Why Agent Workflows Cost More Than a Single Prompt
An agent loops: each tool call resends the growing context. A task with 8 steps can use many times the tokens of one prompt.
Typical monthly workload
- Input tokens
- 35,000,000
- Output tokens
- 4,000,000
- Requests / day
- 1,200
Illustrative estimate. Adjust the sliders below to match your own volume.
Recommended setup
Use a strong model to plan and a cheap one to execute routine tool steps, with a hard cap on steps per task.
Monthly cost by model for this workload
| Model | Input / 1M | Output / 1M | Est. monthly cost |
|---|---|---|---|
| Gemini 1.5 Flash | $0.075 | $0.30 | $3.83 |
| Gemini 2.0 Flash | $0.100 | $0.40 | $5.10 |
| DeepSeek V3 | $0.140 | $0.28 | $6.02 |
| GPT-4o Mini | $0.150 | $0.60 | $7.65 |
| Qwen 2.5 72B | $0.350 | $0.40 | $13.85 |
| Llama 3.3 70B (Groq/Together) | $0.590 | $0.79 | $23.81 |
| DeepSeek R1 (Reasoning) | $0.550 | $2.19 | $28.01 |
| Claude 3.5 Haiku | $0.800 | $4.00 | $44.00 |
| o3-mini (Reasoning) | $1.100 | $4.40 | $56.10 |
| Gemini 1.5 Pro | $1.250 | $5.00 | $63.75 |
| Mistral Large 2 | $2.000 | $6.00 | $94.00 |
| Grok 2 | $2.000 | $10.00 | $110.00 |
| GPT-4o (Omni) | $2.500 | $10.00 | $127.50 |
| Claude 3.5 Sonnet | $3.000 | $15.00 | $165.00 |
| o1 (Reasoning) | $15.000 | $60.00 | $765.00 |
Computed from list prices and the workload above (35,000,000 input and 4,000,000 output tokens per month), with no caching or batch discounts. Real bills vary.
How to cut this cost
- Cap the number of steps
Set a maximum iteration count and a token budget per task to prevent runaway loops.
- Trim tool output
Return only the fields the next step needs instead of full API or page responses.
Simulate your own spend
Quick Workload Presets
Pick a scenario or fine-tune token parameters and cloud host below.
Token & Usage ParametersLLM API
Switching from GPT-4o ($195.00/mo) to Gemini 1.5 Flash ($5.85/mo) saves 97%.
Lowest cost per token, high-frequency tasks, real-time audio/video
State-of-the-art open economics, high-performance general intelligence
Real-time conversational streaming, vision, multimodal processing at speed
High-volume text processing, fast chatbots, structured JSON extraction
Top open-weights coding, math, and multilingual benchmark performance
Open-weights privacy, self-hosting or ultra-cheap hosted inferencing
Complex math, algorithmic reasoning, open deep-thinking workflows
Rapid code completion, customer service, fast structured parsing
Cost-effective scientific & coding reasoning, high-accuracy logic
Massive context (2M tokens), full video/audio analysis, large codebases
Multilingual tasks, European compliance, enterprise code generation
Real-time information synthesis, creative generation, vision analysis
Complex multimodal reasoning, coding, and production enterprise apps
Software engineering, nuanced writing, complex agentic tool use
Advanced mathematics, algorithmic coding, PhD-level scientific logic
Frequently asked questions
Why do agents burn so many tokens?
Every step re-reads the accumulated context plus tool results. Costs grow with the number of steps, so step caps and context trimming matter most.