← Claude
Jack Roberts
100 Cheap AI Agents vs 1 Expensive AI Agent
⏱ 21 min video · 3 min read9 Jul 2026
TL;DR
Jack Roberts runs Claude Opus 4.8 and Claude 4 (referred to as 'Fable 5' in the transcript, likely Claude 4 Sonnet or a similarly positioned top model) through four practical tests — cold email writing, website design, and dashboard building — to determine when the expensive flagship model is actually worth paying for versus cheaper alternatives. The core finding is that model capability differences are smaller than expected, and prompting strategy plus smart model routing matters far more than always using the most powerful model.
Key points
1
Claude Fable 5 (flagship) costs ~$45 per 2M tokens in / 500K out, versus ~$3.36 for GLM 5.2 and ~$1.30 for DeepSeek — a massive price gap for often marginal output differences.
2
In cold email and website design tests, Fable 5 produced slightly better results than Opus 4.8, but the difference was subtle enough that most people would not notice.
3
Opus 4.8 paired with a 'ragtag team' of sub-agents (Gemini catching mobile issues, DeepSeek improving copy, ChatGPT) produced competitive results against solo Fable 5, demonstrating the power of multi-model orchestration.
4
The swarm approach — many cheap models without an intelligent orchestrator — performed poorly, confirming that a capable orchestrating model is essential for multi-agent systems to work.
5
Jack recommends reserving Fable 5 for taste-sensitive work (initial design), strategic one-way-door decisions, and debugging, while delegating edits, volume work, and grunt tasks to cheaper models like Opus 4.8 or DeepSeek.
Actionable insights
→
Use Fable 5 only for high-stakes tasks: initial design, strategic decisions, and final debugging — roughly 5% of your total model usage. For everything else, Opus 4.8 or DeepSeek is sufficient.
→
Reduce token costs by using tools like Firecrawl to scrape clean text (not HTML), trimming context windows by opening a new window after each completed task, and keeping your core .md system files lean.
→
Use the 'spec first, run once' principle: spend time perfecting your first prompt — possibly sparring with a cheaper model first — because quality and cost degrade significantly with back-and-forth correction loops.
→
Bring in secondary models (Gemini, DeepSeek, ChatGPT) as cross-referencing sub-agents under a smart orchestrator to catch mobile issues, improve copy, and find bugs — this is described as 'the creatine of the AI world.'
→
Always have Codex (or another model) review all code before shipping — Jack says models repeatedly told him everything was fine until Codex flagged critical issues.
Notable quotes
“I think I spent more in Fable 5 tokens asking if it was Greek yogurt than the actual price of the Greek yogurt. Like that is where we are at right now in 2026.”
“Using Fable 5 without knowing the right strategies is like giving somebody a level 99 account and dropping them in a game, but they have no experience of playing. They are a complete noob.”
“Bringing in a secondary model is one of the most overpowered things that you can possibly do. It is like the creatine of the AI world.”
Worth watching?
⏭️
Worth watching the full video?
The key framework and cost-routing strategy are fully captured here — watch only if you want to see the side-by-side website and dashboard visuals Jack references, which are harder to convey in text.
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