A Buyer's Guide to the Leading Providers
Proxy 2026
A Buyer's Guide to the Leading Providers
Seven providers publish prices for the same five products in five incompatible formats. This report reads all of them on one day and puts them on one axis.
What residential, mobile, ISP and datacenter proxy actually cost and cover in 2026, compared across the providers data teams already shortlist — so the choice comes down to your use case, not their marketing.
2026
What's inside
Data and platform engineers choosing or replacing proxy infrastructure, and the people who approve the invoice.
Page 18 before any price table. The headline rates in this market are not comparable, and that page explains why.
DoubleData Proxy is a row like any other, marked in mint so you can weigh or discount it.
The market in
eight numbers
Each figure is sourced and dated in the pages that follow. Together they describe a market that got cheaper, got funded, and got the attention of law enforcement.
IPv4 scarcity is not pushing datacenter proxy prices up. The only public IPv4 price index shows a multi-year decline to decade lows, and proxy prices fell alongside it. Agentic search is not a documented driver of proxy demand. No company in that category publishes anything about proxy consumption, and the largest of them claim their own crawl infrastructure. Both are set out on pages 10 and 11.
Competition, not scarcity, explains the price moves. And the demand that vendors can actually evidence comes from AI labs and enterprise agent deployments, not from the tooling layer.
Who sells what
Product availability as shown on each vendor's own site on 24 August 2026. A gap means the category is not sold as a distinct purchasable product — not that the capability is absent.
shared
dedicated
Mint marks the publisher's own row. IPRoyal's shared datacenter gap is a stated product position: its site argues that "the best datacenter proxy services provide dedicated IPs". Evomi expresses dedication through its Static Residential variants rather than a separate datacenter SKU.
As of 24 August 2026 Bright Data's site lists three proxy products — Residential, ISP and Datacenter — and has no standalone mobile proxy product page or pricing. Its proxy pricing page still advertises "7M+ Mobile IPs" and its FAQ still lists Mobile among its proxy types. Bright Data has published no statement about discontinuing mobile proxies, and we make no claim that it has.
DoubleData Proxy sells all five categories, which is why its row is full — a fact about product range, not a recommendation. Where it lands on price, in full: cheapest of four on datacenter billed per GB, second of four on datacenter per IP, mid-pack on ISP and on residential, and dearest of the five that sell mobile. The tables from page 19 carry all five. Weigh them knowing who published the document.
Residential rates,
side by side
Every provider's published residential ladder, from its entry tier to the best rate it publishes. The width of each bar is the discount a buyer has to grow into.
IPRoyal, 4.0×. $7.00/GB at one gigabyte down to $1.75/GB — but the cheap end requires ten terabytes and is quoted as a custom rate.
Evomi Premium, 1.2×. $3.98 to $3.40/GB across four tiers, and nothing published above fifty gigabytes at all.
Around 10 GB is the only volume where nearly everyone publishes a rate. Above 100 GB three of the seven publish nothing.
Five things that
changed in 2026
Residential prices fell for two years, then turned
Proxyway's 2026 research, covering 13 providers and 3.6 million requests, reports that "between 2023 and 2025, the rates of residential proxies contracted by up to 75%", reversing only in early 2026 as providers withdrew discounts. Proxidize's index states the cause directly: "demand from AI applications hasn't increased prices — instead, competition flooded the market with new suppliers".
Institutional capital arrived
Oxylabs announced $130 million from Warburg Pincus at a $3.6 billion valuation on 9 July 2026 — its first outside investment, through the Warburg Pincus Capital Solutions Founders Fund. No stake size was disclosed by any party, and it is an investment rather than an acquisition. Bright Data separately reported over $300 million ARR growing 50% year over year, unaudited.
Law enforcement reached the infrastructure layer
On 2 July 2026 the FBI seized domains associated with NetNut, a residential proxy service operated by NASDAQ-listed Alarum Technologies, coordinated with Google, Lumen and Shadowserver. Alarum confirmed the seizures, paused services and cut roughly a third of its workforce. No criminal charges have been announced.
Automated traffic overtook human traffic
Cloudflare reported bots at 57.5% against humans at 42.5% of HTTP requests to HTML content in June 2026, with training-purpose crawlers rising from 22% to 52% of crawler activity in about a year. It has moved to permission-based access by default for new domains and set 15 September 2026 for mixed-use crawler blocking on ad-monetised pages for new and free-tier customers.
