Agents are not single requests. Model the steps, tool calls, retries, and run volume that actually drive the bill — then compare every model on cost per run.
100% free · no signup · browser-local
Share on XModel a multi-step run with tool calls and retry overhead.
Cost / run
$0.04
Cost / day
$20.79
Cost / month
$623.70
Tokens / run
11.6K
Input tokens
9.2K
Output tokens
2.3K
Cost / step
$0.0069
Retry waste
$0.0038
15,000 runs per month at the current workload.
| Model | Per run | 1,000 runs | Per month | vs selected |
|---|---|---|---|---|
| GPT-6 LunaOpenAI | $0.0021 | $2.08 | $31.19 | -95% |
| Claude Haiku 5.5Anthropic | $0.0021 | $2.08 | $31.19 | -95% |
| DeepSeek FlashDeepSeek | $0.0028 | $2.77 | $41.58 | -93% |
| GPT-4o MiniOpenAI | $0.0028 | $2.77 | $41.58 | -93% |
| Qwen 2.5 72BAlibaba | $0.0042 | $4.16 | $62.37 | -90% |
| Llama 3.3 70B (Groq/Together)Meta | $0.0073 | $7.28 | $109.15 | -83% |
| Gemini 3.5 Flash-LiteGoogle | $0.0085 | $8.55 | $128.20 | -79% |
| DeepSeek V4 ProDeepSeek | $0.01 | $10.67 | $160.08 | -74% |
| Gemini 3.8 FlashGoogle | $0.02 | $15.59 | $233.89 | -62% |
| GPT-5.4 MiniOpenAI | $0.02 | $17.33 | $259.88 | -58% |
| Grok 4.3xAI | $0.02 | $17.33 | $259.88 | -58% |
| Grok 4.7xAI | $0.03 | $32.34 | $485.10 | -22% |
| Mistral Large 2Mistral | $0.03 | $32.34 | $485.10 | -22% |
| GPT-6.1 SolOpenAI | $0.04 | $41.58 | $623.70 | — |
| Claude Sonnet 5.5Anthropic | $0.04 | $41.58 | $623.70 | — |
| GPT-4o (Omni)OpenAI | $0.05 | $46.20 | $693.00 | +11% |
| Command R+ (Cohere)Cohere | $0.05 | $46.20 | $693.00 | +11% |
| Gemini 3.1 Pro (Preview)Google | $0.05 | $46.20 | $693.00 | +11% |
| Claude Opus 5.5Anthropic | $0.08 | $83.16 | $1,247 | +100% |
| Gemini 4 ArgonGoogle | $0.08 | $83.16 | $1,247 | +100% |
| GPT-5.5OpenAI | $0.12 | $115.50 | $1,732 | +178% |
| GPT-6 AstraOpenAI | $0.21 | $207.90 | $3,119 | +400% |
| Claude Fable 5.1Anthropic | $0.21 | $207.90 | $3,119 | +400% |
Retries multiply both input and output tokens. Real agents also pay for failed tool calls, embedding lookups, and orchestration latency; this model covers the LLM tokens only.
A 6-step agent sends six prompts, each carrying the growing scratchpad. Halving the number of steps often saves more than switching models.
Search results, database rows, and API JSON come back as input tokens. Trim payloads before they enter the context, not after.
A failed step that is retried pays for the failed tokens and the successful tokens. Track retry rate as a first-class cost metric.
Planning needs a frontier model. Classification, extraction, and formatting can run on a cheap model. Use the Smart Router tool to price the split.
Start with the cost per run, then multiply by real traffic. A $0.04 run sounds trivial until it happens 50,000 times a month. The simulator shows both numbers side by side so the forecast stays honest.
Once you know the cost, test the architecture: reduce steps, route simple steps to a cheaper model, and trim tool payloads. The Smart Router calculator prices the cascade, and the Token Budget Planner tells you how much headroom the monthly budget leaves.
This simulator prices LLM input and output tokens. It does not include vector database reads, tool API fees, sandbox compute, or orchestration infrastructure.
Compare 100% frontier traffic against a light → general → complex model cascade.
Calculate GGUF weights + KV cache, match RTX / Apple Silicon / Strix Halo rigs, and compare Ollama, llama.cpp, MLX & vLLM.
Find the break-even volume where fine-tuning beats long few-shot prompts.
We'll email you when model prices change or new tools launch. Unsubscribe anytime.