← DeepSeek
Jack Roberts
Hermes + DeepSeek V4 = 100X Cheaper
⏱ 21 min video · 3 min read16 May 2026Worth watching
TL;DR
Jack Roberts shows how to combine the Hermes AI agent with DeepSeek V4 via OpenRouter to run powerful multi-model workflows at roughly 1/100th the cost of Claude Opus 4.7. The core strategy is a 'triad' system where Opus 4.7 plans, DeepSeek V4 executes heavy lifting overnight, and ChatGPT 5.5 acts as a critic — giving near-frontier quality output for a fraction of the price.
Key points
1
DeepSeek V4 costs approximately $0.87 per million tokens versus $75 for comparable frontier models, making it ~100x cheaper for bulk or overnight tasks.
2
The 'triad' (or 'Pantheon') system uses Claude Opus 4.7 as the conductor/planner, DeepSeek V4 as the workhorse running 24/7, and ChatGPT 5.5 as the critic — preventing single-model bias and building in iterative improvement loops.
3
OpenRouter acts as a single API key that unlocks all major models, tracks usage in a dashboard, supports fallbacks, smart routing (nitro, exacto), and lets you bring your own provider keys to avoid rate limits.
4
Gemini CLI can be installed via GitHub and gives Hermes multimodal capabilities, including analyzing YouTube video content visually directly from the command line.
5
Hermes is persistent and self-evolving across your entire life (unlike Claude Code which is session-bound to repos) — the more context and tasks you give it via a soul.md file, the better it understands and serves you.
Actionable insights
→
Set up OpenRouter with a single API key, add your DeepSeek API key under 'bring your own keys' to avoid rate limits, and use the nitro or exacto suffixes on model names for speed or tool-calling accuracy respectively.
→
Create an 'Orpheus' persona in Hermes using the triad prompt template: Opus 4.7 decomposes and briefs the task, DeepSeek V4 grinds through execution overnight, and ChatGPT 5.5 brutally critiques the output before it ships.
→
Populate your Hermes soul.md file with your identity, goals, business metrics, revenue targets, and communication preferences so the agent has deep personal context and improves over time.
→
Install Gemini CLI from its GitHub repo and connect it to Hermes to enable multimodal video analysis and other Google Gemini capabilities without needing a separate API key beyond a Google account.
→
Add fallback instructions to your triad prompts (e.g. 'if tokens run out, default to X') so overnight autonomous runs do not silently fail.
Notable quotes
“Would you pay 1% of the price for 95% of the value?”
“WD40 was the 40th version that actually worked, hence the name. Think about how many incremental improvements you get when you have a critic and review agent running around like this.”
“Free is not free as we say in the USA.”
Worth watching?
✅
Worth watching the full video?
Watch if you are actively building with Hermes or want a concrete, step-by-step system for running cheap overnight AI workflows — the key setup steps and triad strategy are all captured here, so skip the video if you just need the concepts.
Saved you some time? The creator still deserves a like.
Found this by searching? That's the hard way.
summree watches the AI channels you follow and sends you the summary the moment a video drops. Nothing gets past you. From £4/month.
Try it free