Paste your prompt, pick your knobs, see the exact cost across every major model. Accounts for cached-input pricing, batch-mode discounts, and context-window limits.
Cache hit: OpenAI (50% off) & Anthropic (90% off) charge reduced rates for cached input tokens (prefix reuse on multi-turn or RAG). Batch: OpenAI & Anthropic both offer 50% off for non-realtime batch processing with 24h turnaround.
| Model | Context | Input / Output per 1M | Cost / Call | Monthly | vs Baseline |
|---|
Pricing source: Official provider pages, verified April 2026. Rates are US dollars per 1,000,000 tokens. Cache hit % applies to input tokens only (OpenAI 50% off cached, Anthropic 90% off cache reads). Batch mode applies 50% off both input and output on OpenAI & Anthropic endpoints that support it. Self-hosted Llama rates are rough cloud-GPU inference estimates, not a direct API price. Rows highlighted in red exceed the model's context window and aren't usable for this prompt size. For production billing, always double-check with the provider's official calculator — volume discounts and enterprise rates are not modeled here.
This calculator turns a prompt into a side-by-side cost comparison across every major LLM API — GPT-4o and friends, the Claude family, Gemini, DeepSeek, and hosted Llama. Paste text (or enter token counts directly), set an expected output size and monthly call volume, and see per-call and per-month cost for each model.
It also models the pricing details that dominate real bills: cached-input discounts (50% on OpenAI, 90% on Anthropic), the Batch API's 50% async discount, and each model's context-window ceiling.
Say each call sends 1,200 input tokens and generates 500 output tokens, at 100,000 calls per month. On a frontier model priced around $2.50/M input and $10/M output, that's roughly $0.008 per call — about $800/month. The same workload on a small model at $0.15/M input and $0.60/M output runs about $48/month. If 80% of your input is a reusable prefix, caching cuts the frontier-model bill by another chunk — the calculator applies each provider's specific discount so you can compare honestly.
Generating tokens costs providers much more compute than reading them, so output is typically priced 3–5× higher than input. Workloads that produce long responses (writing, code generation) are dominated by output cost; classification and extraction workloads are dominated by input.
When consecutive calls share a long identical prefix — a system prompt, few-shot examples, or a document — providers can reuse the computation. OpenAI charges 50% and Anthropic 10% of the normal input rate for those cached tokens. The cache-hit slider models what fraction of your input qualifies.
Batch mode is 50% off in exchange for asynchronous delivery (results within 24 hours, usually much faster). It's ideal for backfills, evaluations, nightly summarization — anything without a user waiting on the response.
Prices are maintained against the providers' published pricing pages, but they change often. Treat the output as a planning estimate and confirm the current rate card before signing off on a budget.