Proxy providers became defendants, not just suppliers
Scraping litigation has historically named the party doing the collecting. In Reddit, Inc. v. Perplexity AI, SerpApi, Oxylabs UAB and AWMProxy, filed in the Southern District of New York in October 2025, proxy and SERP providers are named as co-defendants rather than as suppliers. The case is at an early stage; nothing has been decided on its merits.
How this was put
together
Oxylabs, Bright Data, IPRoyal, Decodo, Webshare, Evomi and DoubleData Proxy. Selected for consistent visibility on buyer shortlists and for publishing enough pricing openly to be compared at all. A selection, not a ranking, and not a claim about market share.
Pricing and specifications from each vendor's own public pages, read on 24 August 2026. Reputation from G2, Trustpilot and Capterra. Market context from court records, regulator texts, CDN and security-vendor telemetry, and company announcements.
On Proxyway. Affiliate-ranked "best proxy" listicles were not used as a source of fact. Dated benchmark research was — principally Proxyway, which also carries affiliate links; every such citation names its publisher and date.
This report is published by DoubleData Proxy, and DoubleData Proxy is one of the seven providers compared in it. Its row is highlighted in mint in every table and grid so you can see it, weigh it, or discount it. Its rates were read from its own public pages on the same date and by the same method as everyone else's, and it receives no favourable treatment in the commentary.
Two of the seven share an owner. Oxylabs acquired Webshare in September 2022; Webshare was stated at the time to "continue to operate as an independent entity". Readers treating them as independent data points should adjust.
No success-rate, latency or throughput benchmarking — including for the publisher's own product. No total-cost or cost-per-record modelling; that needs volume commitments and request weightings we do not have. No assessment of IP pool quality. No ranking or winner in any category.
Every specific statement about a provider is sourced to that provider's own published material, to a court or regulator record, or to an attributed and dated third-party publication. Where reviewers make allegations we could not verify, we report the theme and not the allegation. Where a vendor's own pages contradict each other we show both figures. Nothing here is an assertion of wrongdoing by any company named.
Prices move in weeks. Every rate carries its access date; two vendors ran promotions that day and a third labels every mobile tier with a permanent-looking discount. Deeper tiers are gated. Several providers publish only a "from" rate and put the ladder behind registration; we say so rather than estimating it. Practitioner forums are absent. Reputation draws on review platforms only, and forum sentiment here runs more critical.
The market
Competition, not scarcity, moved prices
The IPv4 story most of the market repeats is contradicted by the only public price index.
The demand story is not the one usually told
No agentic-search company publishes anything about its proxy consumption.
Enforcement reached the infrastructure layer
A NASDAQ-listed provider had its domains seized in July 2026.
Your television may be part of the supply
Two platform owners banned proxy SDKs within six weeks of each other.
Twelve months that
reshaped the market
Twelve months to August 2026: enforcement, consolidation, a category exit by the largest vendor, and a chief executive moving between two of the companies in this report. Every entry is sourced and dated in the pages that follow.
Really Simple Licensing 1.0 launches — a machine-readable licensing protocol for crawlers, backed at launch by Reddit, Yahoo, Ziff Davis, O'Reilly, Medium and Fastly.
Reddit sues Perplexity, SerpApi, Oxylabs and AWMProxy in the Southern District of New York. The first time proxy providers are named as co-defendants rather than as suppliers.
Cloudflare suffers a global outage, followed by a second on 20 February 2026 — a reminder that one company is now a single point of failure for both the targets and the defences.
Google disrupts IPIDEA, describing thirteen ostensibly independent proxy brands under common control and 550+ distinct threat groups using its exit nodes in a single week.
Nebius agrees to acquire Tavily for up to $400m, folding an agentic-search layer into an AI cloud. Two weeks later Nimble raises $47m and pivots from proxies to search for agents.
The FBI and European police seize SocksEscort — 34 domains, 23 servers, roughly 124,000 customer accounts. The same day, the FBI issues a public advisory naming app SDKs, free VPNs and compromised streaming boxes as the main ways consumer devices are enrolled into proxy networks.
A chief executive changes sides. Vytautas Savickas leaves the top job at Decodo and becomes CEO of Oxylabs Group the same week; Decodo appoints Vaidotas Juknys. Two companies compared in this report, one career.
