Costs and margins
What a finished article actually costs to produce, built from the model rate card in the Austen repository and from Proofed's own published editing price. The model spend is small and knowable. The editing is 98% of the cost of goods, and it is the number you must replace with a measured one first.
- $0.79 of model spend per finished 1,500-word article, including a hero image, whose square crop is the OG image.S9
- $52.50 of human editing per 1,500-word article, at Proofed's own published retail rate of about $0.035 a word.S10 This is an assumed transfer price, not a measured cost.
- 77% gross margin on the $2,500 pilot paid on a UK card, 74% on a US card once Stripe's international and conversion fees are counted.S37
Where the model rates come from
Austen prices every AI call it makes. config/ai_pricing.php is the price book: AiPricing reads it, AiMeter stamps a dollar figure on every recorded call, and /admin/costs aggregates it. The file carries a version, 2026-09-23, and a model missing from it is recorded unpriced and flagged rather than treated as free.S5
The models below are the defaults in config/llm.php. Production can differ: an environment variable or an admin override at /admin/llm-settings changes the model for any purpose, and the price book itself notes that production's standard model may still be GPT-5.2 ($1.75 in, $14.00 out). Since 23 September Claude Opus 5.5 ($4.00 / $20.00) and GPT-6 Astra, Sol and Luna are routable too. Check the routing page before trusting this table. Every model now runs at high effort, which is what these defaults already ran at.S5
| What runs where | Model | Rate |
|---|---|---|
| Article generation and every revision of the draft, competitor analysis | claude-opus-4-8 (the power model) | $5.00 per million input tokens, $25.00 per million output |
| Plans, briefs, SEO and GEO scoring, translations, social, derivatives | claude-sonnet-5 (the standard model) | $2.00 per million input, $10.00 per million output |
| Latency-sensitive inline calls, including keyword suggestion | claude-haiku-4-5 | $1.00 per million input, $5.00 per million output |
| Hero image, medium quality (the default), 1536x1024, on request | gpt-image-2 | $0.041 per image ($0.165 at high quality; the OG image is a square crop of the hero, not a second image) |
| Research grounding | Server-side web search | $0.01 per search ($10 per 1,000) |
| Transcription | gpt-4o-transcribe-diarize | $0.006 per minute of audio |
| Embeddings | text-embedding-3-small | $0.02 per million tokens |
Scoring deliberately does not run on the power model. The quality gate can score an article up to six times, and running that on Opus would multiply the largest recurring cost in the pipeline for no quality gain.S6
Model cost of one finished 1,500-word article
The pipeline steps are read from ArticleGenerationPipeline and ArticleQualityGate. The token counts are estimates and are the weakest part of this table.S8 Both improvement passes are counted because they are the norm: in the 60 days to 30 September 2026 only 3 of 35 articles reached the 90 GEO target, so the gate almost always spends both.S6
| Step | Model | Input tokens | Output tokens | Cost |
|---|---|---|---|---|
| Brief drafting | claude-sonnet-5 | 3,000 | 1,200 | $0.018 |
| Plan generation | claude-sonnet-5 | 4,000 | 1,500 | $0.023 |
| Research: 5 web searches plus synthesis | web search + claude-sonnet-5 | 6,000 | 2,000 | $0.082 |
| Article draft | claude-opus-4-8 | 12,000 | 4,000 | $0.160 |
| SEO and GEO scoring, three rounds each | claude-sonnet-5 | 36,000 | 7,200 | $0.144 |
| Improvement passes, both of the two | claude-opus-4-8 | 24,000 | 8,000 | $0.320 |
| Text and search subtotal | $0.747 | |||
| Hero image, medium, 1536x1024 | gpt-image-2 | 1 image | $0.041 | |
| OG image, square crop of the hero | none | no model call | $0.000 | |
| Total model cost per finished article | $0.788 | |||
Replace this table in week one
Open /admin/costs and put the measured average cost of an article here. Every AI call is already metered and priced on main, production is already generating articles, and enforcement is correctly still off. php artisan credits:usage --models adds a per-account view once gtm/pilot-offer is merged.S7 There is no excuse for this staying an estimate.
The editing cost, stated honestly
This is the assumption the whole margin model rests on, so it gets its own section rather than a footnote.
| Basis | Figure | Per 1,500-word article | Type |
|---|---|---|---|
| Proofed's published retail price | $105 for 3,000 words at one-day turnaround, about $0.035 a wordS28 | $52.50 | Published price, used as an assumed cost |
| Proofed's published managed-service rate | $65 an hour, falling with retainer sizeS28 | $33 to $65 at 30 to 60 minutes per article | Published rate, applied to an assumed duration |
| What Austen charges the customer for it | $55 per 1,000 words, $15 minimumS2 | $82.50 | Published |
| True group cost of an editor's time | Not published anywhere and not in this repository | Unknown | Must be measured |
This guide uses $52.50 per 1,500-word article, Proofed's own retail rate, as the cost line. That is deliberately conservative in two directions at once. It is too high, because Proofed and Austen are the same owner and the real group cost is the editor's time, not the retail price. It is also the only defensible number available, because the editor cost is not published and inventing one would be exactly the kind of number this guide refuses to make up.
