Every proxy provider publishes a price per gigabyte, every comparison table ranks on it, and on its own it is close to useless for choosing a network.
Nobody buys bandwidth for its own sake. You buy results: a product page with the price in it, a search results page with results on it, a profile that actually loaded. Bandwidth is only the unit the invoice happens to be written in. A network at $1.50/GB that fails a third of the time on your targets can cost more per result than one at $3.00/GB that rarely fails, and no rate card will ever show you that.
This guide works through what sits between the sticker price and the cost of a delivered result, in the order it usually matters: failed requests, the size of those failures, how success changes by target, how much of the plan you actually use, and the billing rules in the small print.
What a failed request costs you
When a request through a proxy fails, the bytes still move, and you still pay for them. Failures are not all alike, though, and the difference matters for the arithmetic:
- Challenge pages. A CAPTCHA or JavaScript challenge is a full HTTP response, often with scripts, fonts and images behind it. It can be as large as the page you wanted, sometimes larger.
- Soft blocks. A search page that loads with zero results, or a product page with the price missing, looks like a success to a status-code check. It is usually the full page weight, billed in full, and it delivers nothing.
- Hard blocks. A short 403 or 429 is small. It costs a request slot and some time, but few bytes.
- Timeouts and resets. These bill whatever was transferred before the connection died, which can be anything from a handshake to most of the page.
The expensive failures are the ones that look like successes until you check the content. That is why success has to be judged by what came back, not by the status code. The success rate guide covers how to write those checks.
The retry tax
If a share p of your requests succeed, you need on average 1 ÷ p attempts for every result you keep. The premium over the sticker price, if failures are as large as successes, is:
retry tax = (1 ÷ success rate) − 1
| Success rate | Attempts per result | Retry tax |
|---|---|---|
| 99% | 1.01 | 1% |
| 95% | 1.05 | 5% |
| 90% | 1.11 | 11% |
| 80% | 1.25 | 25% |
| 75% | 1.33 | 33% |
| 60% | 1.67 | 67% |
| 50% | 2.00 | 100% |
| 40% | 2.50 | 150% |
| 30% | 3.33 | 233% |
The curve is the point. Between 99% and 90% the tax barely moves. Below 80% it climbs fast, and below 50% the failures cost more than the results. Two networks that look close on success rate, say 92% and 84%, are further apart in cost than their rate cards suggest.
Failures are not the same size as successes
The simple formula assumes every failure weighs as much as a success. Real traffic rarely does. The more precise version counts bytes:
bytes per result = success size + ((1 − p) ÷ p) × failure size
Take a 450 KB page and a 60% success rate:
| What the failures look like | Bytes per result | Retry tax |
|---|---|---|
| Short 403s, about 20 KB | 463 KB | 3% |
| Full pages with no data, about 450 KB | 750 KB | 67% |
| Heavy challenge pages, about 900 KB | 1,050 KB | 133% |
Same success rate, three very different bills. This is why the same network can be cheap on one site and expensive on another: the target decides how it fails. A site that answers unwanted traffic with a quick 403 is cheap to fail on. A site that answers with a full challenge page is expensive to fail on, and the sites that do this tend to be the valuable ones.
Cost per successful request
Put the pieces together and the number worth comparing, for per-gigabyte billing, is:
cost per 1,000 results =
1,000 × bytes per result ÷ 1,048,576 KB per GB × price per GB
With failures the same size as successes, that reduces to the familiar form:
cost per 1,000 results =
(1,000 × payload in KB ÷ 1,048,576) × price per GB ÷ success rate
Every input has to come from measurement rather than from a sales page:
Payload. Measure it from your own traffic. The HTML of a page and the same page loaded in a real browser can differ by an order of magnitude, and if you drive a browser you pay for every image, script and font it pulls.
Price per GB. Use the rate for the plan you will actually be on, not the “from” price, which usually belongs to a tier most buyers never reach. Atlas uses each network’s 200 GB plan for exactly this reason: it is a volume real teams buy, and the networks it measures publish a price for it.
Success rate. Measured on your targets, with content checks. It is the input that changes the answer most often, and the one most often copied from a vendor’s homepage.
The cheapest network depends on the target
A single success rate per network hides the most useful part of the picture. Success varies far more between targets than between networks, so the ranking can flip as your mix of sites changes.
Here are three illustrative networks, priced on a 450 KB page:
| Network | Price per GB | Success, easy sites | Success, hard sites | Per 1,000, easy | Per 1,000, hard |
|---|---|---|---|---|---|
| A | $1.50 | 92% | 31% | $0.70 | $2.08 |
| B | $3.00 | 96% | 84% | $1.34 | $1.53 |
| C | $4.50 | 98% | 95% | $1.97 | $2.03 |
On easy sites, A costs about half as much as B. On hard sites, A is the most expensive of the three. The rate card ranks them A, B, C every time, and it is right only on easy targets.
