Proxy Pricing Explained: Per GB, Per IP, Per Request and What You Really Pay

On 27 July 2026, Webshare listed 100 datacenter proxies at $2.99 a month while Bright Data listed residential bandwidth at $8 per gigabyte. Both are current list prices from the same market, and side by side they tell you almost nothing, because one meters addresses and the other meters bytes. The market prices in four separate units: per gigabyte transferred, per IP per month, per thousand successful requests, and a flat fee with no volume meter at all. Published comparison tables, including Proxyway's provider rankings, reprint those headline numbers in one shared column, which makes a $1.00/GB vendor and a $2.99/month vendor look like two points on the same scale.

This guide is the arithmetic that fixes that. It converts every pricing model into one comparable figure, cost per successfully scraped page, then covers the commercial terms that move that figure more than the headline rate does: minimum commits, bandwidth expiry, overage rates and what changes when you leave self-serve for a contract.

There are no speed or block-rate rankings here. We run no first-party benchmarks and would not ask you to trust ours if we did. There is also no cost-per-page figure for your workload, because that number depends on your page weight and your success rate, and only your logs know those. You get the formula and the list of inputs to collect for it.

Every price below was read from the vendor's own live pricing page on 27 July 2026 unless stated otherwise.

The four units the market actually prices in

Every vendor in the proxy networks category prices in one of four ways, and several price in three of them across different products.

Per gigabyte transferred. The dominant model for rotating residential. You pay for bytes over the wire, in both directions, including headers and TLS overhead. DataImpulse lists $1/GB residential, $0.50/GB datacenter and $2/GB mobile, with a $5 entry offer for new accounts and a pricing table stating traffic never expires. PacketStream lists $1 per GB with a $50 minimum purchase, $500 for resellers, and states that purchased balance does not expire. Evomi lists residential from $0.49/GB, though that rate carries a 100 GB monthly minimum at $49.99, alongside datacenter from $0.35/GB, mobile from $3.75/GB and static ISP from $1.00 per IP.

Per IP per month. The dominant model for datacenter and static residential. You rent addresses, not bytes. Webshare is the clearest example: $0.0299 per proxy per month at 100 datacenter proxies, falling to $0.0179 at 60,000, plus static residential at $0.30 per IP falling to $0.225 at 10,000, and a free tier of 10 proxies. Webshare also sells rotating residential per GB, from $3.50/GB at 1 GB down to $1.40/GB at 3,000 GB, so a single vendor spans two units.

Per thousand successful requests. The model unblocking APIs use, and the only one where the vendor absorbs your retry cost. Bright Data's Web Unlocker and both major vendors' SERP products sit here. This unit is already close to what you want, which is why unblocking APIs are easier to budget than raw proxies even when they look more expensive per unit.

Flat monthly fee, no volume meter. Geonode lists unlimited residential proxies at $1,800 per month. Volume stops being the scarce resource and your own throughput becomes the constraint.

Here is the same set with the unit made explicit, which is the column comparison tables usually collapse.

Three things in that table matter more than any individual number in it.

First, the range on residential is roughly 16x between the cheapest published floor and the most expensive published pay-as-you-go rate. That is not a quality signal. It is a difference in what the rate includes, which pool you are routed into, and how much sales cost is amortised into the price.

Second, "from $X/GB" is frequently not a rate you can buy. IPRoyal advertises from $1.75/GB and lists its actual subscription tiers at $7.00/GB for 1 GB, $5.95 for 2 GB, $5.25 for 10 GB and $4.90 for 50 GB, with pay-as-you-go between $7.35 and $5.15. The $1.75 figure describes bulk purchase far above those tiers. Evomi's $0.49/GB similarly requires a 100 GB monthly commitment. Compare the tier you will buy, not the footnote.

Third, Bright Data's page on 27 July 2026 displayed every rate twice: a struck-through list price and a discounted price under a 50% promotional code. Budgeting against a promotional rate that can lapse is how a forecast quietly doubles. The Bright Data versus Oxylabs comparison covers how the two differ beyond price, and Oxylabs is the sharper case of the same problem: its pricing page names Starter, Advanced, Premium, Venture, Business and Corporate plans without publishing what any of them cost, so a firm number requires entering the sales funnel. That is itself a pricing fact worth recording.

