Token savings
Your agents send the model a lot of content it already has. Gate sits between them and the provider, sends the repeats once, answers identical requests from its own cache, and stops a runaway job at a spend cap you set. The model gets the same content and returns the same answer. You are billed for less of it.
On by default, with nothing to install in your code. Free to start, $20 per user a month on Pro.
Long agent sessions. A coding agent resends the whole conversation on every turn and re-reads the same files after every edit, so the repetition compounds as the session runs. The longer the job, the more Gate takes out of it. The three ways it saves
Compression and caching are free, so on the free plan the whole saving is yours. Pro is $20 per user a month and adds spend caps, per-team breakdown and the stronger compression modes. See pricing
The same content. Gate keeps the first copy of anything intact and replaces only the later copies, so nothing is summarized away and nothing is guessed at. The exact edit, in JSON
Agent traffic is repetitive. The same file is read again after every edit, the same command is run around a loop, and the same tool schemas ride along on every message. Providers bill all of it at full rate. Gate removes the repetition on the way out, serves what it can from its own cache, and stops a job at the cap you set.
A coding agent reads the same file after each edit. Every re-read is billed again at full input rate. Gate keeps the first copy intact and replaces the later ones with a pointer back to it, so the model can still see the content and you pay for it once.
Saves the largest share of the total on coding traffic.
A retry, a refresh, or a loop sends a request Gate has already answered. When the request matches one it has stored, Gate returns the stored response and never contacts the provider. You choose how long an answer stays valid.
Saves 100% of the tokens on a hit.
A batch job hits a failing call and retries all night. At the cap you set, Gate returns a refusal the agent can handle, so the loop exits instead of compounding until someone wakes up and finds the bill.
Saves the spend you never intended to make.
Put your monthly model spend in and see the saving at the rate Gate measures on agent traffic. Long agent sessions save the most, because they repeat the most.
An estimate, not a quote. 20% is the reduction the Pro plan measures on a typical agent request, the request drawn in Fig. 1 above.
Read the case study“As we add AI-powered tools, we need to understand both what those systems contribute and what they cost. Gate AI gives our engineering team a clearer view of the workflows we’ve connected and helps us identify opportunities to operate more efficiently without compromising the experience we provide to job seekers all while saving money!”
Here is one real request, before and after. Your coding agent sent the same file on turn 1 and again on turn 7. Gate sends the first copy in full, and replaces the second with a line pointing back to it. Everything else goes upstream untouched.
{
"role": "user",
"content": [{
"type": "tool_result",
"tool_use_id": "toolu_014c",
"content": "// src/app.ts
import { createServer } from './server'
import { loadConfig } from './config'
... 412 more lines ..."
}]
}{
"role": "user",
"content": [{
"type": "tool_result",
"tool_use_id": "toolu_014c",
"content": "// src/app.ts
[identical to the copy in
message 4 of this conversation]"
}]
}The model already read this file earlier in the same conversation, so it does not need a second copy to answer. Everything Gate did not touch goes upstream byte for byte.
tools[] is verbose prose with examples and warnings.
Gate minifies it without changing which tools the model can call.
Token prices keep falling and AI bills keep rising, because better models attract more work. An agent you trust does not answer one question and stop. It works through the whole task, and the teams that trust it run it far more often than they planned to. Cutting waste out of every request is what keeps that affordable.
Pay 80 cents for what cost a dollar and the same budget runs 1.25 times the work. On $40,000 a month that is $10,000 of inference you were not going to get.
From the first request, Gate reports what it removed and what its cache answered, priced against what you paid, and broken down by team, user and model across pay-as-you-go and your own provider keys.
Example view · figures are illustrative
Compression and caching are on the free plan, so if you stay on Free the saving costs you nothing. Pro is $20 per user per month and adds spend caps, per-team and per-model breakdown, and the advanced compression pipeline. Model usage is pay as you go through Gate, or bring your own provider keys and keep buying tokens directly. The estimator above subtracts the Pro seats you enter, so the figure it shows is net.
Both are, and they sit on the same record. Whenever compression actually changes a request, Gate stores the body your client sent alongside the body it forwarded upstream, so you can put them side by side rather than reconcile two systems. The forwarded body is what the model received, which is the one an auditor is asking about. Entries are hash-chained and anchored, so a gap in the chain is itself evidence.
Cost tools read the bill after it arrives. They report what you spent and who spent it, which is worth having. They cannot make the next request cheaper. Gate sits in the request path, so it removes tokens before they are billed and answers repeat calls without contacting the provider. The two compose: Gate lowers the bill, your cost tool still reports it.
Only traffic that routes through Gate. Anything your teams send straight to a provider is invisible to it. In practice that is the argument for pointing every tool at one endpoint: Gate reports spend by team, user, and model across pay-as-you-go and your own provider keys, which is hard to assemble from several provider consoles.
The default mode is the safe one. It removes content the model cannot use: terminal color codes, repeated lines, editor scaffolding, pretty-print spacing, and copies of content already earlier in the same conversation. The model still has that content, in full, at its first appearance. Your own prose is never rewritten.
Before anything is removed. Gate evaluates the request your client sent for prompt injection and for leaked credentials or PII, and only then does a savings stage touch it. Turning compression off does not turn the scanner off, and turning the scanner off is a separate decision.
No, and it is designed around that. A repeated block is only replaced once the model has already replied past it, so the prefix Gate forwards stays byte-identical on every later turn and your provider's cache keeps hitting. Gate also inserts Anthropic cache_control markers for clients that do not send their own.
No. Savings are on by default and there is nothing to install in your code. Compression is a single toggle on the Token Savings page in the dashboard, and the cache lifetime is set next to it.
Free to start, no card. Savings are on by default, and the dashboard shows what they were worth on your own traffic.