Get the real editing cost out of Proofed's books
Cost per 1,000 words of editor time, at the article lengths and turnaround Austen actually sends. One number, from your own accounts. It moves the pilot's gross margin somewhere between 77% and about 95%, which is the single largest uncertainty in this guide.
The customer-facing editing add-on at $55 per 1,000 words against a $35 assumed cost is a 36% gross margin on that line. That is thin for a software company and normal for a service. It is another reason the pilot and the retainer, where editing is bundled and hidden, are better business than selling editing as an add-on.
The margin table
Direct cost only: model spend, editing at the assumed rate, and Stripe fees. It excludes your time, hosting and everything in the running-costs table below. Stripe is modelled at 1.7% on a UK card and 5.15% on a US card, being 3.15% plus 20p plus 2% currency conversion.S37
| Offer | Revenue | Model | Editing | Stripe | Gross profit | Margin |
|---|---|---|---|---|---|---|
| Pilot, 10 articles, UK card | $2,500 | $7.88 | $525.00 | $42.50 | $1,924.62 | 77.0% |
| Pilot, 10 articles, US card | $2,500 | $7.88 | $525.00 | $128.75 | $1,838.37 | 73.5% |
| Retainer, 8 articles at $2,000 | $2,000 | $6.30 | $420.00 | $103.00 | $1,470.70 | 73.5% |
| Retainer, 8 articles at $2,500 | $2,500 | $6.30 | $420.00 | $128.75 | $1,944.95 | 77.8% |
| Starter $99, all 15 articles used | $99 | $11.82 | none | $5.10 | $82.08 | 82.9% |
| Starter $99, 8 articles used | $99 | $6.30 | none | $5.10 | $87.60 | 88.5% |
| Pro $299, all 50 articles used | $299 | $39.40 | none | $15.40 | $244.20 | 81.7% |
| Pro $299, 25 articles used | $299 | $19.70 | none | $15.40 | $263.90 | 88.3% |
| Editing add-on, per 1,000 words | $55 | none | $35.00 | excluded | $20.00 | 36.4% |
| Free plan, 5 articles lifetime | $0 | $3.94 | none | none | -$3.94 | Acquisition cost |
What breaks the margin
| Risk | What it does | The guard |
|---|---|---|
| A heavy account on Pro | The 50-article allowance is enforced, but the AI spend around it (optimisation, images, translations, derivatives) is measured and never capped, because credits are not enforcedS7 | Read /admin/costs every Friday. Set an allowance from the measured data before turning enforce on, never the other way around |
| Long articles on a retainer | Editing is per word, so a 3,000-word pillar doubles the editing cost on that piece | The retainer definition counts a 3,000-word article as two. Enforce it (Chapter 3) |
| Rewrites after delivery | A second full edit is a second editing charge with no revenue | Brief approval before drafting, and one edit included. Both are stated in the pilot agreement (Chapter 12) |
| Rush turnaround | Proofed's own urgent rate is three times standard: $315 against $105 for 3,000 wordsS28 | Never promise under five working days without pricing it |
| Paying in USD, collecting on international cards | Up to 5.15% instead of 1.7%S37 | Offer invoicing for larger deals: 0.4% capped at $2.00 per invoice |
| Trim and tell-rewrite passes firing on top of the two improvement passes | Up to four more power-model calls, about $0.16 each | Nothing. At worst it is $6.40 across a whole pilot. Ignore it |
Business running costs
These sit outside the per-article model. Amounts are placeholders until invoiced, and each is labelled.
| Item | Cost | Type | Note |
|---|---|---|---|
| Application hosting and queue workers | Existing | Not separately measured | Austen already runs in production. There is no incremental hosting cost from selling ten more pilots |
| Stripe | 1.7% UK, up to 5.15% international; Invoicing 0.4% capped at $2.00S37 | Published | Modelled per deal above |
| Upwork Agency Plus | $20 per month, plus 15% to 20% all-in on won workS30 | Published | The one marketplace worth paying for |
| Fiverr Pro | No listing fee; 20% seller commission on every orderS31 | Published | A $2,500 pilot sold there nets $2,000, which is still above the $2,000 discount floor. Just |
| Proofed editorial capacity | Internal | Must be measured | The real constraint on the plan, not a cash cost |
| Your time | 8 to 12 hours per pilot | Estimate | Measure it on the first three and put the real figure here |
The four numbers to measure before month three
- Real model cost per finished article, from /admin/costs, where it is available now. Replaces the whole estimate table above.
- Real editing cost per 1,000 words, from Proofed's own books at the lengths Austen sends. Replaces $0.035 a word.
- Average delivered article length. Everything on this page assumes 1,500 words. If the real average is 2,200, the editing cost rises by 47% and the pilot margin falls to about 67%.
- Hours per pilot, measured. It decides how many pilots and retainers you can carry at once, which is the binding constraint in Chapter 5.