Paid API Service Categories in the Agent Economy: A Spend Taxonomy

1. Overview

Autonomous AI agents have become net consumers of paid APIs across four dominant categories: (1) inference (LLM token generation), (2) search and retrieval (web search, vector DBs, scrapers), (3) research and structured knowledge (curated data feeds, research subscriptions, financial data), and (4) compute and execution (sandboxes, browsers, code runners). Empirical observation of public agent frameworks — AutoGPT, LangChain, CrewAI, OpenAI's Agents SDK, Anthropic's computer-use traces — and vendor pricing pages indicates that inference still dominates agent budgets by 60–80%, but the fastest-growing share is going to search/retrieval and research APIs as agents shift from chat-style assistants to long-horizon task executors [P4]. This synthesis ranks the four categories by spend volume, adoption rate, and cost-per-task, and identifies which categories behave as commodity layers versus durable infrastructure.

2. Key Findings

• Inference is the largest spend bucket but the fastest-commoditising. Frontier model pricing has fallen roughly an order of magnitude every 12–18 months: GPT-4 launched at $30/1M input tokens in March 2023; GPT-4o landed at $2.50/1M input tokens (OpenAI pricing — https://openai.com/api/pricing/); Gemini 1.5 Flash sits at $0.075/1M input tokens (Google pricing — https://ai.google.dev/pricing); DeepSeek-V3 at $0.27/1M input tokens (DeepSeek pricing — https://api-docs.deepseek.com/quick_start/pricing). For a typical agent loop consuming 50–200K tokens per task, raw inference cost-per-task has fallen from ~$3 to ~$0.05 in 24 months. [EMPIRICA ANALYSIS]