First, separate model from harness
In the two DeepSeek Harness demonstrations, the presenters describe the model as the “brain” and the harness as the surrounding context, tools, plugins, permissions and memory. Jack Roberts calls the new harness a model-agnostic alternative to Claude Code; Sharbel A. demonstrates configuring another provider, GLM, rather than being locked to DeepSeek. These are the presenters’ explanations and demos, not an independent technical audit. Jack, 0:31–1:09 · Sharbel, 0:28–0:57 · Sharbel, 3:04–3:29
Jack’s April video describes the other route: combine the Claude Code workflow with DeepSeek V4. That keeps the familiar coding environment while changing the model being used. It is a different proposition from adopting a new harness. Jack, 0:00–1:51
What the creator recommends for each kind of work
Jack’s April walkthrough recommends Claude for design-heavy work and polished interfaces, DeepSeek for automations, quick scripts, algorithmic tasks and unit tests, and Codex for review. In his August harness video, he makes a similar distinction: use Claude or Codex when client-facing polish matters; use the cheaper “workhorse” option for volume-heavy tasks where price matters. April, 7:43–8:16 · August, 8:20–9:47
That is one creator’s workflow advice repeated across two videos—not two independent evaluations. It is useful as a hypothesis for choosing a task, not proof that one model is consistently better at a category of work.
Treat the price headline as a dated example
The April video’s title says “100X Cheaper.” In the narration, Jack compares a quoted $62.50 Claude Opus session with about $0.20 for a DeepSeek V4 Flash session, while noting a promotion due to end on May 5. The August video later describes DeepSeek as delivering “90–95%” of the performance for a fraction of the cost. Those are the creator’s time- and setup-specific claims; the two videos come from the same channel. April, 1:16–1:50 · August, 4:48–5:05
The corpus does not provide a matched task, billing record or independent benchmark that would make those numbers a general cost or quality comparison. Don’t treat the title or quoted session amounts as a reusable price ratio.
A practical way to use this
If you want to keep your existing coding workflow, test a model substitution. If you want to change how tools, plugins or permissions are composed, test a separate harness. Compare them on the same small set of tasks and inspect the outputs and actual bills before moving important work over. That is a practical next step—not a result established by these videos.
What this sample does not establish
This is a synthesis of creator demonstrations, not a controlled benchmark. It does not establish current product availability, today’s prices, privacy or security properties, or general coding quality. No outside web research has been added.
A transcript search found 26 videos that co-mentioned Claude Code-like and DeepSeek-like terms. This sample uses the three videos whose main subject is the direct workflow or harness comparison. The other co-mention candidates still need relevance screening; they are not counted as evidence here. This is not yet a complete topic review.
Sources
- Jack Roberts, “DeepSeekV4 + Claude Code = 100X Cheaper”, 2026-04-30. Creator walkthrough; relevant sections at 0:00–1:51, 7:43–8:16, 15:34–16:37.
- Jack Roberts, “DeepSeek Harness = Claude Code for $0”, 2026-08-21. Creator demonstration and personal workflow advice; sections at 0:31–1:40, 4:48–5:05, 8:20–9:47.
- Sharbel A., “DeepSeek Harness Setup: A Free Claude Code You Own In 10 Minutes”, 2026-08-28. Creator setup demonstration; sections at 0:00–1:06 and 3:04–3:44.
Source dates and metadata come from the corpus records. Timestamp links point to the relevant video sections. Product, pricing and performance claims are attributed to the presenters and were not independently verified.