Type what you spend on model calls each month — the numbers move as you type. We apply the reduction we actually measured on production, not a projection.
Calculated at ~90% off the model bill: measured 92–96% on a replay of 11,794 real coding-agent calls, and 89–93% of context once a project holds about a hundred facts. Long-running agents save the most; short one-off calls save less. Your provider bills you directly; Iconia’s fee is shown separately, so what you see below is what you keep.
every month, after Iconia’s fee — your model bill drops by $0
Keep your own model and key. Change one line.
Every call to every model starts from nothing. The model does not remember your codebase, your decisions, your customers, or the conversation you had with it an hour ago. The only way it can know any of that is if you send it again — and you are billed for every word of it, every single time. That resend is not a small line item. On a mature project it is most of the bill.
The context that makes your agent useful — conventions, architecture, prior decisions — is not stored anywhere the model can reach. It rides along on each call, and it is charged at full price each time, even though nothing about it changed.
A one-line question against a mature project costs many times what the same question costs on day one. The work did not get harder. The payload did. Left alone, the price of every interaction rises for as long as the project succeeds.
When everything must travel on every call, the overwhelming majority of each payload has nothing to do with the request. You are paying to transmit context the model will not use — and burying the part it needs inside it.
Iconia gives the work a memory that outlives the conversation. What your project knows is held once, and each request is accompanied by the part of it that request actually needs — not the whole archive. The model still receives everything necessary to answer well. It simply stops being handed everything else.
What you send reflects the question, not the accumulated history of the project. This is where the money goes back in your pocket, and it is the reason the saving widens rather than narrows as you keep working.
On questions answerable only from project knowledge, models with no memory got 13% right. With Iconia, the same models on the same questions got 100%. Cheaper is only worth having if the answers hold — these got better, not worse.
The same memory serves Claude, GPT, Gemini, Llama and models you run yourself. Switch providers for price or capability without rebuilding the context that makes your agent worth using.
The reduction is not a fixed discount. It is the gap between a payload that grows and one that does not — so the longer a project runs, the wider it gets. Measured against carrying full context, at three project sizes:
| What the project knows | Carry everything | With Iconia | Reduction | Answers correct |
|---|---|---|---|---|
| A new project · 8 facts | $0.0079 | $0.0058 | 19–45% | 9 / 9 |
| A month in · 40 facts | $0.0310 | $0.0071 | 75–86% | 9 / 9 |
| A real codebase · 100 facts | $0.0718 | $0.0072 | 89–93% | 9 / 9 |
Read the second column: carrying everything rose ninefold as the project matured. The Iconia column barely moved. That flatness is the whole product — and note the last column, which never moved either.
Those rows measure a project as it accumulates facts. Session length is the second axis, and it is the steeper one. Over a real twenty-turn session the conversation grew to 14,393 tokens while the memory Iconia injected held at 247 — from 37% of the payload down to 1.7%. That is where up to 98% comes from, and it keeps climbing the longer a session runs.
A number without its baseline is not evidence. Each of these carries the comparison it came from, and every one is reproducible.
1,000 memories free. Your provider keys stay yours.
The figures above are real and reproducible. They are also bounded, and we would rather you learn the bounds here than discover them later.
The saving comes from what you would otherwise resend. If your project knows almost nothing yet, there is little to avoid sending — 19–45% at eight facts. The calculator assumes a project with real accumulated knowledge.
Memory reaches a 7B model correctly and it answers well — until a large agent tool payload is also present, at which point the model cannot hold everything at once. That is a limit of the model, not of the memory, but it is a real limit.
Six to nine questions per condition, consistent and repeatable across models and runs. Directionally strong; not a published benchmark suite, and we will not present it as one.
We have measured that each call gets smaller. Whether an agent also finishes a task in fewer steps is plausible, untested, and therefore not claimed.
No. Your model calls run on your own provider account, so your provider bills you directly, at the lower amount — your key stays yours and we never substitute ours for it. Add it once in your dashboard, or send it alongside your Iconia key. Iconia is priced as a flat subscription, not a cut of your inference, so we have no incentive to send more tokens than your question needs.
It did not. Correctness rose from 13% to 100% on project-specific questions, and held at 9 out of 9 at every project size we tested. Removing what a request does not need makes the part it does need easier to use, not harder.
Neither, in testing from 8 to 1,000 memories. The right memory ranked first at every size, and the time added stayed under ten milliseconds and flat. Growing stores are where most memory products degrade, so we measured that one hardest.
You point your existing setup at Iconia and keep your tools, your models and your code. There is no library to adopt and nothing to migrate. If you turn Iconia off, you are exactly where you started.
No. What you store is yours, readable by your account, and exportable whenever you want it. Nothing is harvested from your traffic — what gets remembered is what you choose to store.
Iconia saves you up to 98% of it — on your model, with your key.
Start saving free