Video summary

Everyone's Testing Claude Fable 5.1 On Code. It Made Me A 37-Second Film.

Main summary

Key takeaways

Finance

Finance-focused summary

The speaker tests an AI model workflow (Anthropic “Fable 5.1” versus other variants like “Fable,” “Soul,” and “GPT-5.6”) by instructing each model to generate an executive-ready post-acquisition financial model using a discounted cash flow (DCF) in Excel, then convert the results into a PowerPoint deck.

The main finance takeaways center on:

  • Model verification (how easily the work can be audited)
  • Scenario design (base/bearish/bullish outcomes)
  • Assumption traceability (how well sources and logic are linked)
  • How increasing “effort/settings” changes outputs and decision usefulness

Deals / tickers / instruments mentioned

GoPro acquisition scenario

  • GoPro: “GoPro is being bought by a company called Starman.”
  • Starman: referenced as the acquirer (no ticker provided in the subtitles)

Share-price outputs (DCF scenario results) for GoPro

  • Low setting (Fable 5.1 at low effort):
    • Base case: $15/share
    • Bearish: $14/share
    • Bullish: $44/share
  • Extra setting (Fable 5.1 at extra effort):
    • Value per share reported as $1.03 (“a dollar and three cents”)
  • Soul (another model tested):
    • Baseline: $1.21/share

Entry/exit valuation mechanics mentioned

The subtitles describe common deal-model components and checks, including:

  • WACC (weighted average cost of capital)
  • Multiple exit checks
  • Likelihood that the deal would close
  • Financing needs and cash burn considerations

Methodology / step-by-step framework used (as described)

The models are instructed to:

  1. Research the deal (the acquisition transaction)
  2. Build a post-acquisition discounted cash flow (DCF) model in Excel
  3. Create an executive PowerPoint that includes:
    • Transaction overview
    • Assumptions
    • Scenario analysis (base/bearish/bullish)
    • A final recommendation driven by valuation outputs

The workflow also incorporates uncertainty items such as:

  • Cash burn
  • Possibility the combined company needs additional financing
  • Deal close probability

What changes with higher “effort/extra” settings

At higher effort/extra levels, the speaker notes stronger modeling rigor, including:

  • More explicit WACC usage
  • Multiple exit checks
  • Separate timing: “value today vs value at the time of closing”
  • More source traceability, described as linking to “26 different sources”
  • Additional verification artifacts in outputs, such as:
    • Source sheets
    • Checklists
    • Model verification tools described as “checksums” (in the model outputs)

Key finance-specific observations / recommendations

Low-effort runs may be “complete” but harder to verify

The speaker notes that at the low-effort setting the model output can be usable, but verification is harder, including:

A missing “source sheet” and no checklist at the low-effort level.

They can follow formulas, but confirming assumption provenance and whether workbook logic is fully checked takes more work.

Extra-effort runs increase rigor and verification

The “extra” run is described as producing a larger workbook/deck and adding mechanics such as:

  • Deal close likelihood
  • Financing needs
  • Additional valuation mechanics like WACC and exit checks

This run also produced a different per-share value, reported as:

  • $1.03/share (with the subtitles emphasizing timing: today vs closing)

“Soul” vs “Fable 5.1”: emphasis on auditability vs presentation

“Soul” is characterized as having stronger audit features:

  • A separate source sheet
  • A separate checklist
  • A clear passing score
  • Easier-to-review verification artifacts (e.g., sources, checksums)

However, the speaker downplays any claim that the exact stock price output is inherently “the” correct valuation—emphasizing instead how easily another analyst can verify and audit the work.

Caution about “model ranking”

The speaker explicitly avoids claiming a definitive winner:

  • They are not saying one model is “best” in a universal sense.
  • For knowledge work, you often need more than one option.

Price / token-cost numbers (operational constraints)

These are described as cost constraints for running the models to produce analysis (not as finance metrics for GoPro).

Fable 5.1 API token pricing (as stated)

  • $10 per million input tokens
  • $50 per million source tokens (wording in subtitles is slightly unclear; it reads like source-token/output-source pricing)

Cash reading price

  • Dropped to $0.25 per million tokens (versus “a dollar” previously)

Relative cost comparison claims

  • Typical workload costs about 25% less than Fable 5
  • Multi-agent jobs about 45% less
  • Versus “Opus 5”: Fable 5.1 described as about double Opus 5 I/O cost

Subscription note

  • On subscription, “cash price doesn’t matter,” but the speaker claims limits are tighter than OpenAI’s.
  • Token efficiency still matters to keep tasks within those limits.

Disclosures / disclaimers (as reflected in subtitles)

  • No explicit “not financial advice” language appears in the subtitles.
  • The speaker says they are not claiming which model provides the “correct” valuation for GoPro.
  • Key deal inputs were incomplete (e.g., Starman does not publish the financial history typically required).
  • The emphasis is on scenario-building under missing information and verifiability rather than definitive valuation accuracy.

Presenters / sources mentioned

  • Nate (the speaker)
  • Claude / Claude Fable 5.1 / Fable 5
  • GPT-5.6
  • Soul
  • Anthropic (pricing/efficiency claims and model comparisons)
  • OpenAI (mentioned as a next comparison; Astra referenced)

Original video