Bright Data exits mobile proxies, per Proxyway's market research — a category exit by the vendor claiming the largest residential pool in the market.
Spur reports proxy SDKs inside smart TV apps — roughly a third of the LG webOS apps it unpacked. LG and then Samsung move to ban them within six weeks.
Cloudflare sets 15 September 2026 for default blocking of mixed-use crawlers on ad-bearing pages. A week later the European Data Protection Board publishes draft guidelines on scraping for generative AI.
Warburg Pincus invests $130m in Oxylabs at a $3.6bn valuation — the first outside investment in the company's history.
Cloudflare ships Kitesurf, a cloud browser built for AI agents. The company that sells the blocking now also sells the client.
Prices fell for two years.
The reason you've heard
is wrong.
The most repeated explanation for datacenter proxy pricing in 2026 is IPv4 scarcity. The available price data points the other way.
IPv4.Global, the largest public transfer marketplace, reported /16 blocks dipping below $15 per address for the first time in 2025 — a decade low — with /22 to /24 blocks falling to just under $30 from around $35. Its July 2026 report describes prices as having "declined steadily from late 2025 into the opening months of 2026" before a partial recovery that remains "well below previous cyclical peaks".
Registry exhaustion is real: ARIN and RIPE both run waiting lists. But both have operated in this form since 2019–2020, and nothing in the 2026 registry data marks a step change.
Proxy prices fell alongside IPv4 prices, not against them. Residential rates contracted by up to 75% between 2023 and 2025 on two independent accounts, reversing only in 2026 as discounts were withdrawn.
The cost floor did rise — in the cloud, not on the transfer market. AWS began charging $0.005 per hour for each in-use public IPv4 address on 1 February 2024, about $43.80 a year per always-on address, citing a 300% increase in its own acquisition cost over five years. Peers followed. That raises the floor cost of datacenter capacity, and it is a different claim from "the transfer market is tightening".
No public time series of datacenter proxy prices exists. We looked. Any statement about the trend in that category — including a favourable one — rests on assertion.
Oxylabs
$130m from Warburg Pincus at a $3.6bn valuation, 9 July 2026. First outside investment; the company describes itself as the second unicorn from the Tesonet accelerator. Stake size not disclosed by any party.
Bright Data
Reports over $300m ARR, up 50% year over year, targeting $400m by mid-2026, and claims to serve "14 of the top 20 global LLM labs". Company statements, unaudited, repeated by trade press.
Webshare
Owned by Oxylabs since September 2022, operating as a separate brand with a separate price list. The two appear here as separate rows and share an owner.
Two years of falling prices trained the market to expect discounts, and 2026 is the year that stopped. If your current rate came from a 2024 or 2025 negotiation, it is probably better than what the same provider would quote you today — which is an argument for checking renewal terms early rather than for switching.
Where the demand
actually comes from
Automated traffic growth is the best-evidenced fact in this market. The popular explanation for who generates it is much less well evidenced than it sounds.
Each bar measures a different population with a different bot definition and they must not be averaged into one number. All figures are vendor telemetry, unaudited, from companies that also sell bot mitigation or CDN services.
Fastly's observation that 51% of AI requests hit origin against under 9% of human requests is the most useful number here. AI traffic is different in kind, not only in volume: it defeats caching, so it converts directly into origin cost for the site being crawled. That mechanism is behind every access-control change of the past year.
The common account is that agentic search and LLM-oriented crawling tools drive proxy demand. None of Exa, Tavily, Parallel, Linkup, Firecrawl or ScrapingBee publishes anything about proxy spend, vendors or routed volumes. The chain is entirely inferred — and Exa states it runs its own crawling infrastructure, tracking "over 500 billion urls".
The documented demand story runs through a different channel. Proxy vendors attribute growth to AI labs and enterprise agent deployments — model training and production retrieval — not to the tooling layer. Bright Data names LLM labs as customers and reports handling "more than 100 million AI agent interactions daily"; Warburg Pincus framed its Oxylabs investment around AI agents beginning to "navigate the web far more than humans ever have". These are vendor and investor statements rather than independent measurement — and the fact that they are the best evidence available is itself the finding.
Enforcement reached the
infrastructure layer
The FBI seized hundreds of domains associated with NetNut, a commercial residential proxy service operated by Alarum Technologies, listed on NASDAQ as ALAR. Google acted in coordination, disabling accounts and applications tied to the network; Lumen and Shadowserver also participated. Google put the associated botnet at "at least 2 million devices", recruited through SDKs embedded in consumer hardware.
