NEW Design partner program · UAE first

More AI work.Same budget.

Tokpad sends every AI request to the lowest-cost model that still meets your quality bar and your region rules. Keep your provider keys. Change one line of code.

Get early access Estimate my savings
  • Your keys, your invoices
  • No token markup
  • Nothing to self-host
routing simulationillustrative
tokpad.
Requests0
All on premium$0.00
With Tokpad$0.00
Simulated prices per request. Not live traffic.

Built for the models you already use

The paradox

Tokens got cheaper. Your AI bill still went up.

Agents chain many calls, reasoning models use more tokens per answer, and more apps move from pilot to production. Most teams send all of it to one premium model, because nobody has time to decide request by request.

>280×

Drop in inference cost for GPT-3.5-level performance between 2022 and 2024.

Stanford AI Index 2025
10×+

Typical price gap between premium and efficient model tiers for the same task.

Public provider price lists
70%

Of teams building multi-model apps expected to use an AI gateway by 2028, up from about 10%.

Gartner, Market Guide for AI Gateways (2025)

Cost

Premium models handle routine work like classification, extraction and summaries, at 10× the price they need to.

Quality

Switching blindly to a cheaper model causes retries and rework. You need evidence on your own workloads, not on benchmarks.

Location

Sensitive prompts must stay in approved places. Logs, retries and fallbacks can quietly break that promise.

Cheaper models don’t lower the bill on their own. Something has to decide which model each request deserves. That’s Tokpad.

How it works

Up and running in an afternoon. Not a quarter.

No gateway to host, no routing rules to write, no model-by-model integration. You connect once and we run the optimisation loop for you.

01

Connect

Add the provider keys you already have. Your accounts, contracts and invoices stay yours.

~5 minutes
02

Set your policy

Pick a monthly budget, a quality bar and the regions you allow. Three choices, no rule-writing.

~5 minutes
03

Change one line

Point your OpenAI-compatible SDK at Tokpad. Each request now goes to the cheapest model that passes.

1 line of code
04

See the proof

Track cost per accepted task, failures and a full audit trail. Savings you can show your CFO.

Always on

Your code barely notices.

Tokpad speaks the OpenAI API format, so your existing SDKs, prompts and tools keep working. If a cheaper route ever misses your quality bar, it falls back to your premium model automatically.

You keepProvider accounts, keys, contracts and invoices. We never mark up tokens.
You dropSelf-hosting a gateway, writing routing rules and guessing which model is good enough.
from openai import OpenAI client = OpenAI(    base_url="https://api.openai.com/v1",    base_url="https://api.tokpad.com/v1",    api_key=TOKPAD_KEY,) # Everything else stays exactly the sameclient.chat.completions.create(model="auto", messages=msgs)

Endpoint shown is illustrative. Advanced SDK features are compatibility-tested per pilot.

Savings calculator

What would Tokpad save you?

Move the slider to your monthly AI API spend and pick the kind of work you run. Our fee is already included, so this is what you’d actually keep.

What does most of your AI work look like?
Fine-tune the assumptions
Estimated net savings
$7,320

36.6% of your AI budget, every month · $87,840 a year

Today: one premium model for everything$20,000
With Tokpad, fee included$12,680
Model spend (paid to your providers)Tokpad fee
Strong fit. Most of your traffic could run on lower-cost models.
Model spend$11,600
Tokpad fee$1,080
PlanStarter

Illustrative estimate, not a promise. Formula: spend × eligible share × cost reduction = gross saving; fee = platform fee + % of the optimised model spend that runs through Tokpad (proposed pricing). Real results depend on output length, retries, caching and the quality bar you set. We measure your workloads before routing anything.

Live demo

Policy first. Then price. Then proof.

Pick a request and a region mode, and watch how Tokpad decides. It runs in your browser and calls no real models.

1 · Choose a request

2 · Region mode

POST /v1/chat/completions
“Classify this ticket: I can’t download my last invoice.”
1Policy filter
2Quality and price
3Decision
—
Regional controls · UAE first

Decide where your AI runs. Then prove it.

Banking and health data in the UAE must stay onshore, and new data-protection rules arrive in 2027. Tokpad gives you two clear modes instead of a page of settings.

