New blogMaking an AI Agent You Can Trust With Money
logo

★ “Why is our AI bill so high?”

LLM API Cost Calculator

Put in your traffic and see your monthly bill across 14 current models side by side, including the prompt-caching and batch discounts most calculators forget.

  • Runs in your browser
  • No signup
  • Updated 7 Nov 2026

Workload

tok
tok
%

Repeated system prompt or docs

%

Jobs that can wait hours

Models

Anthropic
OpenAI
Google
Mistral
DeepSeek

$ per 1M tokens, input / output

Monthly cost · 30 days

Cheapest
$105
Claude Haiku 5.5
Priciest
$2,100
Gemini 4 Argon
Spread
20.0×
priciest vs cheapest
Per 1K users / mo
Claude Haiku 5.5Cheapest
$105$52.50
Claude Sonnet 5.5
$2,100$1,050
GPT-6.1 Sol
$2,100$1,050
Gemini 4 ArgonIntro price
$2,100$1,050
Per 1K users assumesrequests per user per day.

Where the money is

If 60% of your traffic is simple enough for Claude Haiku 5.5, routing it there instead of Gemini 4 Argon saves about $1,197/month.

Caching half your input on Gemini 4 Argon (a shared system prompt or document) would cut about $570/month.

You're looking at $105–$2,100/month.

Most AI products I audit can cut that 30–60% with routing, caching and shorter prompts.

Get a cost-cutting audit
01

How the math works

Every provider bills input tokens (your prompt, system prompt and retrieved documents) and output tokens (the reply) separately, per million. Output is usually 4–5× the input price, so chatty apps are output-heavy and RAG apps are input-heavy.

cost per request = (fresh_in × input_price
+ cached_in × cached_price
+ out × output_price) / 1,000,000
× (1 − batched_share × batch_discount)
monthly = cost per request × requests per day × 30

Two adjustments are applied automatically. Introductory prices (currently on some Gemini models) are used until their end date. Long-prompt surcharges kick in when your average input crosses the provider's threshold: 272K tokens on current OpenAI models, 100K on Claude Haiku 5.5.

Not modelled: cache-write premiums, priority or fast tiers, regional surcharges, image and audio tokens, and reasoning tokens a model spends before answering (they bill as output, so add them to your output estimate if you use extended thinking).

02

Worked example

A customer-support bot with retrieval: each request sends about 2,000 input tokens (system prompt, the question and a few retrieved help-centre chunks) and gets back about 300 tokens. It handles 10,000 requests a day.

Monthly cost of a RAG support bot by model
ModelPer requestMonthly50% input cached
GPT-6 Astra$0.0350$10,500$7,800
Claude Opus 5.5$0.0140$4,200$3,060
Claude Sonnet 5.5$0.00700$2,100$1,530
GPT-6.1 Sol$0.00700$2,100$1,530
Gemini 4 Argon$0.00700$2,100$1,530
DeepSeek V4.1 Flash$0.00096$288$200
Claude Haiku 5.5$0.00035$105$78.00
GPT-6 Luna$0.00035$105$78.00

Same traffic, same prompts: GPT-6 Astra costs $10,500 a month and GPT-6 Luna costs $105, a 100× gap. The system prompt and help-centre chunks rarely change, which makes this workload a good fit for prompt caching: the last column shows what happens when half the input is a cache hit.

03

Pricing by model

Each model has its own page with the full price card, caveats and a calculator preset to it.

04

Head-to-head

The comparisons people search for most, at three traffic levels.

05

Questions founders actually ask

Are these prices current?
They were last checked on 7 Nov 2026 against each provider's own pricing documentation, which is linked in the Sources section. Models with prices I could not confirm from a primary source are left out rather than guessed.
What is prompt caching and should I use it?
When the start of your prompt (system prompt, tool definitions, a long document) repeats across requests, providers can serve it from cache at a fraction of the input price, often 90–99% off. If your requests share a large fixed prefix, caching is usually the single biggest saving available. Put the stable content first and the variable content last.
How do I estimate my token counts?
For English text, a rough rule is about 4 characters or three-quarters of a word per token. The reliable way is to log the usage numbers the API returns on a sample of real requests and use the averages. Remember to include the system prompt, tool definitions and any retrieved context in the input count.
Is the cheapest model good enough?
Often for part of your traffic, rarely for all of it. A common pattern is routing: a small model handles classification, extraction and easy questions, and only the hard requests go to a larger model. Measure quality on your own examples before switching; a cheaper model that needs retries or human fixes is not cheaper.
Does self-hosting an open-weight model save money?
Only above a certain steady volume. Hosted APIs for open-weight models are already cheap, and self-hosting adds GPU rental that bills whether or not you have traffic, plus the engineering time to run it. It tends to pay off when traffic is high and predictable, or when data must stay on your own infrastructure.
Why does my real bill differ from this estimate?
Usually because of tokens you are not counting: reasoning or thinking tokens (billed as output), retries, tool-call round trips, and cache-write premiums. Compare the estimate against the usage page in your provider console for a week and adjust the averages.
06

Sources

All prices are list prices from each provider's own documentation, standard tier, in USD, last verified on 7 Nov 2026. Providers change prices often; if a number here disagrees with the source, the source wins.

07

More from the workbench

01 · Plan & scopeMVP Cost EstimatorPick features and who builds it. Get an hour-by-hour estimate, not an agency teaser.03 · Money & fundraisingRunway & Burn CalculatorMonths of cash left, the date it hits zero, and whether you are default alive.04 · Money & fundraisingSaaS Metrics CalculatorsLTV, CAC, churn, NRR, burn multiple and more, each with the formula and benchmarks.

Fine print

Tools: on the house.
Engineer: by the hour.

Knowing your AI bill was free. Cutting it in half isn't, but I'm cheaper than the overspend.

I'm Himanshu, an AI and full-stack engineer. I build MVPs, AI agents and the growth plumbing around them for early-stage startups.

Hire me →My projects