Per-GB pricing stopped describing what you actually pay for

On 17 February 2026, Rayobyte cut its mobile proxy rates by up to 98 percent. The old Starter plan was $50 for 2 GB, or $25/GB. The replacement is $250 for 500 GB, or $0.50/GB. Adam Dubois, reporting it for Proxyway, wrote that the cut "sounds absurd but seems legitimate," on the grounds that the old baseline was very expensive to begin with and that Rayobyte had left its mobile product dormant for several years before rebooting it.

On 15 April 2026 the same outlet covered NodeMaven launching a cashback programme alongside a proxy quality guarantee that pays a customer $1 each time a proxy fails to perform as expected. On 7 July 2026 it covered SOAX rebooting its site, product line and pricing at once.

Those are three unconnected decisions by three companies, and they point the same way. The headline dollar-per-gigabyte figure has stopped carrying the information a buyer needs.

Block rate is the missing number

A residential proxy sells you bytes, and finance gets a bill denominated in bytes. The unit of work, though, is a successful extraction, and what connects the two is something no price page publishes: how often a request comes back usable.

A blocked response still moves bytes. A CAPTCHA interstitial moves more of them than the page you wanted, because challenge pages ship a lot of JavaScript. A soft block that returns HTTP 200 with an empty product grid moves the full weight of a rendered page and yields nothing. If a pool at half the listed rate needs three attempts to land one good record where the incumbent needed one and a bit, the cheaper pool costs more, and nothing on the price page that sold it to you would have said so.

That is the ordinary condition of any target running an anti-bot system, which is most targets worth scraping. The mechanics are covered in our guide on why scrapers get blocked. What matters for the invoice is narrower: block rate multiplies everything else on it, and no vendor quotes a block rate.

Four units that do not convert

Even before retries enter the picture, the market has fragmented to where two quotes cannot be compared without doing arithmetic on your own traffic first.

Webshare's pricing page, checked on 27 July 2026, lists rotating residential from $1.4/GB, static residential from $0.23 per IP per month, and its datacenter proxy server from $0.018 per IP per month. That is two incompatible units on one page. Converting the per-IP figures into anything comparable requires knowing how many gigabytes you will push through each IP, which depends on your concurrency and session length rather than on anything the vendor controls.

Decodo's residential pricing page, checked the same day, shows a different gap. The headline figure is $2/GB. The cheapest subscription actually listed works out to $2.75/GB at 100 GB, and pay-as-you-go is $4.00/GB. Every number there is honest. The headline is a floor reachable at volume, not a rate anyone starting out will pay, and the spread between the advertised number and the entry number is 2x.

SOAX's post-reboot page, also checked on 27 July 2026, makes the per-GB rate a matrix. Bandwidth is priced by country tier and separately by commitment level, across a free sandbox and four paid plans from $200 to $3,000 per month. Tier-one traffic (the page groups the US, UK, Australia and Japan here) is listed at $5.00/GB on the sandbox and falls through $3.00, $2.20 and $1.50 to $0.85/GB at the enterprise plan. Tier three runs $2.00/GB down to $0.35/GB across the same steps. One rate applies to residential and mobile alike, credits roll over 60 days on monthly billing and 365 days on annual, and datacenter proxies are gone from the lineup. There is no single number to put in a comparison column, and that looks deliberate rather than evasive.

Then there are vendors whose whole proposition is that the unit stays still. PacketStream advertises residential bandwidth at $1.00 per GB "metered at the gateway," with a $50 minimum, no subscription, and a balance that does not expire. DataImpulse publishes residential at $1/GB, datacenter at $0.50/GB and mobile at $2/GB, pay-as-you-go, with traffic that never expires and a 1 TB tier at $0.8/GB residential and $1.6/GB mobile. Evomi leads with residential from $0.49/GB, which its page calls the lowest price on the market, while listing mobile from $4.00/GB on the same table. Within one vendor, the spread between the cheapest and dearest product is roughly eight times. All of these are vendor-published figures read on 27 July 2026 and they will drift.

At the far end, some products dissolve bandwidth into an abstraction. Scrapeless bills in Consumption Units drawn down from a prepaid balance, covering browser and crawl, AI scraper, web unlocker, proxies and scraping API through one meter, with the CU-to-dollar rate improving as your package grows and unused funds forfeited in any month you fall below the minimum commitment. The same pattern runs through managed scraping products generally, where the unit is a credit whose cost varies with rendering and geotargeting, as covered in scraping API credits decoded.

