The most expensive sentence in any AI vendor pitch is "we use our proprietary large language model."
73% of the time, that sentence is wrong. In a Towards AI study from late 2025, an engineer reverse-engineered 200 funded AI startups by watching their actual network traffic and decompiling their JavaScript bundles. Nearly three-quarters were calling OpenAI or Anthropic on every query. The "proprietary" part was a system prompt telling Claude or GPT-4 to never admit it's Claude or GPT-4.
That's the headline. The expensive part is the math underneath.
When a vendor's underlying API cost on your workload is $40 a month and they're charging you $3,500, the implied markup is 87x. Most SMB owners never run this number because they don't know how. The AI Vendor Cost Reverse-Calculator runs it for you in about 60 seconds.
What the tool does
Three inputs. One verdict.
- Vendor's monthly quote (e.g., $3,500).
- Your monthly query volume (e.g., 5,000).
- Complexity (light, medium, heavy, or RAG/long-context).
- Most likely underlying model (Claude Sonnet 4.6, GPT-4o, Claude Haiku, etc.).
It calculates:
- Per-query API cost on the assumed model.
- Total monthly API cost (what the vendor is actually paying the model provider).
- Markup multiplier.
- Fair-price band based on the typical 3-5x wrapper margin.
- Annual overpayment vs. the fair-price band.
Then it emits a negotiation prompt. Paste the prompt into Claude or any other LLM and you'll get a one-page email back, ready to send to the vendor's account manager. The tone is professional and benchmark-driven, not confrontational. Vendors respond better to "here's the industry benchmark and here's where you sit" than they do to "you're ripping me off."
Three example scenarios
Scenario 1. The 25-person landscaping company quoted $2,400/month
Inputs: $2,400/month, 2,000 queries/month, medium complexity, Claude Sonnet.
Output:
- Per-query API cost: about $0.012
- Monthly API cost: ~$24
- Markup: 100x
- Fair-price band: $72-$120/month
- Annual overpayment vs fair: $27,360-$27,936
Verdict: strongly overpriced. The negotiation prompt drafts an email asking the vendor to explain the gap between $24 in underlying cost and a $2,400 monthly bill.
Scenario 2. The legal-doc-automation tool at $1,200/month
Inputs: $1,200/month, 8,000 queries/month, heavy complexity, GPT-4o.
Output:
- Per-query API cost: about $0.020
- Monthly API cost: ~$160
- Markup: 7.5x
- Fair-price band: $480-$800/month
- Annual overpayment: $4,800-$8,640
Verdict: on the high side but in the gray zone. The negotiation prompt asks the vendor to itemize the legal-domain expertise, the document templates, and the support that justify the markup above the standard wrapper band.
Scenario 3. The customer-support routing tool at $400/month
Inputs: $400/month, 15,000 queries/month, light complexity, Claude Haiku.
Output:
- Per-query API cost: about $0.001
- Monthly API cost: ~$15
- Markup: 26x
- Fair-price band: $45-$75/month
- Annual overpayment: $3,900-$4,260
Verdict: still overpriced. Haiku is the cheapest reasonable model for routing-style tasks, so the underlying cost is genuinely low. The vendor either needs to switch you to a heavier-touch tier that earns the markup, or come down to the fair band.
What the tool deliberately doesn't do
It doesn't pretend to know the vendor's actual gross margin. The vendor has support costs, sales costs, infrastructure costs, customer-success costs. A 3-5x markup is the rough industry benchmark for wrappers, not a contractual rule.
It also doesn't tell you to leave. Sometimes the wrapper is the right pick (see the vendor markup walkthrough for when). The tool gives you the information to decide. Whether you negotiate, switch vendors, or DIY is your call.
Where this fits in the broader vendor-evaluation workflow
The Reverse-Calculator is one step in a five-step evaluation:
- Run the F12 check. Confirms whether the vendor is a wrapper at all. See Before You Pay an Agency $3,500 a Month.
- Run this tool. Gets the markup math.
- Run the 11-question pre-agency self-audit. Covers integrations, data governance, exit terms, and the rest.
- Decide. Negotiate, switch, or DIY.
- If "DIY," run the post-agency migration playbook. 30-day plan to leave cleanly.
The whole sequence takes a couple of hours of an owner's time and a $20 Claude Pro subscription. The agencies pricing themselves around AI markup are pricing themselves against the information gap between the vendor and the buyer. This tool is one piece of closing that gap.
What changed in the AI cost landscape between 2024 and 2026
Three things have moved fast enough that the math is different now than it was 18 months ago.
Model prices fell roughly 10x. What cost $0.06 per 1K output tokens on GPT-4 in early 2024 now costs about $0.01 on GPT-4o or Claude Sonnet 4.6. Most vendors raised their prices anyway. The gap between underlying cost and customer price widened, even when the customer price stayed flat.
Haiku-tier models became actually useful. Tasks that needed GPT-4 in early 2024 (basic classification, routing, simple drafting) now run fine on Claude Haiku 4.5 at a fraction of the cost. Vendors who haven't moved their stack to the cheapest reasonable model are pocketing the difference.
Anthropic launched Claude for Small Business on May 13, 2026. Fifteen ready-made skills, twelve connectors, $20/month plan. The DIY alternative used to require a developer and a couple of weekends; now it requires a Claude Pro subscription and a willingness to read.
Each of these widened the bargaining gap between what vendors can charge and what buyers should pay. The Reverse-Calculator surfaces the gap so it's negotiable.
The deeper version
The full argument for treating AI vendors like any other capital expenditure (with explicit math, transparency requirements, and exit terms) is in The $20 Dollar Agency (Digital Empire series, $9.99 on Kindle). This tool is the tactical 60-second version. The book is the strategic frame.
Related reading
- Before you pay an agency $3,500/month for proprietary AI, the parent post with the F12 check.
- Pre-agency self-audit: 11 questions, the broader contract-evaluation pattern.
- Post-agency migration playbook, if you decide to leave.
- LLM Retrieval Cost Estimator deep dive, the math on the input side.
- Opus 4.7 rankings and early-adopter cost, model-by-model cost comparison.
Fact-check notes and sources
- 73% wrapper finding per Kusireddy's reverse-engineering study on Towards AI, November 2025.
- OpenAI pricing per openai.com/api/pricing.
- Anthropic Claude pricing per anthropic.com/pricing.
- Claude for Small Business launch details per Inc.com's coverage of the May 13, 2026 announcement.
- Wrapper-margin benchmark (3-5x API cost) is informal industry consensus; not codified. Use as a negotiation reference, not a regulatory standard.
This post is informational, not legal, contract-review, or vendor-due-diligence advice. The negotiation prompt is a starting point, not a substitute for counsel on material contracts. Mentions of OpenAI, Anthropic, and other third-party services are nominative fair use. No specific vendor is being accused of fraud. No affiliation is implied.
