Why everyone’s switching to AI credits
OpenAI and Salesforce aren’t the first to introduce credit-based pricing, but they’ll make it much easier for you to do it
My friends at Metronome just launched a 🔥 new report about what they’re calling The Monetization Operating Model. It unpacks the shift from charging for access to charging for value and why it’s not a simple pricing change. This shift is powerful, but chaotic — impacting sales, marketing, product, eng and finance.
I’ve been talking about what’s happening for a while. This report shows you exactly how to evolve with details on billing, experimentation, pricing logic, data and more. Check it out here.
Microsoft announced AI credits for Copilot in Microsoft 365 in January. Salesforce added a new flexible, credit-based model for their AI agent in May. Cursor shifted to credit-based pricing in June (and faced some real pushback from users). Not to be outdone, OpenAI recently replaced seat licenses with a pooled credit model for its Enterprise plans.
Credit-based models have been gaining popularity for at least the past year. I’ve seen the model adopted by the likes of Adobe, Apollo, Asana, Atlassian, bolt.new, Clay, HubSpot, Google, Lovable, monday, Replit, v0 and many, many others.
Most AI spend is probably now in the form of credits, which are also referred to as tokens, AI units, generative credits and flex credits. While none of the big players are the first to introduce it, they’ll make it much easier for the next company to sell AI credits.
The main roadblock to faster adoption has been getting buyers to accept it. Credit-based pricing can be hard to predict, hard to manage, and feel like a black-box. And nothing scares procurement more than runaway costs with no visibility.
But large incumbents shifting to this model validates that it’s a ‘standard’ way to buy AI apps. Salesforce and OpenAI are doing the hard work of educating the market, making it easier for startups to replicate it. And, presumably, these companies are making these moves after extensive research with customers – indicating that a lack of predictability isn’t an insurmountable roadblock.
Today I’m unpacking why credits are suddenly everywhere and how to get credit-based pricing right.
Why everyone’s switching to AI credits
Many folks predicted that AI costs would drop 10x every year. That would, in theory, allow folks to sacrifice margin today – heck, even take a negative gross margin – and then reap the rewards very quickly. In this world, distribution was the only thing that mattered.
Pricing could be optimized later. And most launched with simple all-you-can-eat subscriptions or traditional seat-based pricing.
This isn’t playing out. Older models are getting cheaper. But the latest models are still fairly expensive and “nobody wants yesterday’s newspaper”. OpenAI’s GPT-5 costs $10 per million output tokens – not much cheaper than where GPT-4o was priced in early 2024.
Meanwhile token consumption is exploding as folks trust AI with bigger, more complex tasks. We’re not just summarizing meeting notes. We’re asking AI for help with coding, reasoning, deep research and agentic tasks.
In my experience, 70-80% of token consumption comes from just 10% of users. These heavy users can become very unprofitable without the right guardrails. And AI gross margins are already challenged based on reporting from The Information, leaving little room for further erosion.
To summarize: AI isn’t necessarily getting cheaper, there’s a wider disparity in costs and capabilities, and there’s increasing potential for skyrocketing usage from power users.
Credit-based models are uniquely suited to accommodate this complexity. Customers see a pool of usage that feels relatively straightforward (“50 credits per month”). Vendors, meanwhile, can easily adapt credit pricing to monetize newer, higher value actions, navigate evolving LLM costs, and nudge customer behavior in a way that’s win-win.
Credit-based models also allow companies to include some AI usage “for free” for existing customers or in free plans, and they let customers decide exactly how they want to use those credits.
What credits actually mean in practice
While credit-based models are suddenly everywhere, each company defines a “credit” differently, which can make it extremely frustrating for folks (or newsletter writers 😉) to figure out what exactly they’re buying. Price per credit can range from a fraction of a penny (see: Salesforce) to upwards of $0.25 (see: Lovable)!
The default starting point is to make credits a cost-based metric, literally passing through AI costs directly to customers. It’s straightforward, relatively easy to calculate, and transparent. This approach lets customers automatically benefit from any decreasing LLM costs and gives them the power to optimize their usage/spend based on their own objectives.
Cursor credits are simply dollars of LLM API usage at the prevailing model API rates.
Clay credits are charged for data points where they have to pay a cost for access. A mobile phone number lookup, for example, might cost anywhere from 2 credits up to 25 credits depending on the data provider.
That being said, I’m generally not a fan of cost-based pricing outside of infrastructure software. Buyers frankly don’t care what your costs are; they care about what they get. They also don’t want to have to do the work of figuring out whether OpenAI or Claude or someone else has the best price-for-performance for their specific task. Isn’t that your job as the vendor?
