Loading page content
Estimate what an AI agent costs to run, accounting for the thing single-call calculators miss. Context grows on every turn, so cost per turn rises through the run.
Estimate the cost of running AI agents across LLM providers. Configure your agent's expected usage including tool calls per run, then compare pricing for Anthropic, OpenAI, Google, Groq, and DeepSeek models.
LLM Calls / Day
400
Agent Runs / Month
3,000
Cheapest / Month
$7.20
Most Expensive / Month
$810.00
| Model | Cost / Call | Daily | Monthly | Yearly |
|---|---|---|---|---|
| GPT-4o-miniOpenAIcheapest | $0.0024 | $0.240 | $7.20 | $87.60 |
| Gemini 2.5 FlashGoogle | $0.0024 | $0.240 | $7.20 | $87.60 |
| Llama 4 MaverickGroq | $0.0028 | $0.280 | $8.40 | $102.20 |
| DeepSeek R1DeepSeek | $0.0088 | $0.878 | $26.34 | $320.47 |
| Claude Haiku 4.5Anthropic | $0.014 | $1.44 | $43.20 | $525.60 |
| Gemini 2.5 ProGoogle | $0.030 | $3.00 | $90.00 | $1,095 |
| o3OpenAI | $0.032 | $3.20 | $96.00 | $1,168 |
| GPT-4oOpenAI | $0.040 | $4.00 | $120.00 | $1,460 |
| Claude Sonnet 4.6Anthropic | $0.054 | $5.40 | $162.00 | $1,971 |
| Claude Opus 4.7Anthropic | $0.270 | $27.00 | $810.00 | $9,855 |
Monthly Cost Comparison
Recommendations
Best for Prototyping
GPT-4o-mini
$7.20/mo via OpenAI
Lowest cost option. Great for development, testing, and low-stakes automation.
Best for Production
DeepSeek R1
$26.34/mo via DeepSeek
Strong balance of quality and cost. Reliable for production agent workloads.
Best for Complex Reasoning
Claude Opus 4.7
$810.00/mo via Anthropic
Top-tier intelligence for multi-step reasoning, complex tool use, and critical decisions.
Prices as of May 2026. Includes tool call overhead (3 tool calls = 4x LLM round-trips per agent run). Batch API and cached prompt discounts not included.
Agent cost is quadratic, not linear, and this surprises nearly everyone the first time. Each turn resends the entire conversation so far, so a ten-turn task does not cost ten times a single call. It costs the sum of a growing context, which is closer to fifty times. Tool results make it worse, because a large result stays in context for every subsequent turn.
Two things bring it down substantially. Prompt caching, which most providers now offer, discounts the repeated prefix heavily and directly targets the quadratic term. And context pruning. Dropping tool results that are no longer needed, or summarizing older turns, stops the context growing without limit. An agent that reruns a failing step ten times is the pathological case, which is what the Agent Loop Cost Predictor models specifically.