DeepSeekV4 + Claude Code = 100X Cheaper

Tech / product concepts covered

• Goal: Combine Claude Code with DeepSeek V4 to run Claude Code workflows at a much lower cost (claimed “100X cheaper”).

DeepSeek V4 details (as stated)

• 1.6T parameters
• ~1M context window
• MIT open weights
• Strong “tool calling” ability, comparable to Anthropic/Claude’s workflow but at lower cost.

Cost comparisons / promos (as stated)

• Claude Opus session: ~$62.50
  • Note: subtitles imply “$6250” but surrounding context suggests $62.50; exact meaning is unclear.
• DeepSeek V4: described as significantly cheaper, with a promo mentioned.
• DeepSeek V4 Flash: ~$0.20 for a 4-hour session, compared to Claude Opus pricing.

Benchmarks + “does it code in practice?”

• Mentions benchmark testing and comparisons against models like Sonnet 4.6 / Opus 4.7.
• Tools mentioned include Terminal Bench and MCP Atlas.
• Emphasis is placed not only on benchmark scores, but on showing the workflow in a live coding / development environment.

Core setup / tutorial: Claude Code + DeepSeek via a GitHub “dual terminal” workflow

Key idea: “dual terminal split-up”

A workflow that splits responsibilities across terminals/windows:

• Run a free model in one terminal/window
• Run DeepSeek in another
• Run Claude Code in a third
• Then switch/tag models depending on the task type, sometimes simultaneously

Problem with OpenRouter (as stated)

• Using DeepSeek through OpenRouter causes heavy rate limiting, described as unusable for the intended workflow.

Preferred approach to avoid rate limits: use DeepSeek’s API directly

Steps described:

1. Create a DeepSeek account on platform.* (subtitle shows “platform.dec”).
2. Add $2–$5 (or get free access via an OpenRouter workaround, with caveats).
3. Retrieve a DeepSeek API key.
4. Clone a trending GitHub repo that acts like a proxy to integrate DeepSeek with Claude Code-style tooling.
5. In the dev environment (example given: Anti-gravity as the IDE/workspace), install/clone the repo and configure it with the DeepSeek API key.
6. Optionally enable “dangerously skip permissions” (explicitly mentioned as optional).

Live tests / features demonstrated

Verifying model routing

• A test prompt is used to confirm the integration works.
• Clarification: if asked “who is the tallest man,” the system may reference Claude Sonnet due to “data it reads,” not because the underlying model truly is Claude.

Skill / tool execution (“skills” work)

• Demonstrates using a bit.ly skill to shorten a URL.
• Claim: compared to prior systems that complained about poor skill usage, this setup successfully executes the skill via tool calling.

Task routing / when to use which model (practical guidance)

• Claude: best for creative / visual / design
  • Positioned as the “design king”
• DeepSeek: best for non-visual backend / algorithmic work
  • automations, quick scripts, unit tests, “dirty work,” refactors, docs, code audits
• Claude Code also used in the pipeline:
  • Suggested strategy: Claude handles design/creative, DeepSeek does heavy lifting, and optionally other models handle additional steps.

Website build walkthrough (tutorial example)

Design system + repo-driven approach

• Uses a design system repo (examples mentioned: BMW/Apple) to provide visual identity/framework components.
• Workflow:
  • Clone an “awesome designs” repo
  • Prompt Claude to build a website “using Apple design framework” and include components from other repos
  • Delegate heavy lifting (e.g., complex functionality) to DeepSeek

Example feature built with DeepSeek

• A multi-variable ROI calculator
• Inputs include:
  • caffeine consumption
  • productivity
  • hourly rate
  • hours slept
  • caffeine sensitivity
• Output includes computed “annual ROI” values and a working UI.

UI / style preservation claim

• DeepSeek can replicate exact HTML/style when given an existing design/style guide.

Additional tip

• Use an HTML website extractor (mentioned: “Google HTML website extractor”) to copy real website source for inspiration.

Optional expansions mentioned

Free-model alternative via OpenRouter

• Integrate an OpenRouter API key into the same Claude Code / dual-terminal workflow.
• Mentions cycling through free models (example text in subtitles: “Qwentry coder, Nvidia, Needron 3 Super”).
• Still warns about rate limiting and cost/quality tradeoffs.

Code review step via ChatGPT

• Proposed pipeline:
  • Claude for design → DeepSeek for heavy lifting → ChatGPT (code review) to improve correctness and ensure the code works
• Claimed benefit:
  • Leverages an existing ChatGPT Plus subscription
  • Mentions “free reviews” in practice (as described)

Limitations / cautions (analysis-style notes)

• Data sovereignty / geopolitics caution:
  • Avoid if you need to discuss sensitive topics (example cited: Tiananmen Square).
• Data handling caution:
  • Do not paste naked URLs or API keys into DeepSeek/Claude chats (described as a major “no-no”).
• UI taste limitation:
  • DeepSeek is described as not great at initial UI unless it’s working inside an existing UI framework.
• Regulated industries caution:
  • Avoid using DeepSeek for corporate, healthcare, finance, or highly regulated data/IP scenarios.

Who the main speaker/source is

• Jack Roberts (the video creator; described as the speaker and system builder).