Choosing a model for your agent: Haiku, Sonnet or Opus

Choosing a model for your agent is the decision that most shapes how it behaves and what it costs. Atroxabot runs agents on three Claude models from Anthropic, each a step up in capability and in credits per message. They share a lot: every one uses tools, reads images as well as text and supports prompt caching. What separates them is how hard they can think and what that thinking costs.

The context window is how much an agent can hold in view at once: its instructions, the conversation and anything it looked up. All three can hold far more than most agents need, so for most jobs the choice comes down to capability and cost.

Claude Haiku 5.5 is the fastest and least expensive model in the lineup. It is the right default for agents that run many times a day, where each task is small and the total adds up: a reply costs about 33 credits, so Pro's monthly credits cover around 7,575 typical messages.

• Quick replies and triage
• High-volume scheduled runs
• Summaries and lookups

Haiku is also the model on the Free plan, which makes it the place to start. Keep its context lean: once a request's prompt passes 100K tokens, the whole request is billed at a higher rate.

Claude Sonnet 5.5 balances speed and capability. It plans well, works through several tools in a row and writes clearly, which covers most agents people actually run. At about 660 credits a message, Pro's monthly credits cover around 378 typical messages.

• Research and writing
• Multi-step tool use
• Support and operations

If you are unsure which model an agent needs and Haiku falls short, Sonnet is the next step. It is available on Pro and Ultra.

Claude Opus 5.5 is the most capable model in the lineup. It holds a plan over many steps and reasons carefully before it acts, which matters when an agent has real work to finish on its own. It is also the most expensive: about 1,260 credits for a typical message, or around 198 messages on Pro's monthly credits.

• Long autonomous tasks
• Complex analysis and planning
• Work where mistakes are costly

Give Opus to the agents whose results you rely on most, not to every agent. A planning or analysis agent that runs a few times a week is a good fit; a briefing that runs every morning usually is not.

1. 1

   Start from what the agent does most

   Look at the job it will repeat, not the hardest thing it might ever face. Frequent, short tasks point to Haiku; research, writing and multi-step tool use to Sonnet; long autonomous work to Opus.

2. 2

   Try the lighter model first

   Run the agent on real work for a few days. If the results are good, you have the cheaper setup; if not, you know exactly what falls short.

3. 3

   Step up only where it helps

   Move an agent to a larger model when its results need it. Other agents can stay where they are: every agent has its own model.

4. 4

   Set the effort level

   Each agent can be set to an effort level from low to max. Higher effort lets the model reason for longer, which is billed as output. By default, Claude Haiku 5.5 uses medium, Claude Sonnet 5.5 uses high, and Claude Opus 5.5 uses medium.

5. 5

   Give it a budget

   Set the share of your credits the agent may use. How Atroxabot credits work shows what each model costs per message and how budgets cap spending.