Vanta CEO on AI Token Markups: Why Per-Token Pricing Loses to Claude
By Matt Cullerton, founder of Mavric and Apso
Vanta CEO Christina Cacioppo argues that AI applications should not bill customers per token, because customers will compare that price with Anthropic's list price and go try Claude directly. A TikTok from @iltb_podcast cuts the moment under the title "You Have Just Marked Up a Claude Token 4x." The 4x is a customer objection she acts out, not a measured price, and Vanta's own policy is to absorb token costs instead of reselling them.
Where is the clip from?
The speaker is Christina Cacioppo, co-founder and CEO of Vanta, on Invest Like the Best with Patrick O'Shaughnessy. The episode is EP.496, "Christina Cacioppo - Building an Enduring Software Company in the Age of AI," published on 20 October 2026 and running about 57 minutes. Colossus, the show's publisher, lists it as "Becoming a Category Winner." The TikTok was posted five days later.
Vanta was founded in 2018 by Cacioppo and Erik Goldman and is based in San Francisco. Fortune reported in April 2026 that it had reached $300 million in annual recurring revenue with more than 16,000 customers. Sacra puts that at 69 percent year-over-year growth. Its last disclosed valuation was $4.15 billion, after a $150 million Series D led by Wellington Management in July 2025.
What did Cacioppo say about token markups?
She describes what Vanta hears from customers: "if you try to charge them based on tokens, they are very quick to say, oh, you've just marked up a Claude token 4x." She continues in the customer's voice: "Are you 4x as good as Claude? I don't know. Let me go try it out."
Her conclusion is that "if you're trying to mark up at all, users are very smart and you will lose on that comparison," and that "anyone who is marking up Claude tokens 4x" and passing that on to users is making a mistake.
The 4x is her example of a customer complaint. She names no company and quotes no price, and the clip shows no pricing data.
How does Vanta price AI features?
The episode's show notes say Vanta "won't mark up a single token." BigGo's coverage quotes her: "I think charging tokens themselves really puts you up against Claude." Scuttlebutt's summary says Vanta "eats the AI token cost rather than disguise a markup as 'credits'." Vanta charges a software fee and absorbs the inference cost.
What would a 4x markup on Claude actually cost?
Anthropic's pricing page, checked on 26 October 2026, lists these prices per million tokens (input / output):
- Claude Haiku 5.5: $0.10 / $0.50 for prompts up to 100,000 tokens
- Claude Haiku 4.5: $1 / $5
- Claude Sonnet 5 and Sonnet 5.5: $2 / $10
- Claude Sonnet 4.6 and 4.5: $3 / $15
- Claude Opus 5.5: $4 / $20
- Claude Opus 5: $5 / $25
- Claude Fable 5.1: $10 / $50
An app reselling Claude Sonnet 5 at four times list would charge $8 / $40. That is more than Anthropic charges for Opus 5.5 or Opus 5. A customer with a calculator would ask why they are paying more than Opus prices to reach Sonnet.
Why list price hides real AI margins
Anthropic's pricing includes three things that make the markup hard to read from either side.
- Prompt caching. Cache hits cost 0.05 times the base input price on Opus 5.5 and Sonnet 5.5, and 0.1 times on most other models. On Sonnet 5.5 that is $0.10 per million cached input tokens against a $2 list price.
- Batch API. Batched requests get 50 percent off both input and output.
- Tokenizer changes. Claude 4.7 and later models use a tokenizer that produces about 30 percent more tokens for the same text than Sonnet 4.6 and earlier, so token counts are not comparable across model generations.
An app could charge exactly Anthropic's list price, show customers a 1.0x multiple and still make a wide margin after caching and batching. An honest pass-through on a long-context workload could also look like a markup. Customers compare against list price, not the vendor's real cost. That supports Cacioppo's practical point: per-token billing invites a comparison the vendor cannot control.
What are agent harnesses, and why do they matter here?
In the full audio, Cacioppo briefly says "we're going to talk about harnesses" before returning to pricing. Anthropic's engineering post on Managed Agents defines a harness as "the loop that calls Claude and routes Claude's tool calls to the relevant infrastructure." How many tokens an application burns depends on its harness, so pricing and harness design are linked.
A note on the "Forex" transcription
Automatic transcripts render "4x" as "Forex." Podcast Alpha's summary of the episode published the quote as "You've just marked up a Claude token Forex." The video's own captions read "4x," which is correct.
Bottom line
Cacioppo's argument is that reselling tokens puts an AI application in a price comparison against the model it is built on, and the application loses. The 4x is a hypothetical complaint, not a market figure. At Anthropic's current prices, a 4x markup on Sonnet 5 would cost more than Opus list, while caching and batching mean list price says little about a vendor's real cost.
Sources
- Anthropic pricing - list prices, caching, batch discount and tokenizer note
- Invest Like the Best episode listing, Muck Rack - episode number and date
- Invest Like the Best, Apple Podcasts - feed listing and runtime
- Becoming a Category Winner, Colossus - publisher's episode page
- Fortune, 29 April 2026 - Vanta's ARR and customers
- Vanta on Sacra - growth rate and funding history
- Vanta (company), Wikipedia - founding and founders
- BigGo Finance, 20 October 2026 - Cacioppo's quote on token pricing
- Podcast Alpha, 23 October 2026 - episode summary with the "Forex" error
- Scuttlebutt, 20 October 2026 - episode summary
- Scaling Managed Agents with stable interfaces, Anthropic - definition of a harness
Published October 25, 2026. Based on a TikTok by @iltb_podcast. Full video notes (markdown)