LLM API pricing calculator.
Per-million-token prices are useless until they meet your actual traffic. Set your request size and monthly volume, and see what Claude, GPT-5.6, Gemini and DeepSeek really cost side by side. Verified against official pricing pages, updated when providers reprice.
≈ 1,504 words of prompt + context
≈ 602 words of response
| Model | $ / 1M in | $ / 1M out | Per request | Per month | |
|---|---|---|---|---|---|
| DeepSeek V4 FlashDeepSeekCache-miss rate; cached input drops to ~$0.003. | $0.14 | $0.28 | $0.0005 | $5.04 | |
| GPT-5.4 NanoOpenAI | $0.2 | $1.25 | $0.0014 | $14.00 | |
| DeepSeek V4 ProDeepSeekCache-miss rate; cached input drops to ~$0.004. | $0.435 | $0.87 | $0.0016 | $15.66 | |
| Gemini 3.1 Flash-LiteGoogle | $0.25 | $1.5 | $0.0017 | $17.00 | |
| Claude Haiku 4.5Anthropic | $1 | $5 | $0.0060 | $60.00 | |
| GPT-5.6 LunaOpenAI | $1 | $6 | $0.0068 | $68.00 | |
| Gemini 3.5 FlashGoogle | $1.5 | $9 | $0.010 | $102 | |
| Gemini 3.1 ProGooglePrompts over 200K tokens bill at $4 / $18. | $2 | $12 | $0.014 | $136 | |
| GPT-5.6 TerraOpenAI | $2.5 | $15 | $0.017 | $170 | |
| Claude Sonnet 5AnthropicIntro pricing of $2 / $10 runs through Aug 31, 2026. | $3 | $15 | $0.018 | $180 | |
| Claude Opus 4.8Anthropic | $5 | $25 | $0.030 | $300 | |
| GPT-5.6 SolOpenAI | $5 | $30 | $0.034 | $340 | |
| Claude Fable 5Anthropic | $10 | $50 | $0.060 | $600 |
List prices per 1M tokens, standard tier, verified July 18, 2026. Batch tiers are typically 50% cheaper; prompt caching cuts repeated input cost by ~90%. Token estimates assume ≈1.33 tokens per English word.
Token costs, decoded.
Every LLM API bills the same way: input tokens (your prompt, system instructions and context) at one rate, output tokens (the model’s response) at a higher one. The spread matters. A summarization workload that sends 10,000 tokens in and gets 300 out lives on input pricing; a code generator that sends 500 in and streams 5,000 out lives on output pricing. That is why the calculator asks for both numbers instead of a single “tokens per request”.
Two discounts change the math at scale and are not in the table: batch processing halves list price on every provider shown here if you can wait minutes-to-hours for results, and prompt caching drops repeated input (your system prompt, your RAG context) to roughly a tenth of list price on Anthropic, OpenAI and Google.
Cheapest model vs cheapest outcome.
The table sorts by monthly cost, and the temptation is to read it bottom-up. Resist it for anything user-facing. The pattern that actually wins in production is routing: a cheap, fast model (Haiku, Flash-Lite, V4 Flash) handles the 80 percent of requests that are easy, and a flagship (Sonnet 5, Opus 4.8, Gemini 3.1 Pro) gets the 20 percent where quality is the product. Teams that do this well typically spend less than teams that standardized on a single mid-tier model, and their hard cases work better.
When we build AI products at techpotions this routing layer is one of the first things we put in, because it is the difference between an API bill that scales with revenue and one that scales with usage.
Common questions.
01How do LLM API prices work?
You pay per token, separately for input (everything you send: system prompt, context, the user message) and output (everything the model writes). Prices are quoted per million tokens, and output is typically 3 to 6 times more expensive than input. A token is roughly three-quarters of an English word, so 1,000 words is about 1,330 tokens.
02How much does the Claude API cost?
As of July 18, 2026: Claude Haiku 4.5 is $1 per million input tokens and $5 per million output, Claude Sonnet 5 is $3/$15 (with intro pricing of $2/$10 through August 2026), Claude Opus 4.8 is $5/$25, and Claude Fable 5 is $10/$50. Prompt caching cuts repeated input to about a tenth of list price, and the batch API halves everything if you can wait for results.
03Which LLM API is cheapest in 2026?
DeepSeek V4 Flash at $0.14/$0.28 per million tokens is the cheapest capable model from a major provider, with GPT-5.4 Nano and Gemini 3.1 Flash-Lite close behind. Cheapest per token is not cheapest per solved task: a stronger model that answers correctly in one attempt regularly beats a budget model that needs three tries and human cleanup.
04How do I estimate tokens from words?
Multiply English words by about 1.33 to get tokens; code and non-English text run higher. For real precision use the provider tokenizer endpoints, but for cost planning the 1.33 rule and a safety margin of 20 percent is what most teams actually budget with.
05How can I cut my LLM API bill?
In order of impact: route easy requests to a cheaper model and keep the flagship for hard ones; use prompt caching so your system prompt and context stop being re-billed at full price on every call; trim what you stuff into context; use batch endpoints for anything that can wait; and cap output length, since output tokens are the expensive ones. Teams routinely cut 60 to 90 percent with routing and caching alone.
06Are these prices current?
Prices were checked against the official Anthropic, OpenAI, Google and DeepSeek pricing pages on July 18, 2026, standard non-batch tier. Providers reprice a few times a year, so treat the comparison as planning-grade and confirm on the provider page before committing a budget.
- llms.txt generatorGenerate a spec-correct llms.txt file for your site, so AI assistants read you right.
- App development cost calculatorScope a web or mobile app and get an honest 2026 cost and timeline range.
- AI receptionist cost calculatorCompare an AI receptionist against in-house staff and answering services, in dollars.
Building something on top of these APIs?
We build AI products for a living: agents, RAG, voice, and the model routing that keeps the bill sane. If you want the estimate turned into an architecture, tell us what you're building.