Credit-based pricing was supposed to simplify AI billing. It did the opposite.
Cursor replaced 500 fast requests with "$20 worth of API credits." Users burned through their allowance in a few prompts. Salesforce launched Agentforce at $2/conversation, pivoted to Flex Credits at $0.005/credit after backlash, and left customers reconciling three pricing models in twelve months. HubSpot priced Breeze at 100 credits per conversation, then dropped credits entirely for $0.50 per resolved conversation in April 2026.
The pattern is consistent. Credits arrive as simplification. They leave as apology letters.
Last Updated: April 2026
The Unit-of-Account Problem
Credits work when 1 credit equals 1 unit of consistent work. A gym visit. A car wash. A laundry cycle. The buyer knows what they get.
Credits break when 1 credit equals a variable amount of compute depending on which model runs underneath. A simple code completion on GPT-4o Mini costs a fraction of a cent. The same prompt routed to Claude 3.5 Sonnet costs ten times more. Both consume "1 credit" from the buyer's account.
The buyer thinks they purchased a currency. They purchased a blind bet on infrastructure decisions they cannot see.
79 of the top 500 SaaS companies now offer credit systems, up 126% year-over-year. The model is winning adoption. The question is whether anyone buying credits understands the exchange rate.
Three Layers of Opacity
The problem is not that credits exist. The problem is that the abstraction between action and cost has three separate failure points, and most products fail at all three.
Layer 1: Published rates nobody reads. Figma charges $0.03/credit. A design system generation costs 30+ credits. A text edit costs 1. The conversion table exists. It lives in a help center article. The customer discovers actual consumption through depletion, not documentation. Publishing rates in a place nobody checks is transparency on paper.
Layer 2: Model routing nobody controls. Vendors route requests to different foundation models behind the scenes. Credit consumption changes without the customer doing anything differently. The buyer chose an action. The vendor chose the cost. Zylo Research found 65% of IT leaders report surprise bills from AI tools. The common denominator: the customer's action stayed the same, but the infrastructure cost shifted.
Layer 3: Forecasting nobody can do. PYMNTS reported CFOs face "opaque, hard-to-forecast expenses" from AI tools, with invoices arriving as "dense ledgers of token counts, model tiers and throughput metrics." A buyer who cannot explain a $10,000 invoice to their CFO will churn regardless of value delivered. This is the same dynamic that makes enterprise buyers demand forecastability in every billing model.
The Evidence: Three Companies, Same Wall
Cursor: Overnight Depletion
On June 16, 2025, Cursor replaced its request-based Pro plan (500 fast requests/month) with "$20 worth of API credits" consumed at varying rates per model. The sticker price stayed the same. The effective value collapsed.
TechCrunch reported users running out of requests "after just a few prompts when using Anthropic's new Claude models." Community members reported going from roughly $100/month to $20-30/day in charges. No usage dashboard existed at launch. No overage consent was required. The meter ran blind.
CEO Michael Truell issued a public apology on July 4, 2025, and offered refunds.
The failure was three missing pieces: variable consumption rates with no dashboard, no projections, and automatic overage billing. Each is fixable. Together they created a trust crisis that pricing alone did not cause.
Salesforce Agentforce: Three Models in Twelve Months
Salesforce launched Agentforce at $2 per conversation. Oliv.ai calculated five agents handling 70 conversations daily would cost roughly $27,000/month.
By May 2025, Salesforce pivoted to Flex Credits: 100,000 credits for $500. Standard actions cost 20 credits ($0.10 each). Voice actions cost 30 credits ($0.15 each). Monetizely called the evolution "whiplash-inducing."
The math explains the pivot. A conversation averaging 10 standard actions costs $1.00 in Flex Credits versus $2.00 flat. A conversation averaging 30 actions costs $3.00 versus $2.00 flat. Salesforce was undercharging on simple conversations and overcharging on complex ones. Credits let them align price with compute.
The tradeoff: customers now model "standard actions" and "voice actions" in credit terms. One product officer described the experience as "a black box. Hard to predict, hard to explain to stakeholders."
HubSpot Breeze: Credits Dropped Entirely
HubSpot initially priced Breeze Customer Agent at 100 credits per conversation ($1.00 each). In April 2026, HubSpot shifted to outcome-based pricing: $0.50 per resolved conversation, $1.00 per qualified lead. The credit layer disappeared.
HubSpot concluded that credits added friction without adding clarity. A support conversation is either resolved or not. A lead is either qualified or not. Binary outcomes do not need an intermediate unit of account. The product that could simplify did.
When Credits Earn Their Complexity
Not every credit system is broken. The distinction is whether the abstraction simplifies a genuinely complex product or obscures a simple one.
