Estimate your agent costs
How agent pricing works
Agent costs aren't just per-token. Each agent run includes: input tokens (system prompt + conversation history + tool definitions), output tokens (response), tool execution overhead (multiple model calls for tool selection + tool results processing), and retries/failures. A single "run" can trigger 3-15 model calls depending on complexity.
What typical agents cost
| Agent type | Per run | 100/mo | 1K/mo | 10K/mo |
|---|---|---|---|---|
| Simple RAG agent (GPT-4o) | $0.0004 | $0.04 | $0.40 | $4.00 |
| Complex research agent (GPT-4o) | $0.003 | $0.30 | $3.00 | $30.00 |
| Multi-agent system (GPT-4o) | $0.02 | $2.00 | $20.00 | $200.00 |
Estimated costs using GPT-4o API pricing. Actual costs vary by system prompt size, tool definitions, and retry logic.
Plugsky comparison
Plugsky's flat plans make agent costs predictable. No surprises — pay one price for unlimited runs.
| Volume | GPT-4o | Claude Sonnet | Plugsky |
|---|---|---|---|
| 1K agent runs/mo | ~$4 | ~$3 | $20 flat |
| 10K runs/mo | ~$40 | ~$30 | $50 flat |
| 100K runs/mo | ~$400 | ~$300 | $100 flat |
At high volumes, Plugsky's flat pricing becomes dramatically cheaper.
Cost drivers
What affects your agent costs the most:
- Number of tool calls per run — each call = additional model interaction, often 2-5× the base cost
- System prompt size — long prompts mean more input tokens consumed on every single run
- Context retention — keeping full conversation history vs. summarization has a massive token impact
- Model choice — smaller models like plugsky-micro are 10-50× cheaper per token than frontier models
- Retry logic — each retry doubles the cost of that step
How to reduce agent costs
Use smaller models for simple tasks (plugsky-micro), limit conversation history, compress tool descriptions, use caching, and batch where possible. Example strategies:
| Strategy | Estimated savings | Effort |
|---|---|---|
| Use smaller model for tool selection | 40-60% | Low (routing config) |
| Compress conversation history | 30-50% | Medium (summarization logic) |
| Cache common tool results | 20-40% | Medium (cache layer) |
| Batch independent agent runs | 15-30% | Low (design pattern) |
| Switch to Plugsky flat rate | 60-90% | Minimal (one API key) |
Self-hosting break-even
When does it make sense to self-host your own LLM infrastructure? The table below shows rough break-even points based on monthly run volume.
| Volume | GPT-4o (API) | Self-host (7B) | Self-host (70B) | Plugsky |
|---|---|---|---|---|
| 1K runs/mo | ~$4 | ~$150 | ~$800 | $20 |
| 100K runs/mo | ~$400 | ~$200 | ~$900 | $100 |
| 1M runs/mo | ~$4,000 | ~$500 | ~$2,000 | $250 |
Self-host assumes GPU rental at ~$1/hr (7B) or ~$3/hr (70B). Break-even for small models starts around 50K runs/mo. See full analysis at Private LLM Cost Estimator.
Frequently asked questions
What is an "agent run"?
An agent run is a single end-to-end execution of an AI agent task, from receiving the user's request to producing a final response. It may involve multiple model calls for planning, tool selection, tool execution, and result synthesis.
How many tokens does a typical agent use?
A simple RAG agent uses roughly 2K input + 500 output tokens per query. A complex research agent can use 10K input + 4K output + multiple tool calls. Multi-agent systems can easily exceed 50K tokens per workflow.
Does Plugsky charge per agent run?
No. Plugsky offers flat-rate plans with unlimited API calls. This means your agent costs are predictable regardless of how many runs, tokens, or tool calls each agent makes.
Can I use different models for different agent steps?
Yes. With Plugsky's one-API platform you can route different agent steps to different models. Use a cheap model for tool selection and a powerful model for final synthesis — all through a single API key.
Is Plugsky cheaper than building my own agent infrastructure?
For most teams, yes. Plugsky's flat-rate plans eliminate per-token costs, infrastructure management, and multi-provider complexity. At 100K runs per month, Plugsky is 4× cheaper than GPT-4o and 3× cheaper than Claude.
Last updated Jul 2026. Prices and availability verified at time of writing — check provider pages for current rates.
Related tools
- Plugsky AI Agents API — build agents with flat-rate pricing
- LLM Cost Calculator — compare per-token prices across models
- Private LLM Cost Estimator — self-hosting break-even analysis