature, and a completed button does not validate a wider product plan.

1. Define the button’s acceptance criteria before checking the result.
2. For the large change, map the user flow, assumptions and bottlenecks, then separate proposed work from changes actually made.

Attach cost conclusions to the recorded run

The description claims that Claude used 70x more credits than DeepSeek in this comparison and discusses choosing precision or lower cost. Record credits alongside task quality and follow-up work. That historical ratio does not establish current model prices, equal token usage or a universal best model.

Keep a compact result sheet for all four tasks: starting state, requested outcome, observed change, necessary follow-ups and credits consumed. This makes it possible to distinguish a cheap first response from an inexpensive completed task when reviewing the comparison later.

The takeaway

Choose from verified task results and the cost of reaching them, keeping implemented changes separate from plans and suspected bugs.