Canva bundles 500 AI credits across all Magic Studio features on its Pro plan ($12.99/month). Background removal, text generation, image expansion, style transfer. Five distinct AI capabilities with different cost profiles, unified under one balance. The credit replaces five separate meters. That is simplification.
Lovable prices AI code generation at $20/month for 100 credits. One core action. One cost profile. The credit replaces nothing. It adds a conversion step between the user and the dollar amount they will spend.
| Scenario | Credits simplify | Credits obscure |
|---|---|---|
| Multi-product platform (5+ AI features) | Unified billing across different cost profiles | N/A |
| Single-action product (code completion, chat) | N/A | Conversion math on a simple transaction |
| Variable model routing | Absorbs backend complexity the customer cannot control | Hides cost spikes from model upgrades |
| Enterprise procurement | Predictable committed spend | Unpredictable consumption against commitment |
The test: does the product genuinely need a unit of account that is not dollars? A platform with multiple product types, different cost profiles per action, and cross-product bundling has a legitimate reason for credits. A product with one core action at a relatively consistent cost added a tax on every purchasing decision.
The Forecastability Gap
Vayu's analysis found 60% of customers cite unexpected pricing complexity as a barrier to scaling AI initiatives. Not cost. Complexity.
The distinction matters. A buyer who understands a $10,000/month bill and considers it fair value will renew. A buyer who receives a $10,000/month bill they cannot explain will churn. This is the same margin problem from a different angle: when the buyer cannot forecast, the seller cannot retain.
Five attributes separate forecastable credit systems from opaque ones:
| Attribute | Forecastable | Opaque |
|---|---|---|
| Conversion rates | Published per action type, linked from the pricing page | Discovered through depletion |
| Usage visibility | Real-time dashboard with projected spend | End-of-month invoice |
| Projections | "At current rate, you'll use $X by month-end" | No projection available |
| Overage handling | Service pauses or requires approval | Automatic billing without consent |
| Rollover policy | Documented on the pricing page | Buried in terms of service |
The FTC's $2.5 billion Amazon Prime settlement established that manipulative billing interfaces create legal exposure. An FTC study found 76% of subscription services use at least one dark pattern. Hidden conversion rates map to "drip pricing." Automatic overage billing without consent maps to "forced continuity." The regulatory categories already exist. A single high-profile case could set credit-specific precedent.
The Quiet Part
If credits hide the cost-to-serve from the buyer, they also hide it from the seller.
A vendor selling "500 credits" controls the exchange rate. They can adjust what a credit buys without changing the sticker price. That flexibility runs in both directions. When the underlying model gets cheaper, the vendor pockets the difference. When the underlying model gets more expensive, the vendor absorbs the loss. Neither adjustment is visible to the buyer. Neither is visible to the seller's own product team unless they have cost attribution by customer and model.
This is the real problem. Credits do not just obscure the buyer's cost. They obscure the seller's margins. A credit system without per-customer cost visibility means the vendor is pricing on averages. Some customers are wildly profitable. Some are underwater. The credit abstraction prevents both sides from seeing which is which.
Metronome's 2025 field report found that "finance likes it. Customers don't know what a credit does." The same report noted customers avoiding free credits because they feared unpredictable costs. Credits designed to drive adoption were suppressing it. The vendor could not diagnose why without breaking the abstraction they built.
The unit economics challenge is hard enough when costs are transparent. Adding a conversion layer that neither side fully understands makes every subsequent pricing decision worse: margin calculations, cost dashboards, migration planning, and multi-model routing all depend on knowing what a unit of work actually costs.
The Credit System That Works
The credit bet pays off under five conditions. Miss one, and opacity fills the gap.
- The product has multiple action types with materially different cost profiles.
- A clear conversion table fits on one page and is linked from the pricing page.
- Real-time usage visibility exists, with projected spend, not just a balance counter.
- Overage policy requires explicit customer consent before charging.
- A finance team can forecast next month's bill from this month's usage data.
HubSpot looked at that list, counted the "no" answers, and dropped credits. Cursor looked at the backlash, added a dashboard, and published model-specific rates. Salesforce is still iterating.
The companies that publish their exchange rates, build real-time dashboards, and pause at limits are making the credit bet honestly. The rest are building on a trust deficit with a timer attached.
Neither the buyer nor the seller can make informed pricing decisions when the unit of account obscures the unit of cost. Transparency is not a feature of a credit system. It is the credit system. Without it, credits are just opacity marketed as simplicity.
If you are building an AI product and need cost attribution by customer, model, and feature, Bear Lumen provides the visibility layer that makes credit pricing honest. See how unit economics work.