The only number that compares: cost per successful page

Normalise everything to this.

cost per successful page = billed units per successful page x unit rate

For a per-GB vendor, "billed units" is bytes, and bytes per successful page has two components most buyers forget to add together:

bytes billed per success = B_success + ((1 / s) - 1) x B_failure

where s is your success rate and B_failure is how many bytes a failed attempt transfers before you know it failed. That second term separates cheap failures from expensive ones. A 403 with a short body costs almost nothing. A full headless render that completes, paints, and only then reveals a bot-detection interstitial costs you the entire page weight for zero data.

You need three inputs, and all three come from your own instrumentation rather than any vendor's marketing:

1. Bytes per successful fetch. Read this from your proxy dashboard or your HTTP client, not from an estimate. Include response headers and any redirect chain.
2. Success rate on the specific targets you scrape. Not the vendor's published network success rate, which is measured against a different target mix and travels badly to yours.
3. Bytes per failed fetch. Usually much smaller than a success, unless you render.

Page weight moves the answer more than the rate card does

Hold the rate constant at $3.00 per GB and vary only how you fetch. One gigabyte is 1,048,576 KB, so pages per GB is that divided by your transfer size.

Those transfer figures are not invented. HTTP Archive's Web Almanac 2025 page-weight chapter reports a median page of 2,164 KB on mobile and 2,412 KB on desktop, a 10th percentile of 516 KB and 1,127 KB at the 25th, and a 90th percentile of 8,337 KB on mobile and 9,179 KB on desktop. Home pages specifically are heavier: the chapter puts the median mobile home page at 2.6 MB, up 8.4% from 2.4 MB in 2024, and the median desktop home page at 2.9 MB, up 7.3% from 2.7 MB.

The spread between the lightest and heaviest row above is roughly 70x. The spread between the cheapest and most expensive residential rate card in this guide is roughly 16x. Your fetch strategy is a bigger lever than your vendor choice, and it is a lever you control unilaterally. Blocking images, fonts, media and third-party analytics in your browser infrastructure layer will cut more from a proxy bill than any negotiation will.

This also flips rankings. At 120 KB per page, the gap between DataImpulse at $1/GB and Decodo's 100 GB tier at $2.75/GB is $0.20 per thousand pages, which is noise against your engineering time. At 2,164 KB per page the same gap is $3.61 per thousand, which at ten million pages a month is a real line item. Cheap per-GB vendors only start to matter once your pages are heavy. If you scrape lean HTML, choose on pool quality and success rate and stop reading price tables.

Where retries land, and why they land differently

Apply the retry term. Assume a 15 KB block page and an 80% success rate, so (1/s) - 1 = 0.25.

For the HTML-only profile: 120 + 0.25 x 15 = 123.75 KB billed per success. Retries add 3%. Negligible.

For the full-load profile, where a failure costs a complete page load: 2,164 + 0.25 x 2,164 = 2,705 KB. Retries add 25%.

So on per-GB pricing, retries cost you in proportion to how far you got before you knew you had failed. Detecting a block from the status line is nearly free. Detecting it from rendered DOM content is expensive. If you run a rendering pipeline, a cheap pre-flight check that catches obvious blocks before the browser commits to a full load is a bandwidth optimisation, not just a latency one. The mechanics of which signals fire when are covered in why scrapers get blocked.

This is also the honest case for per-request unblocking APIs. When a vendor bills per successful response, your retry cost is contractually theirs. Whether that is cheaper than raw proxies depends on the same page-weight arithmetic, and the credit systems those products use have their own traps, which scraping API credit pricing unpacks. But on a target where your success rate is 40% and volatile, forecastability is worth paying for even at a worse unit rate.

Per-IP pricing is a bet on a different constraint

Per-IP pricing takes bytes off the bill entirely, so no amount of arithmetic converts it into a $/GB figure. You have to go all the way to pages:

pages per month = IPs x requests per IP per day before throttling x 30 x success rate
cost per 1,000 pages = (monthly price x 1,000) / pages per month

Plug placeholder numbers into Webshare's $2.99 for 100 datacenter proxies. If each IP sustains 500 requests per day against your target before rate limiting, at an 80% success rate, that is 1.2 million successful pages a month for $2.99. The figures 500 and 80% are placeholders. Substitute your own, and distrust anyone who publishes them as universal constants.