Alarum confirmed the seizures in filings, paused certain network services and cut approximately one third of its workforce, stating it "still does not know the exact root cause". It said it would cooperate fully. No criminal charges against NetNut, Alarum or any individual have been announced, and no finding of wrongdoing has been made.
| Meta v. Bright Data N.D. Cal., Jan 2024 | Summary judgment for Bright Data: Meta's terms bound only logged-in users, so logged-off collection of public data was not a breach. |
| X Corp. v. Bright Data N.D. Cal., May 2024 | Complaint dismissed with leave to amend on preemption and related grounds. |
| Reddit v. Perplexity, SerpApi, Oxylabs UAB, AWMProxy S.D.N.Y. 1:25-cv-08713, filed 22 Oct 2025 | DMCA §§1201 and 1202 claims plus contributory and vicarious liability. Amended complaint February 2026. Reporting dated 31 July 2026 indicates the motion to dismiss was largely denied. Proceeding; nothing decided on the merits. |
| Google v. SerpApi N.D. Cal., Jul 2026 | Two DMCA claims dismissed, one with leave to amend; the court found circumvention adequately pleaded. Not dismissed in its entirety. |
Where your provider's residential IPs originate, and whether it resells capacity it does not control, is an operational risk question — continuity of service, not only ethics.
Bitsight research published 7 May 2026 found active malware infection on a share of residential proxy exit nodes it studied. None of the seven providers profiled in this report were among the networks that research identified. We note it because it is the context in which KYC across this industry tightened.
"Scraping public data is legal" overstates what these cases decide. The Bright Data decisions are district-court rulings from a single district, on contract and preemption grounds, and neither addresses copyright in the collected content. The 2026 actions show plaintiffs moving onto DMCA anti-circumvention theories those earlier cases did not resolve. No court anywhere has yet issued a final merits ruling on whether training a general-purpose model on scraped web data is fair use or covered by the EU text-and-data-mining exception. This is an inventory of public records, not legal advice.
The summer the television
became a proxy
In June 2026 the threat-intelligence firm Spur unpacked 6,038 smart TV applications rather than reading their store listings, and reported residential proxy SDKs inside 2,058 of them, roughly a third of the sample. LG moved to ban them within a month; Samsung followed on 3 August.
Bright Data, whose SDK was among those identified, states that every peer opts in through a dedicated screen and receives value in return, that every customer is vetted, and that its practices had undergone a second independent PwC audit, completed 30 July 2026.
Spur's finding is a technical observation about which SDKs sit inside which application packages. It is not presented as a finding of wrongdoing by any company named, and none of the seven providers compared in this report has been the subject of a regulatory finding on this question.
A banned SDK is a supply shock
Pools sourced through app SDKs shrink when the store bans the SDK, and what remains gets worked harder. GreyNoise, analysing four billion sessions to February 2026, found 78% of residential addresses appeared once or twice and never returned — a thinning pool rotates you into worse addresses.
Provenance is a continuity question
A provider that cannot say where its addresses come from also cannot say what happens to your success rate when a platform closes a door. Put the sourcing answer in the contract and ask what the fallback is. That is a commercial term, not a compliance box.
The offering
Two of the seven do not sell mobile at all
One of them left the category in April 2026 without announcing it.
Pool size is the most marketed and least useful number
It says nothing about availability in the country you need, on the day you need it.
The same product carries different names
“ISP” and “static residential” are used interchangeably, sometimes on a single page by a single vendor.
Session limits are frequently unpublished
Three providers state no maximum sticky duration anywhere on their sites.
Product names, as each
vendor writes them
The same product carries different names across this market, and sometimes two names on one vendor's own site.
