Guides · agent API desk
What an AI agent actually costs to run — per vendor, per step, with the arithmetic.
11 guides for teams paying API rates for tool-using agents. Each one prices a real agent trace on the current catalogue, shows where the money goes, and gives you the cap that stops a runaway loop.
Guide 1 · 9 min read · 2026-09-11
How Much Does an AI Agent Cost to Run? Cost Per Task in 2026
Worked cost traces for 5-, 12- and 30-step tool-using agent runs on Claude Sonnet 5, GPT-5.4, Gemini 3.8 Flash and DeepSeek V4 Flash, and why cost per successful task is the only number that matters.
how much does an AI agent cost · AI agent cost per task · cost to run an AI agent
Read the guide →Guide 2 · 9 min read · 2026-09-11
Cheapest LLM for AI Agents: 30 Models Ranked by Cost Per Run
Every current API model ranked by the cost of the same 12-step tool-calling run, a reliability tier list, the route-cheap-escalate-on-failure pattern with code, and when the flagship is cheaper per successful task.
cheapest LLM for AI agents · cheapest model for tool calling · cheapest LLM API for agents
Read the guide →Guide 3 · 9 min read · 2026-09-11
Claude Agent SDK Pricing: Anthropic Agent API Cost Explained
What a Claude agent loop costs on the Messages API and the Claude Agent SDK: Opus 5, Sonnet 5 and Haiku 4.5 traces, prompt caching arithmetic, Message Batches, and where a Claude Code subscription beats API billing.
Claude Agent SDK pricing · Anthropic agent API cost · Claude Messages API tool use cost
Read the guide →Guide 4 · 9 min read · 2026-09-11
OpenAI Agents SDK Pricing: Responses API Cost for Agents
What an OpenAI agent costs to run: GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5.4 nano and GPT-5.3-Codex on the Responses API and Agents SDK, with cached and batch arithmetic and the ChatGPT Pro and Business break-evens.
OpenAI Agents SDK pricing · Responses API cost · GPT-5.4 agent cost
Read the guide →Guide 5 · 9 min read · 2026-09-11
Gemini API Function Calling Cost: Gemini Agent Pricing Guide
What a Gemini agent costs to run: Gemini 3.1 Pro, 3.8 Flash and 3.5 Flash Lite on a function-calling loop, why 1M context is a cost trap, context caching and batch, the free tier, and Google AI Pro and Ultra versus the API.
Gemini API function calling cost · Gemini agent pricing · Gemini 3.8 Flash agent cost
Read the guide →Guide 6 · 9 min read · 2026-09-11
DeepSeek Agent Cost: V4 Flash and V4 Pro API Pricing for Loops
What DeepSeek V4 Flash and V4 Pro cost on a real 12-step agent run against Claude Sonnet 5, GPT-5.4 and Gemini 3.8 Flash, where DeepSeek tool calling holds up, where it does not, and a hybrid pattern that keeps a flagship for the last step.
DeepSeek agent · DeepSeek V4 API pricing · DeepSeek tool calling
Read the guide →Guide 7 · 8 min read · 2026-09-11
Prompt Caching Cost for AI Agents: Anthropic, OpenAI, Gemini
How prompt caching is priced on the Claude, OpenAI and Gemini APIs, what to put in the cached prefix, and a worked 12-step agent run showing the saving with the arithmetic laid out.
prompt caching cost · Anthropic prompt caching pricing · OpenAI prompt caching
Read the guide →Guide 8 · 8 min read · 2026-09-11
MCP Cost: What Model Context Protocol Servers Add in Tokens
MCP is free, but every connected server injects its tool schemas into every request. This article measures that overhead, shows what ten servers do to a 12-step run on four models, and gives four ways to cut it.
MCP cost · Model Context Protocol token cost · MCP server pricing
Read the guide →Guide 9 · 9 min read · 2026-09-11
Multi-Agent System Cost: Why Orchestrators Cost 3 to 6x
The token accounting behind orchestrator-plus-worker systems built with LangGraph, CrewAI, the OpenAI Agents SDK or the Claude Agent SDK, with a priced 3-worker research task and the model routing that keeps it affordable.
multi-agent system cost · LangGraph cost · CrewAI cost
Read the guide →Guide 10 · 8 min read · 2026-09-11
LLM Cost Control for Agents: 5 Guardrails to Limit Spend
Five guardrails that stop an AI agent from spending your month's budget in an afternoon, a priced example of a 200-step runaway loop on each flagship model, and a 30-line Python runner that aborts at a dollar cap.
LLM cost control · limit AI agent spend · agent budget
Read the guide →Guide 11 · 8 min read · 2026-09-12
Codex Cloud vs Local Cost: GPT-5.3-Codex Pricing Compared
Prices the GPT-5.3-Codex API against ChatGPT subscription and self-hosted usage patterns, using the official 12-step reference run to show what a coding agent actually costs.
codex cloud vs local cost · GPT-5.3-Codex pricing · Codex cloud API cost
Read the guide →Questions before you build
How much does it cost to run an AI agent?
It depends on how many steps the agent takes, because every step re-sends the growing conversation. A 12-step tool-using run costs cents on DeepSeek V4 Flash or Gemini 3.8 Flash and low dollars on Claude Opus 5 or GPT-5.5. The guides below show the arithmetic for each vendor.
Which is the cheapest model for tool calling?
DeepSeek V4 Flash at $0.05 input / $0.16 output per million tokens is the cheapest catalogue row that reliably follows tool schemas, followed by GLM 4.7 Flash and Gemini 3.5 Flash Lite. Claude Haiku 4.5 and GPT-5.4 mini cost more but fail less on ambiguous tasks.
Does prompt caching really cut agent cost?
Yes. Cached input reads are roughly a tenth of the list input price on Anthropic and OpenAI, and most of an agent's tokens are the system prompt, tool definitions and earlier steps sent again on every call. Caching that prefix typically removes half or more of the bill.
Are these prices current?
Every number in these guides comes from the same catalogue that powers the pricing tracker, and the rate panel on each page is rebuilt from the live catalogue on every deploy.