Every agent you create runs on the model you choose for it and spends credits from your balance as it works. Here is what each model handles, what it costs, and which plans include it.
AnthropicFast
Fast and light, for agents that answer often.
Specs
Credits per 1K tokens
≈ 33 credits for a typical message
AnthropicBalanced
The balanced choice for most agents.
Specs
Credits per 1K tokens
≈ 660 credits for a typical message
AnthropicFrontier
The most capable Claude, for the hardest work.
Specs
Credits per 1K tokens
≈ 1,260 credits for a typical message
A credit is the unit your agents spend. Each request costs credits for the tokens the model reads and writes, at that model's rates. Tokens served from the cache, such as an agent's instructions and the conversation so far, cost a fraction of fresh input.
Pro includes 250,000 credits a month and Ultra 1,500,000. Free comes with 10,000 a month to try an agent on Claude Haiku 5.5.
All your agents spend from the same balance, each up to the share you allow it, and you can buy more credits in blocks of 100,000.
A typical message is 8,000 tokens in, 6,000 of them read from the cache, and 800 tokens out.
| Claude Haiku 5.5 | Claude Sonnet 5.5 | Claude Opus 5.5 | |
|---|---|---|---|
| Specs | |||
| Context window | 1M tokens | 1M tokens | 1M tokens |
| Max output | 128K tokens | 128K tokens | 128K tokens |
| Inputs | Text and images | Text and images | Text and images |
| Effort levels | Low to max, medium by default | Low to max, high by default | Low to max, medium by default |
| Plans |
|
|
|
| Credits per 1K tokens | |||
| Input | 5 | 100 | 200 |
| Output | 25 | 500 | 1,000 |
| Cached input | 0.5 | 10 | 10 |
| Typical message | ≈ 33 credits | ≈ 660 credits | ≈ 1,260 credits |
| Capabilities | |||
| Tool use | Yes | Yes | Yes |
| Structured output | Yes | Yes | Yes |
| Vision | Yes | Yes | Yes |
| Prompt caching | Yes | Yes | Yes |
Match the model to the work. Claude Haiku 5.5: fast and light, for agents that answer often. Claude Sonnet 5.5: the balanced choice for most agents. Claude Opus 5.5: the most capable Claude, for the hardest work. Each agent has its own model, so one account can run all of them.
Credits are how Atroxabot measures what your agents use. Every request spends credits for the tokens the model reads and writes, at that model's rates. Monthly allowances: Free 10,000 credits, Pro 250,000 credits, and Ultra 1,500,000 credits. You can buy more in blocks of 100,000.
It is the reference our estimates use: 8,000 tokens in, 6,000 of them read from the cache, and 800 tokens out. That is roughly one agent turn whose instructions and earlier conversation come from the cache, with a reply of several paragraphs. Real messages vary, so treat the figures as estimates.
Each model's credit rates follow what it costs to run, and the larger models cost more per token. A typical message comes to about 33 credits on Claude Haiku 5.5, 660 credits on Claude Sonnet 5.5, and 1,260 credits on Claude Opus 5.5, so the model you pick decides how far your credits go.
Yes. The model is a setting of each agent, so a triage agent can run on a fast model while a research agent runs on a frontier one. All of them spend from the same credit balance, each within the share you allow it.
Once a prompt passes 100K tokens, cached tokens included, the whole request is billed at the long-context rate: 25 credits per 1K input tokens and 125 per 1K output, instead of 5 and 25. Shorter prompts stay at the normal rate.