A more value-aligned approach is to make credits an output-based metric, charging only for successful work products like a task execution, AI summary, image generation, or a conversation resolution. This approach is harder to quantify upfront, especially for multi-purpose apps, and pricing doesn’t always link directly with LLM costs. Today’s prevailing rate is charging ~$0.10 for simple task executions and up to ~$1 for complex, multi-step workflows.
Salesforce Agentforce Flex credits are connected to “business outcomes” like updating a customer record, automating a complex workflow, or resolving a case. This is a more flexible option in addition to their existing $2 per conversation model.
HubSpot Breeze credits can be used for customer agent interactions (100 credits) or simpler data enrichments (10 credits).
While output-based credits are still complicated, the communication to customers is much more straightforward. You can see that for yourself in Salesforce’s press release announcing the new Flex Credit pricing model (snippet below).
Buyers can more easily compare the cost-per-action to their existing approach and can estimate a business case accordingly. And they have peace-of-mind that pricing connects to tangible outputs rather than some arbitrary token calculation from a third party LLM.
How to make credit-based pricing work for you
I believe credit-based pricing is a bridge from flat-rates toward more value- and outcome-based pricing models. But they’re an important bridge for the present moment.
After dozens of conversations with folks exploring or implementing credit-based pricing, here’s some of what I’ve observed about how to do this well.
1. Monetize based on multiple axes, not just credits.
Credits feel like a commodity, particularly when they’re essentially a mechanism for passing-through third-party LLM costs. Commodity products face margin pressure and a race to the bottom on pricing (see: SMS messaging and compute). Multi-axis pricing allows you to offer a competitive credit price while still monetizing what makes your product unique and valuable.
Clay's hybrid pricing model, for instance, has multiple routes to expand customers while keeping pricing relatively simple: more features (subscription packages) and more usage (credits). All of Clay’s plans include unlimited seats.
2. Include a baseline amount of credits in all plans.
Ever get excited about a gift only to discover there were no batteries included? Just me?
I recommend including a base amount of credits across all plans to allow folks to form a habit around your product before getting hit with a credit overage bill. A simple rule of thumb is to ensure that at least 80% of customers have enough credits for their first month (remember that the top 10% of customers usually account for 70-80% of all credit usage).
3. Where possible, tie credits to outputs rather than costs.
In other words, don't automatically charge the customer when they want to try a feature. Charge when their usage is successful and (ideally) recurring. It's about shifting from a mindset of charging for access to charging for the work delivered.
At Replit, for example, credits are primarily drawn down for Agent checkpoints, which means the AI agent has taken an action and it has succeeded. The Agent automatically suggests a next step, which fuels ongoing usage.
4. Shift from monthly to annual usage limits.
Running out of usage in the first month feels like a bait-and-switch. So does signing up for an expensive plan with high limits only to use a fraction of that allowance.
An annual drawdown model gives the customer more time to consume their credits and to forecast what steady state looks like. If they run out of credits in month 9 or month 10, it's simply an early renewal – not a huge overage.
For customers with a monthly usage limit, there’s significant momentum around letting unused credits rollover. Lovable and bolt(dot)new, for instance, announced credit rollovers over the summer. I think that’s a good thing for the current moment.
Rollovers mean more lock-in. Always having credits in an account drives folks back to the product and makes it harder to churn to a different provider. And rollovers reduce barriers to usage, equating to less ‘credit hoarding’ behavior and increasing overall usage.
5. Help admins track and predict their usage.
Admins increasingly expect transparency around which actions drive up their bill. They want to be able to set billing thresholds at both the account and the user level. In a credit-based model, usage and spend needs to be visible within the product.
The more closely you can put usage in the context of solving a customer's problem and helping them see a tangible outcome, the more palatable it becomes. As you help customers predict their usage, you can communicate why more usage means a higher ROI for their business.
Final thoughts
Credit-based pricing isn’t necessarily new. As a consumer, I’ve seen credit models everywhere from games (👋 Fortnite) to fitness (👋 ClassPass) to food delivery (👋 Hungryroot). And as a 90s kid, credits are giving me big cellphone throwbacks (although I’m still waiting for someone to introduce free nights and weekends).
But, honestly, I’m torn about it. I can’t imagine the headache of managing dozens of AI products that all have credit allowances. Or running into unused credits with one vendor while facing punitive overages with another vendor. It’s no wonder that new startups like Cline are opting for a different approach, unbundling the app from the AI inferences with a “bring your own keys” model.
For now AI credits can be a lifeline — just not the endgame. They shift the mindset away from flat-rate pricing and unprofitable customers, and toward a future where customers pay for the work delivered by a combination of software and AI. The future of AI pricing is still being written.
Related reading: The monetization operating model [Metronome], The state of B2B monetization in 2025, A new framework for AI agent pricing, From selling access to selling work.
Interesting reading, I'll add it to the awesome-pricing repo!
Love the article Kyle