LLM cost calculator
What a call actually costs, on every major model at once. Set your typical input, output, and monthly volume; the table computes per-call and per-month costs from the providers' own published rates. Counting a specific prompt instead? Use the token counter.
| Model | Input $/M | Output $/M | Per call | Per month |
|---|---|---|---|---|
| Claude Opus 4.8 Anthropic | $5 | $25 | — | — |
| Claude Sonnet 5 Anthropic intro pricing ($2/$10) through Aug 31, 2026 | $3 | $15 | — | — |
| Claude Haiku 4.5 Anthropic | $1 | $5 | — | — |
| GPT-5.6 Sol OpenAI | $5 | $30 | — | — |
| GPT-5.6 Terra OpenAI | $2.5 | $15 | — | — |
| GPT-5.6 Luna OpenAI | $1 | $6 | — | — |
| GPT-5.5 OpenAI | $5 | $30 | — | — |
| Gemini 3.1 Pro Google prompts over 200K tokens bill at $4/$18 | $2 | $12 | — | — |
| Gemini 3 Flash Google | $0.5 | $3 | — | — |
| Grok 4.5 xAI | $2 | $6 | — | — |
| Grok 4.3 xAI | $1.25 | $2.5 | — | — |
| DeepSeek v4 Flash DeepSeek cache-miss input rate; cache hits are $0.0028 | $0.14 | $0.28 | — | — |
| Llama 4 Maverick Meta (via Together AI) no first-party Meta API; Together AI's hosted rate | $0.27 | $0.85 | — | — |
Rates verified against provider pricing pages on 2026-07-13. Models delisted from their provider's current pricing page are omitted rather than guessed. Compare what fits in each model in the context-window comparison.
LLM pricing questions, answered
How is LLM API cost calculated?
Per token, in two meters: every token the model reads (your prompt, plus any context) bills at the input rate, and every token it writes bills at the output rate. Output rates run three to six times the input rate on most frontier models, which is why long answers cost more than long questions.
Which LLM is cheapest?
Among the models in this table, DeepSeek and hosted Llama are an order of magnitude cheaper than the frontier models, and each provider's fast tier (Haiku, Luna, Flash) undercuts its flagship several times over. Cheapest per token is not cheapest per solved problem, though; a stronger model that answers in one attempt often beats a weak one that needs five.
Why do my real costs come out higher than this?
Because conversations resend context. In a chat loop, every turn re-sends the history, so the input meter grows quadratically with conversation length. The single biggest cost lever is controlling what context each call actually carries, not switching models.
Are these prices current?
Every rate comes from the provider's own pricing page and was last re-verified on 2026-07-13. Models the providers have delisted from their current pricing pages are deliberately absent rather than guessed. Footnotes mark intro pricing and tiered rates.
The biggest line item in that table is resent context. NodePad puts your conversations on a canvas where you choose exactly what each model reads, so the input meter carries your decisions instead of your scrollback.
Try NodePad free