Skip to content
Glossary

What is bandwidth billing for proxies?

Definition

Paying a proxy provider for each gigabyte transferred, rather than per address or per month of access.

The meter counts bytes on the wire, in both directions, for every tunnel you open. It does not count the bytes you kept. Most surprise invoices come from the gap between those two numbers, and you can measure the gap on your own targets before you buy anything.

Per-gigabyte pricing is the common model for residential and mobile plans. How the other proxy types are priced, and how to size a plan from a sample of your own traffic, is covered in Residential vs datacenter vs ISP vs mobile: which to use. This entry is about what the meter sees.

What one request costs on the meter

To make the overhead visible we put a byte-counting CONNECT proxy between curl and five requests. It is the simplest meter there is: it adds up every byte that crosses each tunnel, up and down. Commercial meters may round, exclude some traffic or count differently, so treat this as the floor of what a per-GB plan can bill, and ask your provider how theirs works.

Bytes metered per request through a counting CONNECT proxy, curl 8.7.1 Drawn from the table below.
Bytes metered per request through a counting CONNECT proxy, curl 8.7.1. Bar chart of Bytes metered, up and down by Request.5 rows, drawn from the table that follows this chart.Wikipedia article, uncompressedWikipedia article, uncompressed: 311,522311,522Same article with –compressedSame article with –compressed: 68,72668,726HEAD request for the same articleHEAD request for the same article: 11,92611,926Etsy home page, answered with a 403block pageEtsy home page, answered with a 403 block page: 7,9097,909wikipedia.org redirect towww.wikipedia.org, two tunnelswikipedia.org redirect to www.wikipedia.org, two tunnels: 137,137137,1370200K400K bytesBytes metered per request through a counting CONNECT proxy, curl 8.7.1. Bar chart of Bytes metered, up and down by Request.5 rows, drawn from the table that follows this chart.Wikipedia article,uncompressedWikipedia article, uncompressed: 311,522311,522Same article with–compressedSame article with –compressed: 68,72668,726HEAD request for thesame articleHEAD request for the same article: 11,92611,926Etsy home page, answeredwith a 403 block pageEtsy home page, answered with a 403 block page: 7,9097,909wikipedia.org redirectto www.wikipedia.org, t…wikipedia.org redirect to www.wikipedia.org, two tunnels: 137,137137,1370200K400K bytes
Request Body bytes curl kept Bytes metered, up and down Metered per byte kept
Wikipedia article, uncompressed 298,919 311,522 1.04
Same article with –compressed 56,665 68,726 1.21
HEAD request for the same article 0 11,926 —
Etsy home page, answered with a 403 block page 776 7,909 10.2
wikipedia.org redirect to www.wikipedia.org, two tunnels 122,421 137,137 1.12

Three things stand out. Each new tunnel costs roughly 7 KB before any content for these hosts, because the TLS handshake, including the server’s certificate chain, crosses the proxy and is metered like everything else. That fixed cost is why a 776-byte block page metered at more than ten times its size, and why a job that fails fast across many hosts still burns traffic.

Compression is the largest single lever for a plain HTTP client. The same article metered 311,522 bytes without it and 68,726 with it, a factor of 4.5. curl does not ask for compression unless you pass --compressed; python-requests 2.32.3 asks for it by default.

A redirect to another host opens a second tunnel, so it pays the handshake twice. A redirect within the same host can reuse the open connection, so it costs only the extra response.

A browser multiplies it

A headless browser fetches every script, stylesheet, font and image a page asks for. We loaded three pages in headless Chrome 155 with the cache disabled and summed what the DevTools protocol reported as received. Wikipedia’s “Proxy server” article came to 493,843 bytes across 32 requests, about 8 times the 62,844 bytes of the HTML document itself. The BBC News front page came to 3,075,156 bytes across 182 requests, about 17 times its HTML. GitHub’s curl repository page came to 2,191,098 bytes across 201 requests, about 40 times.

The usual advice is to block images, fonts and media. It helps less than people expect, because scripts are now most of the weight. Blocking those three resource types cut the Wikipedia load by 22 percent, BBC News by 28 percent and the GitHub page by 3 percent. On BBC News, scripts alone were 1,976,075 bytes, close to two thirds of the page. The saving that matters is not rendering at all: find the JSON request the page makes and call it directly.

A fresh browser profile also talks to its vendor. Pointed at our meter to load example.com once, Chrome 155 opened 11 tunnels, and only 2 of them went to example.com. The other 9 went to Google update, account and service hosts and carried about 104 KB, six times the page itself. Launching with --disable-background-networking --disable-component-update cut that to 4 tunnels and about 36 KB, not to zero.

Questions to settle before you commit

Ask four things and get the answers in writing. Are failed, blocked and challenged responses billed? Are the handshake and headers counted, or only bodies? How is each request or session rounded? Does unused traffic expire at the end of the period? Then compare prices at the tier you will actually buy, because headline per-gigabyte prices are often quoted for the largest commitment.

Keeping the number down

  • Ask for compression and check it arrived. Look for a Content-Encoding header on the response, not just the one you sent.
  • Reuse connections. Every new tunnel pays the handshake again, so a client that opens a fresh connection per request pays it every time.
  • Revalidate instead of refetching. A conditional request answered with 304 carries no body. The measured difference is in Proxies for RAG pipelines.
  • Fail fast and cap retries. Every retry is a new request with its own bytes. How to use a proxy with curl shows the timeout flags and how to fetch headers only when that is all you need.

How these figures were measured. 21 October 2026, macOS, one consumer connection. The meter is a 26-line Python asyncio CONNECT proxy that counts bytes in each direction per tunnel. curl 8.7.1 made each request three times and the table shows the median; the redirect row has two runs, because on the third wikipedia.org answered 429 instead of redirecting. Browser figures come from Chrome 155.0.8059.85 in headless mode with the cache disabled, summing encodedDataLength from the DevTools protocol. Five full loads of each page agreed within 3 percent, except one GitHub load that made 42 fewer requests and came in 21 percent lower. The background-traffic test was run twice with default flags and twice with the two flags shown, each from a new profile.

Frequently asked questions

Does a failed request still use proxy bandwidth?
Usually. A blocked response is still a response, and the TLS handshake before it crosses the proxy too. In our metering a 776-byte block page cost 7,909 bytes. Ask the provider whether it bills failures and get the answer in writing.
Why is my usage so much higher than the pages I save?
Handshakes, headers, redirects and retries are all counted, and a browser fetches every asset a page asks for. Headless Chrome loading GitHub's curl repository page transferred about 40 times the size of its HTML.
Does blocking images in a headless browser save much bandwidth?
Less than often claimed. Blocking images, fonts and media cut three test pages by 3 to 28 percent, because scripts were most of the weight. Calling the page's own JSON endpoint without rendering saves far more.

Sources

  1. RFC 9110: HTTP Semantics, the CONNECT method
  2. RFC 9110: HTTP Semantics, content codings
  3. RFC 8446: TLS 1.3, the Certificate message sent during the handshake
  4. Chrome DevTools Protocol: Network domain, loadingFinished and encodedDataLength

All terms, A to Z