FunnelCrafter· AI cost & pricing review · July 2026

What each lead magnet actually costs us to generate.

Real numbers for Eben & Paul before we wire up billing. Built from Anthropic's published API pricing (validated today) and token usage measured from your 20 completed runs across 4 real books — not estimates pulled from the air.

✓ Verdict: the managed-key model is safe — AI is 3–8% of subscription price

Cost per run · median

$0.27/run

Sonnet 4.6, standard pricing. Range $0.26–$0.44 across smallest → worst-case run.

01 · Validated pricing

Today's Anthropic API rates

Pulled from Anthropic's pricing docs today. Our configured model is Sonnet 4.6. Sonnet 5 is cheaper only during its intro window — its new tokenizer emits ~30% more tokens, so it flips more expensive after Aug 31.

Model	Input /Mtok	Output /Mtok	Cache read	Batch (−50%)	Note
Sonnet 4.6 in use	$3.00	$15.00	$0.30	$1.50 / $7.50	Our current model
Sonnet 5 — intro	$2.00	$10.00	$0.20	$1.00 / $5.00	Through Aug 31, 2026 · +30% tokens
Sonnet 5 — standard	$3.00	$15.00	$0.30	$1.50 / $7.50	From Sep 1 · +30% tokens
Haiku 4.5	$1.00	$5.00	$0.10	$0.50 / $2.50	Lever for cheap sub-tasks

Prompt-cache reads cost 0.1× input; the Batch API is 50% off but is async (not usable for live generation).

02 · Measured usage

What a real run consumes

Measured directly from the database: the source books fed in, and the content generated out across every completed run. This is the ground truth the cost model is built on.

Runs measured

completed jobs, 4 books

Book input

5.7–12.8k

tokens · median ~8.4k

Output / run

~5.1k

median · up to ~8k

Emails / run

3–7

median 3 · config-driven

03 · Cost model

Where the money goes, call by call

A run is 6–13 Claude calls. Below is the median run (8.4k-token book, one critique round, 3 emails) priced at Sonnet 4.6. The draft + rewrite dominate; emails and page copy are cheap.

Pipeline step	Input tok	Output tok	Cost
Main draft (rules + brand + full book)	11,109	4,000	$0.093
Critique pass ×1	4,500	800	$0.026
Rewrite-from-scratch	5,100	4,000	$0.075
Landing-page copy	4,500	560	$0.022
Thank-you copy	1,200	130	$0.006
Email sequence ×3	6,600	1,650	$0.045
Median run total	33,009	11,140	$0.266

Cost per run by scenario

Cheapest

$0.26

5.7k book · 3 emails

Median

$0.27

planning number

Typical-max

$0.34

12.8k book · 7 emails

Worst case

$0.44

big book · 7 emails · 2 critiques

On Sonnet 5 intro pricing these are ~15% lower ($0.22 median) until Aug 31 — then higher than 4.6.

04 · Unit economics

Cost vs. subscription price, per tier

The per-tier run cap is the cost governor — it bounds the most the company can ever spend per user. Green is gross margin after AI; the striped sliver is the AI cost. Prices are indicative — confirm against live Stripe.

Free / trial2 runs/mo

AI $0.53–0.88/mo · absorbed as trial CAC

Starter5 runs · $29

92–95% margin$1.33–2.20

Growth10 runs · $79

94–97% margin$2.66–4.40

Pro20 runs · $149

94–96% margin$5.32–8.79

Unlimited20 runs · $199

96–97% margin$5.32–8.79

Gross margin after AI AI cost (worst-case run × cap)

Even at the worst-case $0.44/run every single run, AI never exceeds ~8% of a tier's price. The run cap means a user physically cannot cost more than (cap × $0.44).

05 · Levers

Ways to push cost down further

APrompt caching (biggest at scale). The static writing-rules and the library book are re-sent on every run. Caching them drops repeat-book runs to ~$0.24, and at volume (many users pulling the same library books) cache reads are 0.1× input — a large input saving.
BHaiku for the sub-tasks. Landing copy, thank-you copy and emails don't need Sonnet. Routing just those to Haiku 4.5 takes the median run from $0.27 → $0.22 (−18%) with no quality loss on the core draft.
CModel choice is a dial, not a default. Sonnet 4.6 is the right balance today. Sonnet 5 only wins on price during the intro window; after Aug 31 its +30% tokenizer makes it dearer for this workload.
DThe run cap already protects you. No caching needed to be safe — margins are 92–97% before any optimization. Treat A & B as upside, not prerequisites.

06 · Recommendation

The credential model to ship

Friction-free managed default: nobody hits an "add your API key" wall. Who pays for what:

AdminsEben · Paul

Their own Claude Pro subscription via the CLI token (already live). Restricted to admins — keeps subscription use as legitimate individual use.

Trial & paid usersfree → pro

The company API key by default — the managed service, cost baked into the subscription. Their run cap bounds the spend. Optional "bring your own key" stays available for power users.

Cost controlbuilt-in

Per-tier run limits are the hard cap. Reserve ~15% of revenue as AI COGS and margins hold at 90%+.

One action to get exact numbers: add the company ANTHROPIC_API_KEY to the Modal secret. The moment it's live, every run logs real input/output tokens to ai_usage_log — replacing this model with metered truth to finalize pricing.

07 · Methodology & caveats

How honest are these numbers?

Grounded, but a model — not a metered bill. What's measured vs. estimated:

✓Measured: every book's input size and every run's output size, pulled straight from the DB (20 runs, 4 books). The static writing-rules prompt (732 tok) is measured from source.
~Estimated: brand/voice context (~1,500 tok) and the intermediate draft/critique sizes that aren't persisted. Token counts use the ~4-chars/token rule (±~15%).
~Indicative: tier prices ($29–$199) are from marketing — the .env currently holds placeholders. Confirm live Stripe amounts.
!Why not exact: ai_usage_log is empty — every run so far used the CLI, which isn't metered. Enabling the company API key fixes this immediately.

FunnelCrafter · AI cost & pricing analysis · prepared for founder review, July 2026. Figures computed from real usage; treat as directional until ai_usage_log confirms.

Sources: Anthropic API pricing docs · Sonnet pricing announcement