Product
Industries
Roles
Research
Resources

Research · Product Releases

A prepaid wallet for agent spend

Oxagen 5.5 prices each model call by size before it runs, pays from a monthly allowance and then a prepaid wallet, and holds new runs when the wallet is dry while started runs keep their rules.

A stella run works a ticket overnight. Each step is a model call, and each call buys tokens from a model provider. By morning the run has made hundreds of calls. Finance asks what it cost and who agreed to pay for it. The honest answer used to be that nobody knew until the meter caught up.

#Why token billing arrives after the spend

Model providers price by the token. Input and output tokens have separate rates, and the rates differ from one model to the next.1 The token count is known only once the response ends. So a bill built on tokens is written after the money is already spent.

For one chat, that gap is small. For an agent that loops through a task, the gap is where a budget breaks. Kapoor and colleagues argue that agent evaluations should report cost next to accuracy, because agents that score alike can differ in cost by a wide margin.2 The FinOps Foundation puts it in operating terms: most cost practice carries over to AI, and tokens, volatile pricing, and new stakeholders are what change.3

Oxagen had the same gap in its own billing. A model call on Oxagen's key was priced after it ran. So a call could spend credits the organization did not hold. A second defect cut the other way. When the wallet ran dry, a running agent's next tool call was refused, so an agent could lose its rules in the middle of a task. Version 5.5 fixes both.

#How a model call is priced

A model call on Oxagen's key, such as a stella step or an Oxagen AI feature, now costs a fixed number of credits set by its size. One credit is $0.01.4

Credits per model call on Oxagen's key, by size

Call sizeCredits
Light (Claude Haiku 5.5)1
Standard (Claude Sonnet 5.5)15
Deep (Claude Opus 5.5)30
The published call prices. One credit is $0.01. Source: Oxagen billing docs.

The size comes from the model family, or from the tier the model serves. An embedding or a batch call is Light. A model outside both is priced as Standard. Because the price depends only on size, Oxagen knows it before the call and never corrects it afterward.

The charge runs before Oxagen contacts the provider. If the wallet cannot pay, the call stops before a token is bought. If the call fails before the model returns anything, the credits come back.

How a model call on Oxagen's key is paid

  1. PriceLook up the call's size and its price in credits.
  2. MandateCheck the agent's daily credit budget. Over it, the call asks a person or is denied, as the rule says.
  3. AllowanceIf the month's credits cover the price, they pay.
  4. WalletIf not, the prepaid wallet pays, with no room below zero.
  5. RefuseIf neither can pay, the call is refused before it runs.
Illustrative. The order of checks before one model call.

The second step matters to the person who owns an agent's budget. An agent's mandate can carry a daily credit budget. A call that would pass it either waits for a person to raise the budget or is denied, and the mandate's rule says which.

#How usage is tracked

Seats are paid at the start of the term. Each operator holds one seat, and each seat adds a monthly allowance that the workspace pools.4

PlanPrice per operatorEach seat adds a month
Team$119 a month billed monthly, or $99 a month billed yearly ($1,188)50,000 agent actions and 1,000 credits
Business$149 a month billed yearly ($1,788), from 10 operators100,000 agent actions and 2,500 credits

An agent action is one tool call, prompt, response, or approval that Oxagen records for an agent. The allowance pays first. It resets on the first of each month and does not roll over.

Past the allowance, every 50 agent actions cost 1 credit, which comes to $2 per 10,000. Claude Code, Codex, and Cursor run on your own model account, so your provider bills you for their tokens. Oxagen counts their agent actions only.

Agent actions get a margin that model calls do not. Recording an action costs Oxagen almost nothing, so the overage may take the wallet 500 credits, or $5, below zero. The next top-up pays that back first. A model call gets no such room, because Oxagen pays the provider for it.

Every change to the balance is a row in the wallet ledger: a grant, a purchase, a usage debit, an expiry, a refund, or an adjustment. Rows are never edited or deleted, and a mistake is fixed with a new row. The balance is the sum of the rows, so a question about a balance is answered by reading them. Prepaid credits for usage are a common billing pattern. Stripe's billing docs, for one, describe a credit grant as a set of prepaid credits given to a customer.5

#What happens when the wallet is empty

An empty wallet stops new spend. It does not take governance away from a run that has already started. Oxagen holds work back at the start instead.

A new run waits in two cases. The first is a wallet 500 credits below zero. The second is a seat payment that failed. Runs already going finish, still under their rules, and every agent action they take is recorded.

The rule bends for runs on your own model account, because those cost Oxagen nothing to record. Inside the month's allowance, such a run starts even when the wallet is empty. After a failed seat payment, it keeps starting for 7 days, then waits. A run on Oxagen's key waits from the first failure.

If auto top-up is on and a card is saved, the check before a new run tries one top-up before it holds the run. If the check cannot read the wallet, it holds a run on Oxagen's key and admits a run on your own key. A guess that admits a paid run would spend money nobody paid.

#How allowance alerts work

A daily billing sweep reads each month's allowance. When a workspace reaches 50, 80, or 100 percent of its agent actions or its credits, the billing owners get a notice in the app and an email. A month that jumps from 40 to 100 percent in a day sends one notice for 100, not three. A failed email is retried on the next sweep, and the in-app notice stays. The billing docs list the other alerts, such as a wallet that falls below the auto top-up threshold.4

#What this release does not cover

A model call on Oxagen's key never runs on credit, even inside a started run. If the wallet empties mid-run, the next stella step is refused with "Wallet empty. Add credits or turn on auto top-up to continue." The run's tool calls are still recorded and still governed. But a stella run makes no progress until someone adds credits. Auto top-up closes that gap.

Pricing by size trades precision for a price you can know in advance. A short Deep call costs the same 30 credits as a long one. A team that sends many tiny prompts to a Deep model pays more than the token count alone would suggest.

The alerts run once a day. A workspace that burns through its allowance in an afternoon hears about it at the next sweep, not at the moment it crosses the line. The mandate's daily credit budget is the control that acts in the moment.

These controls apply to calls and actions routed through Oxagen. Spend on your own model account is billed by your provider, and Oxagen does not forward that bill.

The other changes in this version are in the Oxagen 5.5.0 release post.

#References

  1. Anthropic (2026). Pricing. Claude documentation. https://docs.anthropic.com/en/docs/about-claude/pricing ↩

  2. Kapoor, S., Stroebl, B., Siegel, Z. S., Nadgir, N., & Narayanan, A. (2024). AI Agents That Matter. arXiv. https://arxiv.org/abs/2407.01502 ↩

  3. FinOps Foundation (2026). FinOps for AI Overview. FinOps Foundation working group paper. https://www.finops.org/wg/finops-for-ai-overview/ ↩

  4. Oxagen (2026). Billing and usage. Oxagen documentation. https://docs.oxagen.sh/docs/billing ↩ ↩2 ↩3

  5. Stripe (2026). Billing credits. Stripe documentation. https://docs.stripe.com/billing/subscriptions/usage-based/billing-credits ↩