| Provider | Residential | Mobile | ISP / static residential | Datacenter shared | Datacenter dedicated |
|---|---|---|---|---|---|
| Oxylabs | Residential Proxies | Mobile Proxies | ISP Proxies; Dedicated ISP Proxies | Shared Datacenter — per-IP and per-GB variants | Dedicated Datacenter Proxies |
| Bright Data | Residential Proxies | No product page or pricing | ISP Proxies — shared and dedicated | Datacenter Proxies, Shared | Datacenter Proxies, Dedicated |
| IPRoyal | Residential Proxies | Mobile — rotating and dedicated | Static Residential Proxies | Not offered — stated position | Datacenter Proxies |
| Decodo | Residential Proxies | Mobile Proxies, 3G/4G/5G | ISP Proxies — shared and dedicated | Datacenter — Pay/GB and Pay/IP | Datacenter Proxies, Dedicated |
| Webshare | Rotating Residential Proxy | Not offered | Static Residential; Private and Dedicated variants | Proxy Servers | Dedicated Proxy Servers |
| Evomi | Residential (Core) and Premium Residential | Mobile Proxies | Static Residential — Shared, Private, Virgin | Datacenter Proxies, billed per GB | No separate SKU |
| DoubleData Proxy | Residential Proxies | Mobile Proxies | ISP Proxies — shared and dedicated | Datacenter — IP-based and traffic-based | Dedicated Datacenter Proxies |
Mobile is the sharpest structural split. Five providers sell a full mobile product with published tiers. Webshare has none. Bright Data has no mobile product page or pricing while still advertising "7M+ Mobile IPs" on its pricing page, and has published no statement about it.
Shared datacenter is a stated refusal at IPRoyal, not a gap. Evomi has no dedicated datacenter SKU and bills datacenter per GB where most bill per IP — a structural difference that makes its datacenter product hard to compare directly.
"ISP proxies" and "static residential proxies" are used interchangeably, sometimes by the same vendor on the same page. Decodo uses both names for one product. IPRoyal markets "Static Residential" and also calls them ISP proxies. Evomi's variant names — Shared, Private, Virgin — encode a reputation claim rather than a network characteristic.
When you compare "ISP proxy" prices across vendors, confirm what is actually sold: how many users share the address, what the bandwidth allowance is, and whether the address is newly issued. Those three move the price more than the label does.
Claimed pool sizes
All figures are vendor claims, verified nowhere in this report. Pool size is the most marketed number here, and the least useful.
| Provider | Mobile · ISP · datacenter pools | Geographic coverage | Sticky session maximum |
|---|---|---|---|
| Bright Data | Mobile 7M+ cited, no product · 1.3M ISP · 1.3M datacenter | Residential 195 countries; datacenter 98 | Not published |
| Oxylabs | Mobile 20M+ · 1.9M dedicated DC · 36–56K shared DC | Residential 195; mobile 140+; dedicated DC 188 | Residential 10 min, up to 1,440; mobile 24 h; ISP unlimited |
| Decodo | Mobile 10M+, 700+ carriers · 100K+ shared DC | 195+ locations; mobile 160+; dedicated DC in 4 countries | 1,440 min per documentation |
| Webshare | No mobile product · ISP and datacenter not published | Residential 195 countries | Not published |
| IPRoyal | Mobile 4.5M+ · 500K ISP | Residential 195+; ISP 31+ | Residential up to 7 days; ISP unlimited |
| Evomi | Mobile 3M+ · ISP and datacenter not published | 150+ countries across products | 30 min, up to 1,440 min |
| DoubleData Proxy | Mobile not published · 700K+ ISP · 770K+ datacenter | 195 countries across products | Unlimited concurrency stated; duration not published |
Test depth, not headline size
Pull a thousand addresses in each country you target and count the distinct ones. Proxyway did that to Evomi Core and found under 100,000 distinct US addresses across 1.2 million requests. The test takes an afternoon and it is the only number that maps to your workload.
Do not pay a premium for a pool figure
The largest claimed pool here is over seven times the smallest, and nothing in this report shows the difference translates into success rate. Where two providers are close on price and far apart on claimed pool, treat the gap as marketing until a trial says otherwise.
Pricing
“From $X” means something different at every vendor
At some it is the best rate, at others the entry rate, at one it is unreachable.
Per GB and per IP are different risk models
One success rate produces a cost problem in the first and a scheduling problem in the second.
Fair-use caps are the second price
Allowances behind similar headline rates differ by an order of magnitude.
Three providers publish nothing above 100 GB
Past that volume the price list is replaced by a sales conversation.
Four reasons proxy
prices don't compare
Read this before the tables. Putting seven headline rates side by side produces a number that is wrong in several directions at once.
"From $X" means something different at every vendor
At Oxylabs, Bright Data and DoubleData Proxy the headline is the best rate: Oxylabs' "from $2.50/GB" is the 1 TB price, while its entry tier is $6.00/GB. At Decodo the advertised floor is not reachable in any published tier — residential is advertised "starting at just $2/GB" against a cheapest published tier of $2.75/GB. One phrase, three incompatible meanings.
