Why Your AI Costs Should Be Yours (And Visible)

Here's a fun exercise. Go to any AI app builder's pricing page. Find the word "credits" or "tokens." Now try to figure out what one credit actually buys you.

You can't.

That's by design. Credits are an abstraction layer between you and the actual cost of the AI model underneath. And that abstraction exists for one reason: margin.

The Credit Shell Game

Let's reverse-engineer what's happening. Most AI app builders use Claude (from Anthropic) or GPT-4 (from OpenAI) under the hood. The wholesale cost of these models is public knowledge:

- Claude Sonnet 4: ~$3 per million input tokens, ~$15 per million output tokens

- GPT-4o: ~$2.50 per million input tokens, ~$10 per million output tokens

A typical "generate a React component" operation uses about 2,000-4,000 tokens. That's roughly $0.01-0.03 at wholesale prices. A penny or three.

Now look at what these tools charge:

- Lovable: One "credit" per generation. 150 credits for $25/month. That's ~$0.17 per generation.

- Bolt.new: Token-based, but with overhead that makes each effective operation cost $0.05-0.15.

- v0: ~30-60 generations for $20/month. That's $0.33-0.67 per generation.

The markup ranges from 5x to 20x over wholesale API costs.

Is that unreasonable? Not necessarily. These platforms provide value beyond raw API calls — UI, orchestration, hosting. But the credit system hides the markup. You can't see it. You can't compare it. You just know your credits are gone.

What Happens During Debugging

The credit system's real damage shows up during debugging. When the AI gets something wrong and needs to retry, each attempt costs credits. But the AI isn't doing fundamentally different work — it's making the same type of API call. The wholesale cost of a retry is the same penny or three.

But to you, it's another credit gone. A $0.17 charge for something that cost the platform $0.02.

Multiply this across a debugging session where the AI tries 10-15 approaches, and you're paying $2-3 for what cost the platform $0.20-0.30.

When users report spending $200-1,000 on debugging spirals, a significant chunk of that is the credit markup — not the actual AI cost.

The BYOK Difference

Chorus does something fundamentally different. We're a platform that facilitates building — we provide the orchestrator, the agent team, the feedback loop, and the deployment pipeline. But the AI itself? That's powered by your Anthropic API key.

You plug in your own key. Our platform uses it to run your AI agents. You pay Anthropic directly at their published, wholesale rates. Chorus never touches, marks up, or resells your token costs. We're the facilitator, not the middleman.

When you build an app and it costs $3.87, you can go to your Anthropic dashboard and see exactly how those dollars were spent. Which messages cost what. Which operations were expensive. Which were cheap.

This transparency changes your relationship with the tool. You're not anxiously watching a credit counter tick down. You're seeing real costs that make intuitive sense. "Oh, that big feature design conversation cost $0.40. The code generation was $1.20. The debugging session was $0.80. Total: $2.40."

The Incentive Alignment Problem

Here's the part that should bother you most about credit-based pricing: the platform profits when the AI wastes your money.

Think about it. When Lovable's AI gets stuck in a debugging loop and burns 20 credits instead of 2, Lovable makes 10x the revenue from that session. They're charging you $3.40 for ~$0.20 worth of API calls.

Their financial incentive is not to fix the debugging problem. Every failed retry is revenue.

This is why the BYOK model matters beyond transparency. Since Chorus is a platform that facilitates building — not a middleman reselling tokens — we have zero financial incentive for the AI to use more of your tokens. In fact, we're motivated to make the orchestrator more efficient — because happier users stay longer, and because we actually care about building something good.

Efficiency isn't just a nice-to-have for us. It's what happens when you separate the platform business from the AI cost.

What About Chorus's Revenue?

Fair question. If we don't touch AI costs, how do we make money?

The Hobbyist tier is free. The Pro tier is $49/month. That subscription pays for the platform — the multi-agent orchestrator, the feedback loop, the deployment pipeline, the quality gates, the AI team coordination, project management, and all the infrastructure that makes BYOK actually work.

You pay Chorus for the platform. You pay Anthropic for the AI. Two clean, separate costs. No bundling. No cross-subsidization. No hidden margins.

We make money when you choose Pro. Not when your agents use more tokens. Not when they get stuck debugging. When you decide the platform itself is worth paying for. That's a business model we can be proud of.

The Transparency Test

Here's a simple test for any AI tool you're considering: can you calculate the actual AI cost of your last operation?

If the answer is "no, I just see credits disappearing," that tool is hiding something. Maybe the markup is reasonable. Maybe it's not. You literally cannot tell.

If the answer is "yes, I can see every API call and its cost in my provider dashboard," you're dealing with a tool that trusts you enough to show you the truth.

We think trust is a better foundation for a product relationship than obfuscation. And we think the market will eventually agree.