Weight by your own traffic and the answer follows. With 30% of requests on hard sites, A still wins at $1.11 per 1,000 against $1.40 for B. Above roughly 54% hard traffic, B becomes the cheaper of the two. C costs the most at every mix here, but on hard sites it is only a third dearer than B, while its rate card is 50% higher and three times A’s.
Two practical consequences:
- Price your traffic, not an average. Split requests by target, measure success per target, and weight the cost by volume. A blended success rate quietly assumes your mix matches whoever made the blend.
- Watch how the mix moves. Teams add harder targets over time, rarely easier ones. The network that is cheapest today on easy work can become the expensive one a quarter later, without any change to its price.
The real price per GB is not on the rate card
Before a single request fails, the price per gigabyte you pay can already differ from the one you were quoted.
Utilisation. Most residential plans sell a monthly allowance, and traffic you do not use often expires at the end of the period. A $300 plan with 100 GB is $3.00/GB on paper. If you use 60 GB, you paid $5.00 for each gigabyte that did any work. Compare what you would pay at the volume you really use, and check whether unused traffic rolls over.
Minimum spends. Below a certain volume, the commitment rather than your usage is your bill. A network at half the rate with a $500 monthly floor costs more than one at full rate with no floor, if you spend $200 a month. Work out the bill at your actual volume, not at the volume where the discount looks best.
What counts as traffic. Networks generally bill bytes in both directions, and on small requests the request itself and its headers become a visible share of the total. If your payloads are small, how a network counts can matter more than a few cents on the rate. Compare the bandwidth on the provider’s dashboard with your own count for a day before trusting either.
Overage. Going past the allowance is often billed at a higher rate than the plan. If your usage is spiky, the effective price per GB is set by your busiest month, not your average one.
Other billing models
Per-gigabyte billing is the most common model for residential traffic, not the only one, and the same reasoning carries over:
- Per request. Payload size drops out, and the question becomes whether failed requests are billed. If they are, cost per result is
price per request ÷ success rate, and the retry tax applies in full whatever the page weighs. If only successes are billed, the network carries the failure cost, and you should expect to see that in its price. - Per IP or per port. Static and dedicated offers price access rather than traffic. Here cost per result depends on throughput: how many successful requests one address can deliver before a target starts refusing it. Compare results per address per day, not price per address.
- Unlimited plans. These move the limit from bytes to threads, speed or fair-use rules. Cost per result is the plan price divided by the results you can actually push through those limits. The unlimited proxies guide covers where those limits usually sit.
Time is a cost too
A failed request does not only cost bytes. It costs the time it took to fail, and the time to try again.
If a network answers in 1.5 seconds and succeeds 60% of the time, each result takes about 2.5 seconds of connection time on average. One that answers in 0.9 seconds and succeeds 95% of the time needs under a second. To deliver the same results per minute, the first needs about two and a half times the concurrency: more threads, more workers and, on per-port plans, more ports. The latency guide shows why the slow tail, not the median, sets this.
Then there is the engineering. A network that fails a quarter of the time needs retry logic, backoff, deduplication, dead-letter handling and someone to investigate when the failure pattern changes. That is a recurring cost in salary, and over a year it is often larger than the difference between two rate cards.
Putting it into one number
A monthly cost model that holds up in a budget review looks like this:
monthly cost = max(minimum spend, traffic used × effective price per GB)
traffic used = sum over targets of (results needed × bytes per result)
bytes per result = success size + ((1 − success rate) ÷ success rate) × failure size
Divide by the results you needed, add the engineering time each option requires, and compare that. Not the rate card.
A checklist before you commit
- Split your traffic by target and estimate the volume for each.
- Measure success per target with content checks, not status codes.
- Measure the size of successes and of failures from your own traffic.
- Price each candidate at the plan you will really use, including unused traffic and any minimum spend.
- Compute cost per 1,000 results for each target, then weight by your mix.
- Check how the ranking changes if hard targets grow.
- Estimate the retry handling and concurrency each option needs.
- Only then look at the rate card, as a sanity check rather than a ranking.
How Atlas calculates it
The true cost page publishes cost per successful request for every network Atlas measures, next to the rate card, so you can see exactly where the order changes. Each figure uses the network’s 200 GB plan price and the success rate Atlas measures across all of its targets with content checks, and it counts the retries a failed request needs, as your bill would. Minimum spends are left out, because they depend on your volume rather than on the network.
The methodology shows how each figure is calculated, and the full dataset behind Atlas downloads as a single file if you want to check the arithmetic or rerun it with your own traffic mix.