Tokens, credits, and hybrid are three ways AI products charge for usage, and they differ in who absorbs the swing when model costs change. Token-based passes raw model cost through, credit-based sells a prepaid balance that abstracts it, and hybrid pairs a subscription base with metered usage. This article covers how each works and when hybrid fits.
Why is the pricing model a margin decision?
For an AI product, the pricing model is a margin choice. AI-native products run gross margins of 50 to 60 percent, against 80 to 90 percent for traditional SaaS. Model inference alone accounts for roughly 23 percent of revenue at scaling stage, a share that does not shrink as you grow (Bessemer Venture Partners, February 2026). When the cost of serving a request is that large and that variable, the way you charge decides whether each new customer adds margin or quietly removes it.
Three pricing models dominate AI products today: token-based, credit-based, and hybrid pricing. They differ in one thing that matters more than format: who absorbs the swing when model costs move.
How does token-based pricing work?
Token-based pricing charges for each unit of model consumption, usually the input and output tokens a request uses, billed at or near the provider's published rate. It is the most transparent model: the customer pays for what the model actually processed, and the price maps almost directly to the underlying cost.
OpenAI's API is the canonical example of token-based pricing. Access is metered per million tokens, and cumulative spend unlocks higher rate-limit tiers, from $5 at Tier 1 up to $1,000 and a $200,000 monthly ceiling at Tier 5 (OpenAI API docs, as of April 2026). Anthropic prices its latest model at $5 per million input tokens and $25 per million output tokens (CNBC, April 2026).
The transparency has a cost, and the seller pays it. If you price your product as a markup on tokens, your margin moves every time a provider reprices output tokens or a new model draws more tokens per task. The risk is not theoretical. Anthropic's own $200-per-month Max plan needed weekly usage caps in mid-2025 after fewer than 5 percent of paid subscribers ran agentic coding workloads that were uneconomic to serve (TechCrunch, July 2025). Even a large flat fee could not contain token-level cost.
One honest counterweight: inference is getting cheaper. The cost of a fixed level of capability has fallen roughly 10x per year (a16z, November 2024). That softens token exposure over time, but it does not remove two problems. A switch to a more expensive model erases years of cost gains overnight, and you still pay the provider before your customer pays you.
How does credit-based pricing work?
Credit-based pricing sells the customer a prepaid balance of credits that they spend as they use the product, with the vendor setting how much actual usage each credit buys. The credit is an abstraction layer between the raw model cost and the price the customer sees.
That layer is the point. ElevenLabs charges 1 credit per character on its standard model and 0.5 credits on its faster, cheaper models, includes 10,000 free credits a month, and lets paid credits roll over for two months (ElevenLabs, as of April 2026). Salesforce sells Agentforce Flex Credits at $0.10 per action in packs starting at $500 (Salesforce, May 2025). In both cases the headline credit price can stay fixed while the vendor adjusts what a credit actually buys as model costs shift.
Kyle Poyar of Growth Unhinged describes credits as sitting in the middle of the spectrum between charging for access and charging for outcomes (Growth Unhinged, January 2026). That positioning is part of why credit-model companies in the top 500 SaaS and AI vendors grew 126 percent year over year (PricingSaaS via Growth Unhinged, January 2026).
The same abstraction creates a visibility risk. Because credits hide the raw cost, a vendor without per-customer cost data cannot see which buyers are consuming at a loss. Cursor's June 2025 move to a $20 base fee plus a $20 credit pool charged at API rates exposed that agentic coding tasks consumed far more than ordinary completions, and some users on the old flat plan had been structurally unprofitable.
Customers ran out of credits after a few prompts on new models, and the company issued an apology and refunds (Cursor; TechCrunch, July 2025). Cursor kept credits. The takeaway was that credits need published burn rates and a live balance the customer can see, or they feel like a meter running in the dark.
Credits also let you price uneven workloads without renaming plans, the way Replit moved from a flat $0.25 per checkpoint to effort-based pricing that ranges from $0.06 to several dollars depending on the work done (Replit, July 2025). If your product is really charging in a named unit rather than dollars, the mechanics of how to bill in custom units instead of dollars are worth a closer look.
How does hybrid pricing work?
Hybrid pricing combines a subscription base, for access and predictability, with metered usage charged above an included threshold. The base gives both sides a floor: predictable revenue for the vendor, a predictable bill for the customer. The usage layer tracks the part of cost that actually varies.
