Get API key

For developers and teams paying for LLM APIs

Cut your LLM bill without putting a proxy in front of your app.

Send every prompt to the cheapest model that can handle it.

Your keys never leave your app. FewerTokens picks the cheapest model that can do the job, never pricier than your default. You pay 5% of what you save, and nothing if you don't save.

Get a free API key Estimate my savings
$ pip install fewertokens

Python 3.10+ · works with your OpenAI, Anthropic and Google Gemini keys

One request, routed

Example

Prompt

"How do I reset my password? I signed up with Google."

Your default model

gpt-4o

$2.50 in · $10.00 out
per 1M tokens

Model we pick

gpt-4o-mini

$0.15 in · $0.60 out
per 1M tokens

  1. SDK sends the prompt and conversation messages to FewerTokens and gets back a model name.
  2. Your app calls OpenAI directly with your own key.
  3. We record token counts and show the saving in your dashboard.

Prices: FewerTokens model catalog v1.2.1, USD list prices.

If most of your traffic goes to GPT-4o or Claude Sonnet, you're probably paying top prices for easy questions.

  • Support and FAQ bots
  • RAG and knowledge assistants
  • Ticket triage and tagging
  • Extraction and rewriting

How it works

Keep your keys. Keep your default model. We just pick a cheaper one when it's good enough.

1

Add the SDK with your keys

Install the Python SDK, pass your provider keys and the model you use today. That model is your default and the baseline we measure savings against.

2

We pick the model for each prompt

For every request, FewerTokens reads the prompt and picks the cheapest model that can handle it, from our list of 66 models. Never more expensive than your default.

3

Your app calls the model directly

The SDK calls the provider from your own server with your own key. Afterwards it reports token counts so you can see what you saved.

Your app + FewerTokens SDK

  • Holds your OpenAI / Anthropic / Gemini keys
  • Receives the full model response

To FewerTokens → prompts and conversation messages

← model name back

To your provider → your key + prompt

← the response

FewerTokens

Your keys never leave your app.

Prompts are sent to FewerTokens for classification.

OpenAI · Anthropic · Google

Billed to your own account, as today

Live demo

Try it: see which model we'd pick for a prompt.

This runs our real prompt classifier and model list. No LLM is called, so nothing is answered, but you see the decision and both prices.

Pick an example or write your own

How demanding is this prompt?

ExampleSimple

SimpleModerateComplex

Your default

gpt-4o

$2.50 in · $10.00 out per 1M tokens

Model we'd pick

gpt-5-nano

$0.050 in · $0.40 out per 1M tokens

Estimated cost of this request · 412 input + 256 output tokens

$0.00359$0.00012

97% cheaper on this request ($0.00347 saved; our fee would be $0.00017).

Reply length assumed at 256 tokens. Prices are list prices per 1M tokens from our model catalog (v1.2.1), input and output priced separately.

Savings calculator

What would you save each month?

Pick the model you use today, enter your usage, and choose roughly how hard your prompts are. We use real list prices from our model catalog.

gpt-4o · $2.50 in · $10.00 out per 1M tokens

2. Your monthly usage

3. Your prompt mix

Complex (stays on your default)40%

Estimated monthly result

Today, all on your default
$135.00
LLM cost with routingpaid to your providers, as today
$58.86
FewerTokens fee 5% of the savings
$3.81
New total per month
$62.67

You keep (net savings / month)

$72.33

54% lower

Prompt typeModel used
Simple · 0%gpt-5-nano$0.050 in · $0.40 out
Moderate · 60%gpt-4o-mini$0.15 in · $0.60 out
Complex · 40%gpt-4o (your default)$2.50 in · $10.00 out

Estimate only. The mix is your assumption; the live router decides prompt by prompt and may also choose cheaper models from other providers you've connected. If nothing is saved, the fee is $0.

Measured in production

Picking the model adds a fraction of a second.

On 13 Sep 2026 we sent 60 routing requests to our production API (api.fewertokens.in) with a mix of short support questions and long ~32,000-character RAG prompts. Here's what we measured.

Successful calls

60/60

0 errors

Median (p50)

374 ms

all 60 calls

95th percentile (p95)

552 ms

slowest: 568 ms

By prompt size (p50)

153 ms short

530 ms ~32k chars

Time for the routing call only (the SDK asking FewerTokens which model to use). Your model's own response time is unchanged.

Pricing

5% of what we save you. $0 if we save nothing.

Savings = what a request would have cost on your default model, minus what it actually cost on the model we picked, at list prices. No plans, no seats, no per-request fee. You keep paying OpenAI, Anthropic or Google directly, as you do today.

fee = 5% × (cost on your default − cost on routed model)
See pricing details

Worked example

Illustrative

A support bot on gpt-4o uses 30M input + 6M output tokens a month. Assume 60% of it is simple enough for gpt-4o-mini.

Today, all on gpt-4o
$135.00
Saved on the 60% slice
$76.14
FewerTokens fee (5%)
$3.81
New total per month
$62.67 (54% lower)

Security and data handling

We're not a proxy. Your keys stay with you.

  • Your keys never leave your app.
  • Routing adds p50 374 ms, p95 552 ms (60 production calls, Sep 2026).

Stays in your app

  • Your OpenAI, Anthropic and Gemini API keys.
  • The model call itself. Your app calls OpenAI, Anthropic or Google Gemini directly with your own key, so the model request goes straight from your app to the provider.
  • Your provider bill. You pay your provider directly, and we only invoice 5% of what you save.

What FewerTokens sees

  • Prompts and the conversation messages are sent to FewerTokens to decide which model fits.
  • Token counts after the call, to work out savings and the fee.
  • Our servers and database run on Oracle Cloud in Mumbai, India (ap-mumbai-1). Request logs are stored with Axiom, outside India. Your provider processes prompts wherever it runs.
Open-source SDK (MIT) on PyPI

66 models across 10 providers, kept up to date for you.

When providers release models or change prices, we update the list. You don't change any code. OpenAI, Anthropic and Gemini have built-in SDK options; the others connect through extra_providers.

  • OpenAI 16
  • Anthropic 9
  • Google Gemini 8
  • DeepInfra 9
  • Together 8
  • Mistral 7
  • Groq 3
  • xAI 3
  • DeepSeek 2
  • Cohere 1

Quickstart

Running in about five minutes.

  1. 1

    Get a FewerTokens API key

    Sign in and create a key in the dashboard. It starts with tr_live_.

  2. 2

    Install the SDK

    pip install fewertokens
  3. 3

    Add your provider key and default model

    Use the model you call today. Savings are measured against it.

Read the docs
app.py
from fewertokens import FewerTokens

ft = FewerTokens(
    api_key="tr_live_...",         # from your dashboard
    openai_api_key="sk-proj-...",  # stays in your app
    openai_default_model="gpt-4o", # what you use today
)

response = ft.chat.completions.create(
    messages=[{"role": "user", "content": "Reset my password"}],
)

print(response.content)
print(response.routing.model_used)      # e.g. "gpt-4o-mini"
print(response.routing.baseline_model)  # "gpt-4o"
Also works with anthropic_api_key / gemini_api_key, streaming (stream=True) and async (acreate).

Stop paying top-model prices for easy prompts.

No card needed to start. You only pay when we save you money.

Get a free API key