Even with pessimistic substitutions the result lands a different order of magnitude from any per-GB residential quote. That makes the choice between per-IP and per-GB a question about what your targets tolerate rather than a question about price. If your target does not block datacenter ASNs, per-IP is so much cheaper that residential pricing is irrelevant. If it does, per-IP is infinitely expensive, because cost per successful page is undefined at a zero success rate. Choosing a proxy type by target is the prior question to all of this.

Webshare's free tier of 10 proxies exists precisely to let you answer that question before spending anything. Use it as a target-tolerance probe, not as a production allocation.

Where per-IP is wrong: any workload needing broad geographic diversity or high IP churn. Renting 100 static addresses gives you 100 fingerprints. A rotating residential pool gives you a fresh one per request. For account-adjacent work or aggressively fingerprinted commerce targets, the per-IP saving evaporates on the first ban wave.

Flat pricing bills your utilisation instead of your volume

Geonode lists unlimited residential at $1,800 per month with no volume meter. The normalisation is utilisation-bound rather than byte-bound:

cost per 1,000 pages = (monthly price x 1,000) / pages you actually complete that month

Run 100 million pages through it and the cost is $0.018 per thousand, which no metered vendor approaches. Run 2 million and it is $0.90 per thousand, which is competitive with cheap per-GB residential only if your pages are heavy. Run 200,000 and it is $9.00 per thousand, which is worse than almost everything else in this guide.

The break-even against a $1/GB vendor is where pages x KB per page / 1,048,576 = 1,800, roughly 1.89 million GB-equivalents. At 2,164 KB per page that is about 872,000 pages a month. Below that, meter. Above it, flat.

"Unlimited" is also a statement about the volume meter and nothing else. It says nothing about achievable throughput, concurrency limits, or which pool you are routed into. We could not reach Geonode's documentation for the unlimited product on 27 July 2026 to confirm the country coverage or any speed cap, so treat those as open questions to ask in writing before committing.

Flat pricing is wrong for anyone running a few hours a week. If your crawl is a nightly two-hour job, you pay 24/7 for 8% utilisation, and a metered vendor with a non-expiring balance will be cheaper by a wide margin.

The commercial terms nobody puts in a comparison table

Four contract terms move your effective price more than the headline rate does, and vendors put none of them in the column you are comparing.

Utilisation is a hidden multiplier on every committed plan

effective $/GB = plan price / GB actually consumed

Buy Decodo's 100 GB tier at $2.75/GB, spend $275, and consume 60 GB. Your effective rate is $4.58/GB. That is above Decodo's own $4.00/GB pay-as-you-go rate. You paid a premium to commit. Nothing on the invoice will tell you that, and this is the correction buyers most often skip.

Tier discounts are only real if your volume is predictable. Volume that swings 3x between months is structurally mismatched with committed tiers, and you should be looking at pay-as-you-go or non-expiring balances even at a worse sticker rate.

Whether unused bandwidth expires

Non-expiring balances collapse the utilisation problem to zero. Unused gigabytes are still yours next month, so effective rate equals sticker rate by definition.

Three vendors here publish that term explicitly, all checked 27 July 2026. PacketStream's pricing page states that purchased balance does not expire and that "nothing renews, nothing expires". DataImpulse's pricing table states that traffic never expires. IPRoyal states that traffic never expires and pitches bulk purchase on exactly that basis.

Subscription-tier vendors more typically reset the allocation at the end of the cycle. Decodo's residential pricing page did not state a rollover policy either way on the date checked, which is the common case. Where a vendor does not state the policy on the pricing page, assume it expires and get the answer in writing before you commit.

Overage is priced at the rate you were trying to avoid

The standard structure: exceed your tier, get billed the difference at pay-as-you-go. Pay-as-you-go is, on almost every rate card in this market, the highest number the vendor publishes. Decodo is the clean documented example. Its pay-as-you-go rate of $4.00/GB sits above every one of its monthly tiers, which run from $3.75/GB down to $2.75/GB. Bright Data shows the same ordering: $8/GB list pay-as-you-go against $7, $6 and $5/GB list rates on its three committed tiers.

The practical consequence is that a plan sized to your average month is a trap if your traffic is spiky. Sizing to your p75 month and eating some unused allocation is often cheaper than sizing to the median and paying overage in the peaks. Run that comparison against your actual monthly distribution before you pick a tier. If your peaks come from a single vendor failing rather than from real demand, failover across two vendors is the cheaper fix than a bigger commit.

