Estimate multi-step AI agent pipeline costs. Configure planner, worker, and critic LLMs — see exact cost per run and monthly totals. Free, no signup.
Define your pipeline steps · Enter run volume · Results update live
Typical 3-step agent: planner → worker → critic. Monthly cost for 1,000 runs.
| Pipeline | Step 1 (Planner) | Step 2 (Worker) | Step 3 (Critic) | Cost / 1K runs |
|---|---|---|---|---|
| BudgetCHEAPEST | GPT-5.4 nano | GPT-5.4 nano | GPT-5.4 nano | ~$0.62 |
| Balanced | GPT-5.4 nano | GPT-5.4 ($2.50/$15) | GPT-5.4 nano | ~$35 |
| High Quality | Haiku 4.5 | Claude Sonnet 4.6 | Haiku 4.5 | ~$96 |
| Maximum | Sonnet 4.6 | Claude Opus 4.8 | Sonnet 4.6 | ~$330 |
Use cheap models (GPT-5.4 nano, Gemini 3.1 Flash-Lite) for planning and verification. Reserve expensive models (GPT-5.4, Claude Sonnet 4.6) only for the main worker step. This cuts costs 5–10× with minimal quality loss.
Deploy your AI agent infrastructure:
Model a 3-step AI agent pipeline and find the optimal cost/quality tradeoff.
A code review agent (planner: GPT-5.4 nano, worker: Claude Sonnet 4.6, critic: GPT-5.4 nano) processing 500 PRs/month: Planner = 500 × (1,500 × $0.20 + 300 × $1.25)/1M = $0.27. Worker = 500 × (3,000 × $3.00 + 800 × $15.00)/1M = $10.50. Critic = 500 × (2,000 × $0.20 + 150 × $1.25)/1M = $0.29. Total: ~$11.06/month for 500 code reviews.
How this agent cost model works.
Pipeline model: The calculator models a 3-step sequential agent: Planner (Step 1) → Worker (Step 2) → Critic (Step 3). Each step is modeled as a single LLM call. Real agents may have more steps, parallel branches, or variable step counts — use this as a baseline and scale by your actual step count.
Token assumptions: Default token counts represent typical production agent pipelines. Planner: 1,500 input / 300 output. Worker: 3,000 input / 800 output. Critic: 2,000 input / 150 output. The actual token count in your pipeline will depend on system prompt length, context accumulated across steps, and task complexity. We recommend profiling your agent in staging to get accurate per-step token counts before using the calculator for budget planning.
LLM prices: All model prices are official, pay-as-you-go rates as of June 2026. See the LLM API Cost Calculator for the full model list with context window sizes.
What is excluded: Vector database costs for agent memory or RAG retrieval, tool API costs (search APIs, code execution, external services), orchestration framework overhead (LangChain, LlamaIndex, custom orchestration), and compute/serverless infrastructure for running the agent. A complete AI agent cost estimate adds these components; the calculator models the LLM portion only.
Cost depends heavily on model choice and step complexity. A budget 3-step agent (all GPT-5.4 nano) costs approximately $0.00062 per run. A balanced pipeline (nano planner, GPT-5.4 worker, nano critic) costs ~$0.035/run. A quality pipeline using Claude Sonnet 4.6 for the worker step costs ~$0.105/run. For 1,000 runs/month: $0.62 to $105 depending on configuration.
GPT-5.4 nano at $0.20/$1.25 per 1M tokens is the cheapest capable model for agent tasks in 2026. Gemini 3.1 Flash-Lite ($0.25/$1.50) is a close second from Google. Both handle planning, routing, and verification tasks well. For the main worker step requiring complex reasoning, GPT-5.4 or Claude Sonnet 4.6 produce significantly better results despite their higher cost.
Key strategies: (1) Use cheap models (GPT-5.4 nano) for planning and verification — reserve expensive models for the worker step only. (2) Cache repeated tool calls and identical contexts (prompt caching on Anthropic saves 90% on cached input tokens). (3) Set a hard maximum steps per run — unbounded agents can cost 10–100× more than expected if they get stuck in loops. (4) Use streaming to detect early termination opportunities. (5) Batch similar agent tasks when latency allows (50% Batch API discount).
An AI agent pipeline is a multi-step LLM workflow where each step uses a separate model call: (1) Planner decides what actions to take based on the user's request, (2) Worker executes the main task (reasoning, tool use, code generation), (3) Critic reviews the result and decides if it meets quality criteria or should be retried. Total pipeline cost = sum of all LLM call costs per run × number of monthly runs.
No — this is the most common and most expensive mistake in agent design. Using Claude Opus 4.8 for all three steps of a simple pipeline is like using a surgeon for triage work. Optimal strategy: cheapest capable model for planner (routing/planning doesn't need top-tier intelligence), best model for worker (where quality directly impacts output quality), cheap model for critic (binary accept/retry decision). This "model mixing" approach typically reduces agent costs by 5–10× with minimal quality loss.
Several components add to LLM costs in a production agent: (1) Vector database for RAG memory storage — see Vector DB Cost Calculator; (2) Embedding API for converting documents to vectors — see Embedding API Cost Calculator; (3) External tool APIs (search, weather, database queries) — varies widely; (4) Serverless compute to run the orchestration — typically $5–20/month at modest scale; (5) Logging and monitoring for agent runs — $5–50/month depending on volume and verbosity. Total non-LLM costs often equal or exceed LLM costs for well-optimized pipelines.