Available from day one

Flexible

Use any region and provider you approve. The most savings and the fastest start.

  • Your approved list of providers and regions
  • Unapproved routes blocked automatically
  • Every decision logged and exportable
Strict UAE

Nothing leaves the UAE.

Inference, logs, retries and telemetry stay in-country. If no approved local model can handle a request, it stops. Fail-closed by design.

  • Local models like Falcon and Jais on UAE infrastructure
  • In-country logs and retries
  • Third-party audit before launch

The boundary covers the whole lifecycle of a request:

InferenceLogsRetries and fallbacksObservabilityStorageSubprocessors

Strict UAE mode launches only once local inference is independently verified. A UAE gateway alone does not guarantee in-country processing. Not legal advice.

Pricing

You pay when we save you money.

A flat platform fee plus a small share of the model spend that runs through Tokpad. Your providers keep billing you directly, with no markup on tokens.

Starter

For $10k–25k monthly AI spend

$500 /mo + 5%
  • Bring your own keys
  • Flexible regions
  • Budgets and cost dashboard
  • Monthly savings report
Join as design partner
Most teams

Growth

For $25k–100k monthly AI spend

$1,500 /mo + 3%
  • Everything in Starter
  • Evaluation engine and quality gates
  • Unlimited apps and teams
  • SSO
Join as design partner

Sovereign

For regulated and enterprise teams

From $5,000 /mo
  • Strict UAE mode
  • Dedicated deployment
  • Audit exports
  • SLA and priority support
Talk to us
Measure before you pay for routingDesign partners start with a baseline of their real workloads before any traffic is moved.
We tell you if it won’t pay offIf your work is mostly complex, savings after our fee can be small. We say so up front.
No lock-inYour keys and provider accounts never move. Switch back with the same one-line change.

Proposed pricing in USD, confirmed per pilot. Percentages apply to optimised model spend routed through Tokpad.

Roadmap

From gateway to regional control plane.

Now · design partners

Gateway

OpenAI-compatible endpoint, your own keys, budgets, cost dashboard, flexible regions.

Next

Evaluation engine

Workload baselines, quality gates, route recommendations and savings reports.

Then

Strict UAE mode

Verified local inference, in-country logs and retries, fail-closed, third-party audit.

Later

Agents and GCC

Per-agent budgets and tool-call governance. Saudi region policy and Arabic evaluation sets.

Questions

Straight answers.

Can’t find yours? Ask it in the form below and a founder will reply.

How much work is it for my team?
About an afternoon. Add your provider keys, choose a budget, quality bar and region mode, then change the base URL in your code. We handle routing, monitoring and fallbacks. Some workloads need a short benchmark first.
Will a cheaper model hurt quality?
Only routes that pass your quality bar on your own workloads are used. If a cheaper model misses, the request falls back to your premium model automatically, and you see every decision.
Can you guarantee savings?
No, and we won’t pretend to. Savings depend on your mix of work, model prices and quality bar. That’s why we measure a baseline first and tell you up front if the numbers don’t work.
Do you store our prompts?
Not by default. Each customer is isolated, and retention is something you switch on, not something you have to opt out of. The calculator and demo on this page run entirely in your browser.
Does my data stay in the UAE?
In Flexible mode, it goes only to the providers and regions you approve. Strict UAE mode keeps inference, logs and retries in-country and blocks anything else. It launches once local inference is independently verified.
Which models are supported?
Any OpenAI-compatible provider, including OpenAI, Anthropic, Google and open-weight models such as Llama, Falcon, Jais and DeepSeek. The exact list is confirmed for each pilot.
Design partner program

Start with one workflow.Prove the savings.

We’re onboarding a small group of teams already spending on AI APIs. You get hands-on help from the founders and shape what we build.

  1. A 20-minute callWe learn which AI workloads you run and what you spend.
  2. A baseline of your real trafficWe measure cost and quality before anything is routed.
  3. Savings you can verifyTurn on routing for one workflow and see the difference in your own invoices.

Get early access

A founder replies within one business day.

No sensitive or regulated data, please. We only use this to contact you.

Thanks, you’re on the list.

A founder will email you within one business day to set up a short call.