Per gigabyte, per IP per month, per prepaid unit, per credit. Four units, no exchange rate, and vendors moving between them faster than procurement cycles turn over.

Normalising against your own logs

The only unit that survives all of this is cost per successful extraction, and its useful property is that you can compute it from data you already have. Nobody has to publish a benchmark for it to work.

Take one full billing period, per vendor, per target class. Divide total invoiced spend on that vendor by the count of records that passed your own validator in the same period. Not requests issued. Not HTTP 200s. Records that a downstream check confirmed were the thing you wanted: price field present, product ID matching, row count inside the range you expect. Soft blocks return 200 constantly, so a status-code denominator flatters every vendor equally and tells you nothing.

A few things keep the number honest.

Use whole billing periods. Per-IP products cannot be divided into per-request costs without knowing throughput per IP, and throughput per IP moves with concurrency and session length. A monthly boundary sidesteps the conversion entirely.

Segment by target class rather than by vendor alone. A pool that is strong on retail and weak on travel averages into meaninglessness. Group targets by how hard they defend themselves, roughly along the lines in picking a proxy type by target.

Run the comparison concurrently. Two vendors measured in different months are scored on different site behaviour, not on their own merits. Splitting live traffic is the only way to hold difficulty constant, and it falls out of the dual-vendor setup described in running two proxy vendors.

Count retry bandwidth against the vendor that caused it. If your client retries three times before giving up, all three requests belong in that vendor's spend. A pool that rotates aggressively on failure can burn a lot of bytes producing nothing.

None of this depends on trusting a published rate. It depends on your pipeline already knowing the difference between a fetched page and a good record, which most pipelines do.

Cashback and guarantees are the same admission

Read NodeMaven's April announcement as market signal rather than promotion and it says something specific. The vendor is conceding that listed bandwidth price does not describe delivered value, and offering to true up the difference in arrears.

Proxyway is the source for the details. Cashback starts at 5 percent of monthly traffic and reaches 10 percent at 500 GB or more, paid as extra traffic that never expires. The quality guarantee pays $1 per underperforming proxy, adjudicated by the PixelScan IP checker, capped at 20 percent of monthly usage and paid in traffic rather than cash. Claiming is involved: contact support, send a screenshot, wait for manual review. Dubois flagged two caveats worth carrying rather than smoothing over. NodeMaven's filter only works with sticky sessions, and PixelScan may be connected to the provider through Multilogin, which he wrote risks introducing bias in the adjudication. Those are reported concerns, not settled findings.

Whatever you make of the mechanism, the shape of it is a vendor putting money behind quality because price alone stopped differentiating. Expect more of these, and expect the interesting variation to be in who adjudicates and where the cap sits.

Get the definition of success in writing

Measuring cost per validated record only pays off if it changes a contract. Two asks are worth making of any provider large enough to negotiate with.

Ask for success-based billing, or failing that, a bandwidth credit policy for failed requests. Vendors resist the first because it transfers block-rate risk onto them, which is exactly why it is worth asking. The credit policy is the fallback and is often achievable.

Then get the definition of success in writing, because that is where the value of such a clause lives. Four terms decide it. What counts as a failure: a connection error only, a specific status code, or a body that fails a named validator. Who adjudicates, and whether that party is independent of the vendor. What the compensation cap is, and whether it pays in cash or in traffic that expires. And what counts as a billable byte, including headers, TLS handshakes on connections that never complete, and retries you did not initiate.

What to do about it

For a team collecting web data, a lower advertised rate is a hypothesis rather than a saving.

Stop tracking dollars per gigabyte as a procurement metric and start tracking dollars per validated record, computed per target class from your own invoices and your own validator counts. Instrument it before your next renewal, not after, because the number means nothing without a full billing period behind it.

When a vendor announces a cut, treat it as a reason to run a split test rather than a reason to switch. Rayobyte's mobile move is a real change in the market and it still does not tell you what a mobile IP will cost per usable page on your targets. Only your logs know that.

When the contract comes round, spend the negotiating capital on the failure clause rather than the headline rate. A tenth of a cent per gigabyte is worth less than a written definition of a failed request. The vendors now offering guarantees have already worked this out, which is the most useful thing their marketing has said this year.

For vendor coverage, the proxy networks category lists the providers on this site, proxy pricing explained goes deeper on the billing models, and the web scraping stack in 2026 sets out where the proxy layer sits relative to everything else.