This is where the market is moving, and quickly. In a 240-company survey, hybrid pricing rose from 27 to 41 percent in a single year, the largest shift the survey had recorded, while flat-rate fell from 29 to 22 percent and seat-based from 21 to 15 percent (Growth Unhinged, June 2025). Among AI companies specifically, 58 percent already carry a subscription component, and hybrid adoption is projected to reach 48 percent in 2026 (ICONIQ Growth, January 2026).
Hybrid wins by elimination. Pure per-seat pricing left light users subsidizing power users, producing negative unit economics and churn risk at once, while pure token pass-through made margin hard to predict. The same ICONIQ survey quotes a VP of Product on the result: the subscription model was failing because heavy users ran negative margins while light users risked churn. Their plan was to bundle usage inside a subscription, which is exactly hybrid.
For the mechanics of putting one together, see our guide to designing hybrid pricing for AI products.
Enterprise buying reinforces the same endpoint. CIOs still prefer usage-based pricing over outcome-based, citing unpredictable costs and hard attribution as barriers to paying per outcome (a16z, May 2025). That leaves a subscription floor plus a usage layer as the pragmatic middle.
Tokens vs credits vs hybrid: a side-by-side comparison
The three models trade off the same five things differently. The right choice depends on how variable your costs are and how much margin visibility you have today.
| Model | Cost-to-price tracking | Margin protection | Revenue predictability | Customer experience | Best fit |
|---|---|---|---|---|---|
| Token-based | Direct, close to 1:1 with the provider rate | Low. The seller absorbs every cost swing | Low. Tracks raw usage | Transparent, but a bill is hard to forecast | Developer APIs whose buyers already think in tokens |
| Credit-based | Indirect. The vendor sets the conversion rate | Medium to high, but only if burn rates are tracked | Medium. Prepaid balances smooth cash flow | Simple unit, but opaque if burn rates are hidden | Consumer and prosumer products spanning several model tiers |
| Hybrid | Base is fixed; the usage layer tracks cost | High. The base covers fixed cost, usage covers variable | High. Subscription floor plus metered usage | Predictable base; pay more only when you use more | Mature products serving a mix of light and heavy users |
Read down the margin-protection column and the convergence on hybrid stops looking like fashion.
How should you choose between tokens, credits, and hybrid?
Start from your costs, not from a competitor's pricing page. Four questions narrow the choice quickly.
- How variable is your per-request cost across models? If you run several models with different economics, raw token pass-through exposes you to every swing. A credit abstraction or a hybrid usage layer gives you a place to absorb it.
- Do you, or your buyers, need budget predictability? If either side needs a forecastable number, a subscription base belongs in your model. That points to hybrid.
- Can you see per-customer margin today? In practice, that visibility is rarely in place when a team picks its pricing model. Credit-based pricing is only safe if you can see which customers consume at a loss. Without it you are guessing, whatever model you pick.
- How often will you reprice? Token pass-through forces public, visible changes every time. Credits let you adjust what a credit buys quietly. If you expect model costs to keep moving, that flexibility is worth something.
Whatever you pick, the same infrastructure sits underneath: something that authorizes and bills usage in real time, tracks a pre-funded balance, applies the right price per model, and shows the per-customer cost behind each charge. Building that yourself is the part founders consistently underestimate.
How Credyt handles this
If you are choosing between token, credit, and hybrid pricing for an AI product, Credyt runs all three on one platform, so the model is a configuration choice rather than a rebuild.
With Credyt you can:
- Run usage-based, prepaid credit, subscription, and hybrid models, and switch between them without re-architecting billing.
- Hold multiple assets in one customer balance (USD, tokens, custom units), with a different rate per model through dimensional pricing.
- Query the customer balance through Credyt's Wallet APIs before each action, then submit the usage event and let Credyt price and debit in a single atomic call.
- See per-customer cost and real-time gross margin per plan as you configure pricing.
- Give customers a branded billing portal with live balance, usage history, and self-service top-ups.
That covers the visibility gap behind credit pricing and the cost tracking a hybrid model needs. See how it works for real-time billing for AI products.
Related resources
- How do AI credit systems work? A closer look at credit grants, burn rates, and expiry.
- AI pricing models Definitions for the main ways AI products charge.