The AI Agent Cost Calculator.

The honest answer: anywhere from a few dollars a month for a single-purpose agent to several thousand dollars a month for a high-volume customer support agent. The cost depends on three variables: which model you use, how many tokens each task consumes, and how many tasks you run per month.

The calculator above lets you plug in your specific workflow and volume. As a rough ballpark from our test workflows: a customer-support reply agent running 1,000 tasks a month costs about $1.25 on Claude Haiku 4.5, $7.50 on Sonnet 4.6, or $50 on Opus 4.7. A code-change agent at the same volume costs about $75 on Sonnet 4.6, $375 on Opus 4.7.

For most builder workflows, the right model lands in the $5–$200 per month range. The calculator shows you the side-by-side so you can pick the model that fits your budget and quality bar.

"Build" means two different things and the cost depends on which one.

If you mean using a no-code platform like Lindy or Relevance AI, the cost is the platform subscription ($19–$199/month for most tiers) plus the underlying API costs the calculator covers. Total typical cost: $30–$300 per month all-in.

If you mean building a custom agent with Claude Code or OpenClaw, the cost is your developer time (a week or two for a working v1) plus API costs from the calculator. The platform itself is free or included with a Claude subscription. The total cost depends almost entirely on the developer.

If you don't know which path you should take, run the five-question picker — it asks the right questions and recommends a starting point.

The most cost-effective model depends entirely on what the agent needs to do. There is no single answer.

• For high-volume, simple tasks (classification, short replies, sentiment analysis): Claude Haiku 4.5, Gemini 2.5 Flash, or DeepSeek V4 Flash. Costs in the $0.20–$1 per million tokens range.
• For balanced workflows (sales follow-ups, doc Q&A, code edits): Claude Sonnet 4.6 or Gemini 2.5 Pro. The sweet spot for most builders.
• For complex reasoning where output quality is critical: Claude Opus 4.7 or GPT-5.5. Worth the cost when the cheaper models actually fail at the task.
• For maximum cost optimisation: open-source models like Llama 3.3 70B (via Together AI) or DeepSeek V4 Flash, often 4–10× cheaper than frontier-tier models.

The calculator above shows the spread for any workflow you choose — usually 50–100× between the cheapest and most expensive model for the same task.

The cost scales linearly with task volume. If 1,000 tasks per month costs $7.50 on Sonnet 4.6, then 100,000 tasks per month costs $750. Use the volume preset buttons in the calculator (100, 1k, 10k, 100k) to see the actual numbers.

What changes at scale is the math on which model is right. At 100 tasks per month, the difference between Haiku and Opus might be $1 vs $5 — irrelevant. At 100,000 tasks per month, it's $100 vs $500 — a real budget decision.

Two cost-saving levers builders should know about:

• Prompt caching — repeat queries with similar prefixes can hit cache rates that drop input costs 50–90%. DeepSeek V4 Flash drops input from $0.14 to $0.0028 per million tokens for cache hits.
• Right-sizing the model — most teams default to the most powerful model and never test cheaper ones. Running Sonnet 4.6 instead of Opus 4.7 cuts costs 5× with minimal quality loss for 80% of workflows.

For high-volume, repetitive workflows: yes, by a large margin. A customer-support reply agent processing 5,000 tickets a month costs about $6 on Haiku versus a $4,500/month support agent handling the same volume. That's a 750× cost difference per task.

For low-volume, judgement-heavy work: probably not. The cost gap closes when the AI's mistakes need a human reviewer anyway, or when the task happens infrequently enough that the human's salaried time is essentially free.

The calculator shows the "vs hiring a [role]" card for each workflow with a fully-loaded US salary estimate. Adjust mentally if your team is offshore, contracted, or senior-tier.

Five practical levers, ranked by impact:

1. Drop to a cheaper model. Most teams over-spec. Test if Haiku 4.5 or Sonnet 4.6 handles your task before defaulting to Opus 4.7. The calculator shows the spread — usually a 5–25× cost difference between adjacent tiers.
2. Enable prompt caching if your agent makes repeat queries with similar context. Anthropic, OpenAI, and DeepSeek all support cache discounts of 50–90% on input tokens.
3. Trim the prompt. Most production prompts have unused context, redundant examples, or stale instructions. Halving prompt length halves your input cost.
4. Batch where possible. Some vendors offer batch APIs at 50% the standard rate (Anthropic Batch, OpenAI Batch). Useful for non-real-time workflows.
5. Use open-source for cheap-tier work. Llama 3.3 70B via Together AI is often 4× cheaper than the equivalent frontier model for simple tasks. DeepSeek V4 Flash is even cheaper for compatible workloads.