Token Economy: How AI Markup Pricing Works at DMHub

When we built DMHub's AI features, we had a choice: flat-rate AI pricing (like "$49/mo for unlimited AI") or usage-based pricing tied to actual compute costs.

We chose usage-based, and then we made it transparent. This post explains exactly how it works.

Why Flat-Rate AI Pricing Is a Lie

"Unlimited AI" in SaaS usually means one of two things:

1. The AI is a rules-based decision tree with a language model slapped on top. Cheap to run, limited in capability.
2. The AI is real (GPT-4o, Claude), and "unlimited" hides a fair-use throttle in the ToS. Heavy users get throttled or churned.

We didn't want to hide our costs behind a flat rate. We'd end up either over-charging light users or losing money on heavy users. Neither is sustainable.

What Actually Costs Money

When you use an AI feature in DMHub, we make an API call to either OpenAI or Anthropic. They charge us per token.

Approximate 2026 pricing (before DMHub markup):

| Model | Input (per 1M tokens) | Output (per 1M tokens) | |-------|----------------------|----------------------| | Claude 3.5 Haiku | $0.80 | $4.00 | | Claude 3.5 Sonnet | $3.00 | $15.00 | | GPT-4o | $2.50 | $10.00 | | GPT-4o mini | $0.15 | $0.60 |

A typical customer service AI response — reading 3 knowledge base articles, understanding the question, and writing a 150-word reply — uses roughly:

``` Input: ~4,000 tokens (system prompt + KB articles + conversation history) Output: ~200 tokens (the reply)

Cost at Claude Haiku: (4,000 × $0.80 + 200 × $4.00) / 1,000,000 = ($3.20 + $0.80) / 1,000,000 = ~$0.000004 per response (less than half a cent) ```

For a simple FAQ answer: $0.000004. For a complex multi-tool response (check inventory, create order, send receipt): up to $0.002.

The Credit System

DMHub converts token costs into credits. One credit = $0.01. We round up to the nearest credit to keep the math simple.

Credit rates by feature:

| Feature | Credits | Reasoning | |---------|---------|-----------| | AI agent response (simple) | 1 credit | ~$0.004 cost, 1 credit = $0.01 — light markup | | AI agent response (complex, multi-tool) | 3 credits | ~$0.02 cost at Sonnet rates | | AI Wizard action | 2 credits | Medium Sonnet call | | Voice note transcription | 1 credit | Whisper API, flat cheap | | AI-generated message draft | 1 credit | Short Haiku call | | Sentiment analysis | 1 credit | Haiku | | Auto-summarize conversation | 2 credits | Longer context Haiku |

Why credits instead of dollars? Three reasons:

1. Micro-billing (billing for $0.004 per API call) creates payment processing overhead that exceeds the transaction value.
2. Credits give users a clear budget to manage. "I have 500 credits" is easier to reason about than "I have $4.97 remaining."
3. Credit packs let us offer volume discounts cleanly.

The Markup

We apply a 2.5x markup on raw API costs. Here's why.

Raw cost of a typical AI agent response: ~$0.004 Credit charged: 1 credit = $0.01 Markup: 2.5x

The markup covers:

1. Infrastructure overhead — Load balancers, retry logic, queue management, rate limit handling, failover between providers. This costs roughly 30-40% of raw API cost.
2. Provider diversification — We route across OpenAI and Anthropic. When one has an outage, we fail over to the other. Operating two accounts has overhead.
3. Caching — We cache identical knowledge base lookups to reduce token usage. You pay for the cache infrastructure, but you benefit from lower token counts.
4. Margin — We need to be a sustainable business. 2.5x markup with infrastructure costs is a ~40% gross margin. SaaS businesses typically need 70%+, but AI cost structures are different.

We publish the markup because hidden markup creates distrust. You know what you're paying for.

Prompt Caching

One of the most impactful cost optimizations is prompt caching — a feature from Anthropic (extended cache) and OpenAI (cached prefixes).

When the same system prompt and knowledge base context appear in repeated calls, the provider caches the computed KV states. Repeat calls cost 10-25% of the original input token price.

DMHub implements this automatically:

• Knowledge base articles are pre-indexed into a cached prefix
• System prompts are structured to maximize cache hit rates
• Cache hit rates on typical DMHub deployments: 60-80%

Without caching, a 4,000-token input call at Claude Haiku costs $0.0032. With 70% cache hit, it costs: 70% × $0.0008 + 30% × $0.0032 = $0.00096. 70% savings on input tokens.

We pass these savings through in the credit rate. Simple AI agent responses cost 1 credit (not 3) partly because caching makes them cheap.

Free Credit Allocation

Every DMHub plan includes a monthly credit allocation:

| Plan | Monthly Credits | USD Value | |------|----------------|-----------| | Free | 100 | $1.00 | | Growth | 500 | $5.00 | | Pro | 2,000 | $20.00 | | Business | 5,000 | $50.00 |

Credits reset monthly. Unused credits don't roll over (to keep accounting simple). You can purchase additional credit packs at volume discounts.

Is $1/month of AI credits enough? For the Free plan, yes — if you're just testing. 100 credits = 100 simple AI responses or 33 complex multi-tool responses. That covers a small business handling 10-15 AI-assisted conversations per day.

Auto-Recharge

Running out of credits mid-day means your AI agent goes offline. Auto-recharge prevents that.

Set a threshold (e.g., 50 credits remaining) and a top-up amount (e.g., 500 credits = $5). When you hit the threshold, we charge your card and restore the credits automatically.

Auto-recharge is optional. You can also manage credits manually or set a hard cap.

Why This Matters

Most AI SaaS products hide their token costs in the plan price and hope you don't do the math. We'd rather show you the math.

If you use DMHub's AI features heavily and think you're paying too much, the transparency cuts both ways — you can calculate whether the feature is delivering ROI. If it is (and for most businesses, AI agent automation delivers a 3-5x return on credit cost in staff time saved), the pricing is defensible. If it isn't, you'll know exactly why.

---

Questions about credit pricing or AI feature costs? Talk to our team — we'll show you the actual economics for your use case.