Per GB and per IP are different risk models, not different units
The same drop in success rate lands in a different column depending on how you are billed. Under per GB it lands on the invoice; under per IP it lands on the calendar.
Some of these are promotional prices, and some don't say
Bright Data displayed a code giving 50% off residential for three months; Webshare displayed a "50% OFF" banner. Decodo labels every mobile tier "SAVE 50%" with no expiry stated anywhere, so whether that is a promotion or the effective price cannot be told from the page — we show both columns.
The billing period is often unstated, and fair-use caps are the real price
Decodo publishes monthly billing plus VAT; IPRoyal residential traffic "never expires" with no period published; DoubleData Proxy publishes none for its per-IP products. Separately, bandwidth allowances act as a second price: Bright Data includes 100 GB per IP per month on ISP and datacenter, IPRoyal's datacenter proxies 100 GB per cycle despite a "no restrictions" framing, and Oxylabs drops ISP concurrency from 100 sessions to 10 once you pass 50 GB in a month.
The obvious response is to buy the cheapest addresses and buy more of them. Under per-IP billing that breaks in a specific way: adding capacity does not fix a low success rate, it multiplies the retries, and the constraint becomes wall-clock time rather than budget. Cheap capacity is worth buying when the failure mode is throughput, and avoiding when it is cost.
Residential, per GB
Each panel is one provider's published ladder drawn on the same axes: rate on the vertical, monthly volume on a logarithmic horizontal. A steep line means the discount is real; a flat one means the headline rate is close to the only rate.
Price the tier you will actually sit on
Model your monthly gigabytes first, then read the rate at that volume rather than the headline. On these ladders the difference between the entry tier and the advertised floor exceeds 2:1 at three of the seven providers, and at one of them the floor requires ten terabytes a month.
A flat ladder is a negotiating position
Where a provider's line barely slopes — Decodo, Evomi Premium, DoubleData Proxy — volume gives you little leverage and the entry price is close to the real price. Where it plunges at the right-hand end, the discount is real but so is the commitment required to reach it. Both are worth knowing before the call.
Mobile, per GB
The category where the shortlist narrows fastest: two of the seven do not sell a mobile product at all, and the market rate moved further in the past year than in any other category.
Decodo is shown at its discounted column; every one of its mobile tiers carries a "SAVE 50%" label with no stated expiry, and the list column runs $7.50 to $5.50. IPRoyal also sells dedicated mobile by rental period, from $130 a month at 30 days, capped at 30 GB per day per proxy.
Proxyway's 2026 market research reports that Bright Data discontinued its mobile proxy product in April 2026. As of 24 August 2026 its site lists three proxy products and has no mobile product page or pricing, while its proxy pricing page still advertises "7M+ Mobile IPs" and its FAQ still lists Mobile among its proxy types. Bright Data has published no statement of its own about the change.
Webshare has never sold mobile. That leaves five providers in this chart, and the vendor claiming the largest residential pool in the market is not among them.
The publisher is the dearest here, at every tier it publishes. Its $8.40 entry is a pay-as-you-go rate capped at 10 GB, and its published ladder ends at $6.30 against $5.00 at Oxylabs and $3.20 at Evomi for comparable volumes. Decodo's figures in this chart are its discounted column; its list column runs $7.50 to $5.50. This is the one category where the publisher of this report is the most expensive option in it, and the chart is drawn the same way as every other.
Proxyway's 2026 research reports mobile proxy rates falling by up to 98% year over year — the sharpest correction in any category it tracks, and an order of magnitude beyond the residential contraction over the same period.
If you priced a mobile workload before 2026 and have not re-quoted since, that is the single number in this report most likely to change your budget.
ISP proxies are priced
on four different axes
This is the category where a side-by-side price table does the most damage, because three of the seven providers do not price it by quantity at all. Before comparing any two numbers, work out which of these four structures each vendor is using.
Quantity ladder
The rate falls as you buy more addresses. The only structure where "cheaper at volume" means what a buyer expects.
Rental period
The rate falls with how long you commit, not how many you take. Ten addresses and a thousand cost the same each.
Address class
Shared, private or newly issued — priced by exclusivity and reputation rather than by volume. No quantity discount is published.
