Video summary

Can a $0.30 AI Beat the World's Best Coding Model? Fable 5 vs DeepSeek V4 Flash

Main summary

Key takeaways

Product Review

Overview

The video compares how well different AI coding models build the same real-world project: a simple, functioning Sales CRM (like Salesforce/Pipedrive) from requirements in agents.md. It uses Pi as the coding-agent interface for most runs (with models such as DeepSeek V4 Flash and Fable 5), and evaluates whether the final app meets success criteria (unit tests, phases completed, and functional/visual checks).

Product reviewed (what the “best cheap coding model” is)

DeepSeek V4 Flash (used through OpenRouter / Pi as an underlying LLM for coding agents).

Key features (as evidenced by the built CRM)

The winning run produced a full CRM app with multiple sections:

  • Dashboard
    • Callouts
    • Activity/tasks
  • Organizations
    • Edit/delete controls (described as slightly janky)
  • Contacts
  • Deals
  • Pipeline
    • Drag-and-drop between pipeline stages
    • Totals updating correctly

It appears to run with an agent workflow that includes self-testing capabilities (via Pi’s “agent browser skill”).

Overall output quality is described as “perfectly decent” rather than top-tier polished UI.

Performance / outcomes (from the video)

DeepSeek V4 Flash run

  • Time to complete: ~15 minutes
  • Declared victory: all phases completed (per its run summary)
  • Produced a working CRM that supports key interactions, including drag-and-drop

Comparison timing (other models)

  • Fable 5: 36 minutes
  • Codex (GPT 5.5): 1 hour 2 minutes
  • Claude Sonnet 5 (in Claude code): declared victory (time not specified)
  • Qwen 3.6 (27B): not shown as successfully completed; hit failures/crash
  • (Earlier mentioned) DeepSeek V4 Pro: previously “pretty good” (cost discussed later)

Costs (major focus of the video)

These are direct “cost to build the whole CRM” figures mentioned:

  • Fable 5: $27.69
  • Opus 48: $21.27
  • Codex (GPT 5.5): “around the same” as Opus 48 (exact value not clearly given; described as an estimate assuming token caching)
  • Claude Sonnet 5: $18.66
  • GLM 5.2: $4.15
  • DeepSeek V4 Pro: $2.56
  • Qwen 3.6 (27B): $1.91
    • Noted it may be runnable locally if you have sufficient RAM; then it would be “no charge beyond electricity”
  • DeepSeek V4 Flash: $0.30 (the “shocking reveal”)

Claimed conclusion: DeepSeek V4 Flash is ~100× cheaper than Fable 5 and is positioned as #1 on OpenRouter’s programming models list (per the narrator).

Pros (DeepSeek V4 Flash)

  • Extremely low cost: $0.30 for a complete working CRM
  • Fast completion: ~15 minutes
  • Functional quality is solid
    • drag-and-drop works
    • totals update correctly
    • major CRM sections work end-to-end
  • Produces a product described as “decent” / “great” / “really decent” despite not being the most polished UI

Cons / limitations

  • UI/UX is not the most fancy/polished
    • dashboard is “not the most fancy”
    • organizations edit/delete is described as slightly janky
  • Even if it works, it may not match the “professional” look of Fable 5

Comparison summary (how it ranks vs others)

Best overall quality / polish

  • Fable 5: described as very professional, with strong UI and strong test/phase results
    • Longer (36 min) and more expensive ($27.69)

Cheapest that still works well (main recommendation)

  • DeepSeek V4 Flash: $0.30, fast, working CRM, correct drag-and-drop behavior

Best “compromise” price/perceived quality

  • GLM 5.2: $4.15
    • described as a best-of-both-worlds option—cheap but still produces a nice implementation

Notable failures / issues

  • Qwen 3.6 (27B)
    • drag-and-drop bug; interaction fails
    • later appears to have server crash
  • General note (narrator): several other frontier models cost much more while being only modestly better in this experiment.

Unique points mentioned (distinct product-related takeaways)

  1. DeepSeek V4 Flash builds a full sales CRM (dashboard, organizations, contacts, deals, pipeline).
  2. Pipeline drag-and-drop works; totals update properly.
  3. Organizations edit/delete UI is somewhat janky but functional.
  4. Dashboard is decent but not highly fancy.
  5. DeepSeek V4 Flash completes in ~15 minutes.
  6. DeepSeek V4 Flash costs $0.30—the central “100× cheaper than Fable 5” claim.
  7. The video asserts DeepSeek V4 Flash is ranked #1 on OpenRouter for programming models.
  8. GLM 5.2 is positioned as the best quality/price compromise ($4.15).
  9. Fable 5 is positioned as highest quality ($27.69) but far more expensive and slower.
  10. Qwen 3.6 may be runnable locally (if enough RAM), but failed due to drag-and-drop/server crash.

Speaker/approach notes

The single main narrator/creator conducts all runs and provides:

  • Agent workflow setup perspective (Pi, agent skill, “agent browser” testing)
  • Cost-accounting perspective (token cost estimates and per-model build cost)
  • Results interpretation perspective (quality/readability and functional correctness)

Aside from the narrator, no distinct additional speakers are clearly identified in the subtitles.

Concise verdict / recommendation

  • Cheapest that still produces a functioning, non-broken CRM: DeepSeek V4 Flash ($0.30, ~15 minutes, working drag-and-drop)
  • Maximum quality/polish: Fable 5 (higher cost and longer runtime)
  • Balanced option: GLM 5.2 is recommended as the best compromise

Original video