LLM API cost calculator
Updated
In short: LLM API cost per request = (uncached input × input price + cached input × cached price + (output + reasoning) × output price) ÷ 1,000,000, prices per million tokens. Multiply by monthly requests. This calculator does it for BroadNet's models and 428 others at providers' own prices, and shows when a flat plan is cheaper.
BroadNet rows use BroadNet Inference's published per-1M-token prices, synced 2026-11-14, before tax. Models added from the list use their cheapest standard price, read from that provider's own pricing page (checked 2026-11-15; every provider per model). A model with no cached price bills cached input at the input price. Pay-as-you-go keys allow 4 requests in flight and 30 a minute. BroadNet Gemma 4 VL 26B is also sold as a flat $0.50 hourly key.
Report a wrong price or number (email to info@broadnet.ai, with this page's link and your inputs filled in).
How it works
- Pick a preset close to your workload, or type your own traffic: requests per day and days per month.
- Enter the average input tokens per request and the share served from cache, then visible output tokens and any reasoning tokens.
- Optionally add any of 428 other models by name (each at its cheapest standard price from the provider's own page), or your current model with its own prices per 1M tokens.
- Read the table: cost per request, per day and per month for every BroadNet model, plus yours if added, cheapest first. Copy the link or download CSV to share the estimate.
The formula
cost per request = ( input × (1 − cache) × input_price
+ input × cache × cached_price
+ (output + reasoning) × output_price ) / 1,000,000
monthly cost = cost per request × requests per day × daysPrices are per one million tokens. A 2,000-token prompt is 0.002 million tokens, so it costs 0.002 times the input price. Reasoning tokens are billed as output. See input and output tokens, cached input tokens and reasoning tokens.
Typical token counts
| Workload | Input / request | Cache | Output / request |
|---|---|---|---|
| Support chatbot | 1,500 | 60% | 300 |
| RAG over documents | 8,000 | 30% | 500 |
| Coding agent (one developer) | 40,000 | 85% | 2,000 (incl. reasoning) |
| Summarise long documents | 30,000 | 0% | 1,000 |
| Classify or tag text | 400 | 70% | 20 |
| Read invoices (vision) | 2,500 | 20% | 600 |
Compare against 428 other models
"Add any model" loads a 428-model price list (about 32 KB) only when you use it. Each model comes in at its cheapest standard-tier price, read from the provider's own pricing page by 41 providers in all, and the row names that provider; batch, flex and priority tiers and free rows are left out. The LLM API pricing pages list every provider's price with its source. Share a comparison with ?model= in the link, for example Claude Sonnet 5.5 and GPT-6 Luna.
These are starting points, not measurements. Log the usage object from a week of real traffic, including prompt_tokens_details.cached_tokens, and use those averages.
Questions
How is LLM API cost calculated?
Providers bill per million tokens, with separate prices for input, cached input and output. Cost per request = uncached input × input price + cached input × cached price + (output + reasoning) × output price, each divided by 1,000,000. Multiply by requests per day and days per month.
Do reasoning tokens cost money?
Yes. Reasoning models write hidden "thinking" tokens before the answer. They are billed as output tokens even though you do not see them, so a short answer can cost far more than its length suggests. Add them in the reasoning field. Why reasoning can empty a reply.
What is a cache hit and why does it lower cost?
When the start of a prompt matches a recent request, the provider reads it from cache and bills it at the cached-input price, which is much lower. Agents and chatbots that resend the same system prompt and history get high cache rates. How to raise your cache rate.
How many tokens is 1,000 words?
About 1,290 tokens for English prose, by our measurement across five real tokenizers, so one token is about 0.78 words. Modern Standard Arabic is about 2,080 tokens per 1,000 words and Python code about 2,126. See the tokenizer efficiency data, or count your own text with the token counter.
When is a flat plan cheaper than paying per token?
For DeepSeek V4.1 Flash, the Full Moon plan costs $60 per session per month with no token meter; one session runs one request at a time. When the monthly per-token cost for DeepSeek V4.1 Flash passes that price, the calculator says so. The break-even math.
Can I compare GPT, Claude, Gemini, Llama or Kimi prices here?
Yes. Use "Add any model" to search 428 models, from GPT, Claude, Gemini and Grok to DeepSeek, Qwen, GLM, Kimi, Llama and Mistral. Each joins the table at its cheapest standard price, read from that provider's own pricing page (41 providers, checked 2026-11-15), and the row names the provider. Every provider's price for each model is on the LLM API pricing pages.
Go deeper
Read as MarkdownAsk ChatGPTAsk ClaudeAsk Perplexity
Sources
- BroadNet Inference machine-readable prices (pricing.json)
- AI21 Labs: pricing (checked 2026-11-15)
- AkashML: pricing (checked 2026-11-15)
- Alibaba Cloud Model Studio: pricing (checked 2026-11-14)
- Amazon Bedrock: pricing (checked 2026-11-14)
- Amazon Nova (Bedrock): pricing (checked 2026-11-14)
- Anthropic: pricing (checked 2026-11-14)
- Baidu: pricing (checked 2026-11-15)
- Baseten: pricing (checked 2026-11-14)
- Cerebras: pricing (checked 2026-11-14)
- Chutes: pricing (checked 2026-11-15)
- Cloudflare Workers AI: pricing (checked 2026-11-14)
- Cohere: pricing (checked 2026-11-14)
- CoreWeave Forge (formerly W&B Inference): pricing (checked 2026-11-15)
- Crusoe Cloud: pricing (checked 2026-11-15)
- DeepInfra: pricing (checked 2026-11-14)
- DeepSeek: pricing (checked 2026-11-14)
- DigitalOcean Gradient (Serverless Inference): pricing (checked 2026-11-14)
- Featherless: pricing (checked 2026-11-15)
- Fireworks AI: pricing (checked 2026-11-14)
- FriendliAI: pricing (checked 2026-11-15)
- GMI Cloud: pricing (checked 2026-11-15)
- Google (Gemini API): pricing (checked 2026-11-14)
- Google Cloud Vertex AI: pricing (checked 2026-11-14)
- Groq: pricing (checked 2026-11-14)
- Hyperbolic: pricing (checked 2026-11-14)
- IBM watsonx.ai: pricing (checked 2026-11-15)
- Microsoft Azure (Azure OpenAI + Foundry Models): pricing (checked 2026-11-14)
- MiniMax: pricing (checked 2026-11-14)
- Mistral AI: pricing (checked 2026-11-14)
- Moonshot AI (Kimi): pricing (checked 2026-11-14)
- Novita AI: pricing (checked 2026-11-14)
- Nscale: pricing (checked 2026-11-15)
- OpenAI: pricing (checked 2026-11-14)
- Parasail: pricing (checked 2026-11-14)
- Perplexity: pricing (checked 2026-11-15)
- SambaNova Cloud: pricing (checked 2026-11-14)
- SiliconFlow: pricing (checked 2026-11-14)
- Together AI: pricing (checked 2026-11-14)
- Venice: pricing (checked 2026-11-15)
- xAI: pricing (checked 2026-11-14)
- Z.ai: pricing (checked 2026-11-14)