How Atroxabot credits work

Atroxabot credits are the one currency on your account. Your plan adds credits every month, top-ups add more when you need them, and every agent you run draws from that same balance to pay for the work its model does. If agents are new to you, start with what an AI agent is.

That work is measured in tokens, the small pieces of text a model reads and writes. Each time an agent handles a message, it reads its instructions, the conversation so far and anything it looked up, then writes its reply. Credits are charged for what that request actually used, so a one-line answer costs less than a long report, and a light model costs less than a frontier one.

• Input: what the agent reads, such as its instructions, the conversation and anything its skills or connectors fetched.
• Output: what the agent writes, including the reasoning it does before it answers.
• Cache: context the agent reuses from an earlier message, which is billed at a fraction of fresh input.

Because every message differs, Atroxabot quotes a reference message: an agent reading about 8,000 tokens of context, of which 6,000 come from cache because its instructions and history repeat, and writing a reply of about 800 tokens. That is a realistic turn for an agent in an ongoing conversation.

The gap between models is large. One Opus message costs about as much as 38 Haiku messages, which is why the model is the biggest decision behind what an agent costs to run. Choosing a model for your agent covers what each model is best at.

Each plan adds a fixed number of credits every month and decides which models your agents may run.

Free is for trying an agent on Claude Haiku 5.5. Pro suits agents for your day-to-day work and unlocks Claude Sonnet 5.5 and Claude Opus 5.5. Ultra is for running many agents, or heavy models, all month. Compare the plans side by side.

Every agent draws from your one balance, but only up to the share you allow it. You set that share per agent, and the agent cannot spend past it, however busy it gets.

An example on Pro, with 250,000 credits a month: a morning briefing agent on Haiku with a tenth of the balance, 25,000 credits, has room for roughly 757 typical messages. The same share on Sonnet covers about 37. Seeing those numbers side by side is often what settles the choice of model.

• No surprises: an agent that loops or gets more requests than expected stops at its budget instead of draining the balance.
• Room for everyone: a busy agent never uses up the credits your other agents depend on.
• Clear costs: because each agent has its own share, you can see what every job costs you.

• Use the lightest model that does the job: for frequent, light work, Claude Haiku 5.5 at about 33 credits a message is hard to beat. Save the larger models for agents whose work needs them.
• Keep instructions stable: every model in the lineup supports prompt caching. When an agent's instructions and history repeat from one message to the next, they are read from cache. On Sonnet, the reference message costs about 660 credits with its cache and 1,200 without, so rewriting instructions constantly costs real credits.
• Keep context lean: everything an agent reads is billed on every message, so long histories and large documents add up. On Haiku, a request whose prompt passes 100K tokens is billed at a higher rate for the whole request.
• Choose effort on purpose: a higher effort setting lets a model reason for longer before it answers, and that reasoning is billed as output.
• Give every agent a budget: it caps the cost of a mistake and shows you which jobs are worth what they cost.