Entry point only
One published rate at a minimum quantity, then a sales conversation. The publisher of this report is alone in this bucket — it publishes an entry rate at ten addresses and a custom enterprise tier, with nothing between.
Webshare's entry rate is $0.30 per address and Bright Data's is $1.80 — but Webshare's entry ladder carries a 250 GB allowance across the whole account, while Bright Data allows 100 GB per address per month, which at twenty addresses is two terabytes. Evomi and DoubleData Proxy state unlimited bandwidth with no metering. Oxylabs is unlimited under fair use, but cuts concurrency from 100 sessions to 10 above 50 GB a month. Four vendors, four different meanings of the same number.
How many people share the address. Oxylabs states up to three users; most publish nothing.
What the allowance is, and what happens past it. Throttling, overage and a hard stop are three different products.
Whether the price is per month. IPRoyal prices by rental window; DoubleData Proxy publishes no billing period for this category at all.
Datacenter, billed
per IP
The widest price spread in this report, and the number that explains the least.
Log scale. The volumes differ — Bright Data and Oxylabs quote 10 addresses, DoubleData Proxy 5, Webshare 100 — so read these as each vendor's way in, not a like-for-like comparison. The ring marks the publisher. IPRoyal declines to sell shared datacenter; Decodo and Evomi bill per GB only.
| Provider | Variant | Published ladder | Terms |
|---|---|---|---|
| Bright Data | Shared | 10 IP $1.40 · 100 IP $1.00 · 500 IP $0.95 · 1,000 IP $0.90 | 100 GB fair usage per IP per month; minimum 10 IPs |
| Bright Data | Dedicated | 10 IP $2.20 · 100 IP $1.70 · 500 IP $1.50 · 1,000 IP $1.30 | 100 GB fair usage per IP per month; above 1,000 IPs on enquiry |
| Oxylabs | Shared | 10 IP $1.20 → 3,000 IP maximum | Product page advertises $0.70/IP at volume; intermediate tiers unpublished |
| Oxylabs | Dedicated | 3 IP $2.25 → 3,000 IP maximum | Advertises $1.20/IP at volume. Five free IPs to trial, no card required. |
| IPRoyal | Dedicated, by period | From $1.57 at 30 d · $1.48 at 60 d · $1.39 at 90 d, per proxy | 100 GB per cycle per proxy, pooled, then throttled — despite a "no restrictions" framing |
| Webshare | Shared | 100 IP $0.0299 · 1,000 IP $0.0269 · 10,000 IP $0.0224 · 100,000 IP $0.0179 | Free tier of 10 proxies, 1 GB monthly, stated as "forever" |
| DoubleData Proxy | Shared / dedicated | Shared from 5 IP $0.75; dedicated advertised from $1.20/IP | Unlimited bandwidth stated. Above 1,000 IPs is custom. Intermediate tiers not published. |
At the same volume — 100 addresses — Webshare lists $0.0299 per IP per month against Bright Data's $1.00: a thirtythreefold difference. We quote it at a common quantity deliberately, because the headline entry rates here sit at 5, 10 and 100 addresses, and comparing those directly is the mistake page 18 warns against. The gap measures exclusivity, address reputation and bandwidth entitlement — and neither vendor publishes the figure that would settle it: how many other customers share the address. Six of the seven offer a trial, which is the only way to resolve it.
Datacenter, billed
per GB
Five providers sell the same capacity by traffic instead of by address, and four of them publish a ladder. The models are not interchangeable, and which suits you depends on whether your bottleneck is concurrency or volume.
Oxylabs also sells datacenter per GB, from $0.59 at 20 GB, and advertises $0.44 at volume; its intermediate tiers sit behind a configurator and are not published as static text, so no ladder is drawn. IPRoyal does not sell this category by traffic.
When traffic is bursty or seasonal, when you cannot predict concurrency, or when you are testing a target before committing. You pay for what you pull and nothing sits idle. The exposure is that a drop in success rate raises the bill directly — every retry is billable transfer.
When volume is steady and high, when you need predictable spend, or when concurrency rather than bandwidth is the constraint. A poor success rate then costs throughput and cycle time instead of money — a scheduling problem rather than a budget one.
