Skip to content

llm_cost_by_usecase

What does an LLM actually cost for your workload?

Token prices alone do not tell the story. A semantic search service at 10K requests/day looks nothing like a research agent at 150. These five real workload profiles are priced across Claude, GPT, and Gemini, every figure computed from the same model as the AI Agent Cost Calculator so they never drift as prices change.

Semantic Search with RAG

Real-time search across documentation or knowledge base. User query → embedding lookup → context retrieval → LLM ranking & formatting.

Daily requests10,000
Avg input1.2k
Avg output400
Tool calls1

cost_across_models

ModelPer requestPer monthvs cheapest
Claude Haiku
Anthropic
$0.0091$2,71911.2×
Claude Sonnet
Anthropic
$0.03$10,17741.8×
Claude Opus
Anthropic
$0.17$50,857208.9×
GPT-4o mini
OpenAI
$0.0016$4802.0×
GPT-4o
OpenAI
$0.03$7,88232.4×
Gemini 1.5 Flash
Google (Vertex)
$0.0008$243Cheapest
Gemini 1.5 Pro
Google (Vertex)
$0.01$3,94516.2×
token_breakdown · baseline

Total input/request: 7.3k (base + agentic context re-send)

Total output/request: 800

LLM turns per request: 2

Customer Support Agent

Multi-turn support conversation. Agent retrieves ticket history + KB → responds → optionally calls tools (create ticket, escalate, fetch order).

Daily requests2,000
Avg input2.5k
Avg output800
Tool calls2

cost_across_models

ModelPer requestPer monthvs cheapest
Claude Haiku
Anthropic
$0.02$1,34211.4×
Claude Sonnet
Anthropic
$0.08$5,02542.7×
Claude Opus
Anthropic
$0.42$25,113213.3×
GPT-4o mini
OpenAI
$0.0039$2332.0×
GPT-4o
OpenAI
$0.06$3,82832.5×
Gemini 1.5 Flash
Google (Vertex)
$0.0020$118Cheapest
Gemini 1.5 Pro
Google (Vertex)
$0.03$1,91616.3×
token_breakdown · baseline

Total input/request: 15.9k (base + agentic context re-send)

Total output/request: 2.4k

LLM turns per request: 3

Code Generation & Refactoring

IDE copilot or PR review agent. Given code context → suggest implementation or refactor → run linter checks → iterate.

Daily requests3,500
Avg input4.0k
Avg output1.2k
Tool calls3

cost_across_models

ModelPer requestPer monthvs cheapest
Claude Haiku
Anthropic
$0.04$4,21711.8×
Claude Sonnet
Anthropic
$0.15$15,81344.2×
Claude Opus
Anthropic
$0.75$79,065221.1×
GPT-4o mini
OpenAI
$0.0068$7152.0×
GPT-4o
OpenAI
$0.11$11,91833.3×
Gemini 1.5 Flash
Google (Vertex)
$0.0034$358Cheapest
Gemini 1.5 Pro
Google (Vertex)
$0.06$5,95916.7×
token_breakdown · baseline

Total input/request: 26.2k (base + agentic context re-send)

Total output/request: 4.8k

LLM turns per request: 4

Research & Summarization Agent

Multi-step research loop: query web API → fetch pages → extract key facts → synthesize → cite sources.

Daily requests150
Avg input8.0k
Avg output2.0k
Tool calls5

cost_across_models

ModelPer requestPer monthvs cheapest
Claude Haiku
Anthropic
$0.14$6786.7×
Claude Sonnet
Anthropic
$0.53$2,41523.9×
Claude Opus
Anthropic
$2.63$11,892117.9×
GPT-4o mini
OpenAI
$0.02$1561.5×
GPT-4o
OpenAI
$0.41$1,88518.7×
Gemini 1.5 Flash
Google (Vertex)
$0.01$101Cheapest
Gemini 1.5 Pro
Google (Vertex)
$0.20$9659.6×
token_breakdown · baseline

Total input/request: 115.5k (base + agentic context re-send)

Total output/request: 12.0k

LLM turns per request: 6

Content Moderation at Scale

High-volume screening: classify text + flags + confidence + optional manual escalation. Runs low-cost model with spot checks by premium model.

Daily requests50,000
Avg input300
Avg output100
Tool calls0

cost_across_models

ModelPer requestPer monthvs cheapest
Claude Haiku
Anthropic
$0.0006$1,0058.1×
Claude Sonnet
Anthropic
$0.0024$3,64529.5×
Claude Opus
Anthropic
$0.01$18,045145.8×
GPT-4o mini
OpenAI
$0.0001$2031.6×
GPT-4o
OpenAI
$0.0017$2,67021.6×
Gemini 1.5 Flash
Google (Vertex)
$0.0001$124Cheapest
Gemini 1.5 Pro
Google (Vertex)
$0.0009$1,35811.0×
token_breakdown · baseline

Total input/request: 300 (base + agentic context re-send)

Total output/request: 100

LLM turns per request: 1

model_selection

When to reach for cheap vs premium

Use cheap models when

  • High volume (10K+ req/day) on a simple task: classification, ranking
  • Fallback or batch processing that tolerates 24h latency
  • The user never sees the model choice (backend only)

Use premium models when

  • Complex reasoning is needed: research, code review, multi-step agents
  • It is user-facing and accuracy beats cost: support, search
  • Tool-calling reliability is critical (agent success equals revenue)

cost_trends

What drives the bill

Semantic search + RAG. Embedding cost ($0.02 per 1M tokens) is often 10–20% of total. High-volume search favors cheap models because LLM cost dominates.

Agentic loops (support, research). Tool calls multiply input tokens. A 5-tool agent can cost 3–5× as much as single-turn. Budget accordingly.

Code generation. Large context (entire files) plus premium models (accuracy matters) gets expensive fast. Batch or offline processing is cheaper than IDE real-time.

Moderation at scale. 50K req/day on cheap models is feasible. Premium spot-checks at a 5% escalation rate add ~$50/mo but catch the edge cases.

next_steps

Customize your numbers: the AI Agent Cost Calculator lets you tune token counts, request volume, and tool calls for your exact workload.

Understand the concepts: read the tokens glossary term or explore model comparisons.

Want this priced against your real traffic?

The calculator runs the numbers; an audit tells you which workload is quietly burning the budget. Book a call, or leave your email and I'll reach out.

Book a call

No spam. You'll get a reply from me.

Prefer proof first? See how this plays out in real case studies →