Moving to a contract changes more than the rate

When you move above self-serve tiers, four things typically become negotiable: the per-unit rate, the commit length, the rollover policy and the overage rate. Buyers spend all their leverage on the first, and the other three usually matter more.

A modestly worse rate with rollover and an overage rate capped at the tier price will beat a headline discount with expiring traffic and pay-as-you-go overage for any workload with variable volume.

Ask specifically about targeting surcharges. Country, city and ASN-level targeting are frequently priced as add-ons rather than included in the base rate, and a base rate that excludes the targeting you actually need is not the rate you will pay. We did not find a published multiplier schedule for any vendor in this set, which is itself the point: get it quoted in writing.

Also read whether prices exclude tax. Several European vendors quote plus VAT, a silent 20%-ish adjustment on a comparison you thought you had finished. And check whether the rate on screen is promotional. Bright Data's page on 27 July 2026 showed every residential rate at 50% off a struck-through list price under a coupon code. Forecast against the list price, not the promotion.

Choosing a model, and who each one is wrong for

Lean HTML at high volume. Cost per page is already low, so optimise for pool quality rather than rate. DataImpulse and PacketStream both list $1/GB with non-expiring balances, which makes them low-risk starting points. Wrong for you if you need granular geo or ASN targeting at scale, where surcharge structure matters more than base rate, or if you need a procurement-grade SLA and a named account manager.

Rendered pages at high volume. Your problem is bytes, not rate. Block subresources first. Then look at per-request unblocking products, where the vendor absorbs retry cost. Wrong for you if your rendering needs are unusual enough that a managed unblocker cannot reproduce them, in which case you are back to raw proxies plus your own browser infrastructure.

Targets that tolerate datacenter IPs. Webshare's per-IP datacenter pricing is a different cost class from anything metered, and the free tier lets you test the assumption for nothing. Wrong for you the moment the target starts fingerprinting ASN, and wrong for geo-diverse work where a fixed set of addresses cannot cover your locales.

Spiky or seasonal volume. Non-expiring balances at PacketStream, DataImpulse or IPRoyal beat tier discounts you cannot consume. Wrong for you if your volume is steady and large, where a committed tier at Decodo or a negotiated contract will beat pay-as-you-go on arithmetic alone.

Sustained, continuous, high-throughput crawls above roughly 900,000 heavy pages a month. Geonode's flat $1,800 removes the volume meter and makes the bill predictable. Wrong for you if you run in short bursts, if your volume sits below the break-even, or if you need country coverage guarantees you cannot get in writing.

Small self-serve budgets on residential. Evomi's $0.49/GB floor is the lowest published residential rate here, but it comes with a 100 GB monthly minimum, so the entry cost is $49.99 rather than pocket change. Wrong for you if you need 5 GB a month, where IPRoyal's $7.00/GB single-gigabyte tier or a per-request API costs less in absolute terms despite the worse unit rate.

Hard targets where failure is the dominant cost. Bright Data and Oxylabs are where teams end up when success rate matters more than unit price, and both are structurally built for that buyer. Wrong for you if you need a firm number today without a sales call, which Oxylabs' unpriced self-serve tiers effectively rule out, or if your budget cannot absorb a four-figure monthly commit while you are still validating whether the data is worth collecting.

A checklist for normalising any quote

Before you compare two vendors, collect these. Most are one line in your logs.

1. Median and p90 transfer size per successful fetch, on your real targets.
2. Transfer size per failed fetch, which is a different distribution.
3. Success rate per target, not per vendor.
4. Monthly volume distribution across the last six months, not the average.
5. Whether the quoted rate is promotional and when the promotion ends.
6. Whether the quoted rate includes tax.
7. Whether unused allocation rolls over, in writing.
8. The overage rate, in writing.
9. Which targeting features carry a multiplier and what the multiplier is.
10. The minimum commit and the minimum contract length, separately, since they are not the same thing.
11. What "unlimited" is unlimited in.

Run two candidates through cost per successful page with your own numbers, then sanity-check the answer against a small real job on both networks before signing.

If you are earlier than this and still deciding whether you need a proxy network at all rather than a web scraping API, start with how to choose a web scraping tool. For how the proxy layer sits underneath everything else, the 2026 web scraping stack covers the surrounding pieces.

Proxy pricing in this market changes often, promotional rates lapse, and rate cards get restructured without notice. Every figure above is dated 27 July 2026. Verify against the vendor before you commit budget.