Atlassian recently began bundling its plans with a set number of credits for AI features while charging users who exceed the limit. This is a part of a bigger trend, where Software as a Service (SaaS) companies that are not AI labs themselves are tying pricing to metered AI usage.
The spread of such strategies and the novel business models they create isn’t only a side effect of AI technology adoption. It functions as an informal experiment for the whole SaaS pricing model, causing businesses to rethink how we charge for software.
The Shift to Usage-Based Pricing
Usage-based pricing (also called consumption-based pricing) is an umbrella term for pricing models in which customers are charged based on how much of the product they use. One of the first major implementations is associated with urban utilities like water and electricity, billed by the unit consumed.
It’s difficult to find such a clear unit of measurement for SaaS products, but tokens, roughly three-quarters of a word each, are becoming a reference unit even when reselling them under other names. Companies are using different forms of usage pricing, but the standard of tokens in AI drives the recent push.
A wave of established SaaS companies have adopted some form of usage-based pricing on top of existing subscriptions. Almost none go pure usage-based, layering a variable charge onto an existing subscription, and that layer takes different forms:
- Tokens are the raw billing unit AI labs use. GitHub Copilot now bills this way. Plans include an allotment of GitHub AI Credits, but those credits draw down at published per-token rates for each model, covering input, output, and cached tokens. The unit is called a credit; the meter is the token.
- Credits are a vendor-defined unit built on top of raw token costs, letting companies bundle AI usage into their own pricing language. Companies like Figma, HubSpot, and Airtable are using some form of these abstractions.
- Outcomes, such as resolutions, conversations, or tasks, mean that the product is billed by results, not inputs. For example, Zendesk charges per resolution and Atlassian's Virtual Service Agent bills per conversation past its credit limit.
- Native usage metrics, like API calls, compute time, GB used, and similar,**** are also worth mentioning. Many companies still rely on them even after implementing AI features.
GitHub is a convenient example for showing why the subscription model needs an addition. Under the old model, their users consumed much more token value than their subscriptions included, and GitHub had to cover the difference. The same risks might arise for other SaaS products implementing AI features.
Usage-based pricing isn't new. It has been the infrastructure default for over a decade: AWS, Twilio, Snowflake, and Cloudflare all bill by consumption, and so does every proxy provider, including IPRoyal. But today, metered billing is moving up the stack into application-layer SaaS: CRM, collaboration, service management, design tools.
Research shows that while most SaaS companies still use subscriptions, over 40% of them have already implemented usage-based or hybrid pricing models. Gartner forecasts expect consumption-based pricing numbers to grow to 70% by 2027 as companies implement more hybrid models, not forego subscriptions.
Part of this is a downstream effect where AI companies are charging SaaS vendors per token, so they need to recoup the costs in a usage-based form to make it financially viable. Yet, the change runs even deeper, as a simple infrastructure shift is not all that is happening.
SaaS Pricing Experiments
Everyone in the proxy industry priced for consumption from the start. Proxy bandwidth is billed by the gigabyte, which means we've spent years on the problems application-layer SaaS is only now discovering.
The lesson we keep relearning is that customers rarely leave over the price itself, but more often over price variance. Token-based pricing has unpredictability that proxies don’t – you always pay the same for a gigabyte, while LLM pricing depends on how many times an agent decides to reread instructions and skills, and even on the model itself.
The way SaaS companies approach usage-based pricing is still largely experimental and cuts in different directions. Some companies are exposing their pricing, keeping it closer to AI tokens, while others are abstracting AI token usage behind credits or outcomes. Various models are implemented to see which one will survive. Ours has always been to create built-in guardrails, such as caps, alerts, prepaid balances – all done so customers don’t spend more than they expect
Billing by usage, instead of account subscription, requires tracking software usage more closely. The billing infrastructure inherited from the pre-AI era wasn't built to track these new, AI-native metrics, like tokens or agent actions, at all.
Platforms, such as Metronome, have already been developed to automate usage-based pricing for software. Every action must be tracked in real time, then processed by applying rating logic and producing billable output.
The problem still persists, since subscription-based platforms were never designed to handle so much variability. Yet usage-based pricing still has incentive because it could open new markets.
Various services are emerging as Model Context Protocol (MCP) servers. MCP servers allow any SaaS to be transferred into a service that can be used by AI agents without requiring custom integrations.
Charging for subscriptions might fall out of fashion if the MCP becomes a standard, as one agent can use up far more resources than a human typically would. Usage-based pricing is a viable solution, but it requires a metering infrastructure for the company to charge users effectively.
Many SaaS companies started to build such infrastructure for their AI features at first, but it’ll likely touch all aspects of products eventually. One reason is that usage measurements also provide valuable data that lets the company optimize their product’s pricing even better.
Every company moving to usage-based pricing is testing how much unpredictability its customers will tolerate. The enterprise cautionary tales are accumulating fast: Uber reportedly exhausted its entire 2026 AI budget by April, and Microsoft scaled back internal Claude Code licenses after per-engineer costs reached $500 to $2,000 a month.
Vendors were quick to implement spend caps, usage alerts, and other safeguards. That’s where marketing experiments begin, as different customers need distinct pricing strategies and guardrails. Each variation is effectively a pricing test run on real customers that can change how SaaS products grow.
Usage-based pricing also carries an appealing promise that the investments, just like basic commodities, will pay off for the users. There is a strong case to be made that even if AI features fail, the results of such pricing experiments and their infrastructure basis will survive.
The Dotcom Precedent
The push for building new infrastructure and pricing models the AI boom created isn’t new. Telecom companies did the same thing in the late 90s, wiring countries for the internet that didn’t yet exist. The number of new websites grew exponentially, so the investments were seen as a safe bet.
When the dotcom bubble burst in the 2000s, the optic fiber and other internet infrastructure built during the bubble stayed unused for a time, but the cables and servers never disappeared. As broadband, streaming, cloud computing, and other online services scaled, it functioned as the cheap backbone of such innovation.
A company like Amazon would not be possible without the foundational layers that were built during the dotcom bubble. Using it, Amazon built its own internal distributed computing system to handle retail operations that became Amazon Web Services, now a major player in the cloud industry.
More traditional SaaS companies that outlasted the dotcom bubble paved the way for the whole business model. Salesforce, founded in 1999, is the clearest case. It survived the crash by making subscriptions look cheap next to six-figure on-premise licenses at the moment enterprise budgets collapsed – their pricing model won because a downturn made it the rational purchase.
The current shifts in infrastructure development and pricing models pushed by leading AI labs may be transformative in a similar manner. If that is the case, there will be new winners even if the AI bubble bursts.
Critics point out extreme overvaluation of AI companies, circular funding, unclear return on investment, physical infrastructure constraints, and other issues as signs of a bubble. It’s hard not to concede that there is some bubble risk. If funding tightens or a cheaper model undercuts today's leaders, the correction could be sharp.
There’s still bubble risk as it’s almost impossible to think otherwise. If funding tightens or a cheaper model undercuts leaders, the correction could be sharp. But our dotcom example cuts both ways – the crash was catastrophic, but its ruins built a new era.
Conclusion
The shift to usage-based pricing might be an early sign of SaaS’s next pricing model. Just as the dotcom era normalized subscriptions for software, the AI boom might bring usage-based pricing. Bubble or not, new business models are already being tested, and someone has to lead them.
Comments
Loading comments…