Iceburg opens soon. Join the waitlist to get in first. Free, no card.
Token optimization for AI apps

Pay for 200 tokens, not 700.

Iceburg sits between your app and your model and cuts what your model is billed for — about half (45–52%) on real chat traffic in our tests, and up to ~75% on long generative prompts (explain, write, analyse) — while your users still get a full, detailed answer. Your model, your output quality. The savings are the product; how we get them is ours.

Measured on our own test runs so far, not yet on customer traffic — your savings depend on your prompts. Every response reports its own savings_percent, so your dashboard shows your real number from the first call. See our runs, including the ones where it saved nothing.

one endpoint · any model
curl -X POST https://projecticeburg.com/api/v1/process \
  -H "Content-Type: application/json" \
  -d '{
    "api_key": "ig_...",
    "model_config": {"invoke_url": "https://api.openai.com/v1/chat/completions",
                     "model": "gpt-4o-mini", "api_key": "sk-..."},
    "prompt": "<your long prompt>"
  }'

{
  "answer": "...",
  "original_tokens": 704,
  "billed_tokens": 216,
  "savings_percent": 69.3
}

Why teams add it

One integration. Nothing changes downstream except your bill.

01 · same answers

Quality your users won’t notice a change in

Iceburg optimizes what your model is billed for, not what it says. The response that reaches your user reads the same as before — full length, full detail.

02 · drop-in

Any model, any provider, one endpoint

Keep the model you already chose. Point your calls at Iceburg, or bring your own provider config. No retraining, no prompt rewrites, no lock-in.

03 · measured

Every response tells you what it saved

Each call returns its own savings_percent, so the cost reduction is a number you can log and audit — not a claim you have to trust.

Where it pays off

Iceburg earns the most on long, context-heavy prompts — the longer the prompt, the bigger the savings.

Best fit

  • Long document summarisation
  • RAG pipelines with heavy retrieved context
  • Chat histories replayed on every turn
  • Repeated boilerplate system prompts

Biggest wins

  • Summarisation and analysis of big inputs
  • Agents that replay context every step
  • High-volume production traffic
  • Anything where the prompt is long

Same quality

Your users get the same full-length, detailed answer they’d get from your model today — you just pay for a fraction of the tokens.

Tokens saved, all users, live
0

across 0 requests and counting

Opening soon

Add Iceburg to your AI and keep the difference

Join the waitlist and you’ll be among the first in. When we open, it’s one endpoint change and every long prompt costs less.