The crossover is worth calculating once. At Bright Data's published rates, a terabyte a month costs $510 on the per-GB plan — its 1 TB tier of $0.51/GB — against $900 for a thousand shared addresses — but those addresses carry a 100 GB monthly allowance each, so the comparison only holds if your volume genuinely sits near a terabyte. Run the same arithmetic against your own monthly volume before choosing a model. It is the one calculation in this report a buyer can complete from published numbers alone.
Trust
Success rate has converged; latency has not
Optimising on block rate means optimising the variable that moved least.
G2 scores higher than Trustpilot wherever the sample is real
True at all four providers with a substantial G2 sample. That is a fact about the platforms rather than about the companies.
Five complaints recur across the whole market
A theme visible at six companies is unlikely to be one unhappy cohort.
Governance is voluntary membership and draft regulation
Neither substitutes for getting the answer written into the contract.
What independent testing
found
This report does not run its own benchmarks — the publisher sells proxies, and a vendor grading itself is worth nothing. What follows is third-party testing, reported with its own caveats.
Residential success rates across the thirteen vendors benchmarked ranged from 95.28% to 99.93%, with average response times between 0.41 and 1.88 seconds. The spread in success rate is under five points; the spread in latency is more than fourfold.
That asymmetry is the practical finding. At the top of this market, success rate has largely converged and stopped being a differentiator. Latency has not. If you are choosing between established residential vendors on their block rate alone, you are optimising the variable that moved least.
Bars are proportional to the measured figure. The first two rows are the extremes of one thirteen-vendor run. The Evomi Core reading comes from a separate Proxyway benchmark published 30 July 2026, is not part of that run and is not directly comparable with it — it is shown in mint for that reason.
Proxyway's July 2026 re-test of Evomi Core recorded better than 99.9% success in every locale except India and a 0.23-second average UK response — among the best figures it has published. In the same test it surfaced fewer than 100,000 distinct US addresses across 1.2 million connection requests — a pool an order of magnitude shallower than the headline claims elsewhere in this report.
Both findings are true at once, and together they describe the trade a cheap pool makes: excellent performance on undefended targets, limited address diversity if your workload needs it.
Proxyway's annual awards, published 17 June 2026, named Decodo Best All-Rounder, Evomi Best Value, Oxylabs Most Scalable, DataImpulse Most Flexible and Byteful Newcomer of the Year. The 2025 edition named Oxylabs Best Enterprise, Decodo Value Choice and Evomi a Newcomer.
Awards from a publication that also carries affiliate links are not independent in the way a laboratory result is. They are included because they are dated, attributed and consistent across two years — and because no better third-party recognition exists in this sector.
What review platforms
record
Aggregate scores as displayed on 24 August 2026. Review scores are time-weighted and drift. Everything here is what reviewers say, not a finding by the publisher about any company.
| Provider | G2 | Reviews | Trustpilot | Reviews |
|---|---|---|---|---|
| IPRoyal | 4.7 | 839 | 3.8 | 2,612 |
| Bright Data | 4.7 | 341 | 4.5 | 956 |
| Decodo | 4.6 | 714 | 4.3 | 1,847 |
| Oxylabs | 4.5 | 424 | 4.0 | 712 |
| Evomi | 4.4 | 4 | 4.3 | 56 |
| Webshare | 4.0 | 5 | 4.2 | 1,094 |
| DoubleData Proxy | Too few public reviews to report a score | |||
G2 scores higher than Trustpilot at all four providers with a substantial G2 sample — and the gap varies by nearly a full point. Webshare runs the other way, on five G2 reviews.
This is a fact about the platforms, not the companies. G2 requires business-email verification and vendors run review campaigns into it; Trustpilot is open and attracts people motivated to complain. A buyer reading one platform is reading a biased sample — and which platform they picked matters more than which vendor they are researching.
The proxy review ecosystem is heavily affiliate-funded: ranked "best provider" listings frequently order vendors by commission rate, and vendor-authored comparisons are presented as independent reviews. Neither was used as a source of fact here.
What the people doing
the work say
Engineers and researchers, in their own words, from public discussion threads and on-the-record interviews. Handles are pseudonymous where the platform is; names are given where the speaker is identifiable.
Hacker News comments are quoted verbatim from public threads and attributed by date rather than by handle, since the handles are pseudonymous and the individuals have not chosen to be named in a commercial document. Named quotes are from press releases and published interviews. Review-platform quotations were deliberately excluded: the wording could not be verified against the source pages to the standard the rest of this report is held to, so those reviews appear as themes on the next page rather than as quotes.
Five complaints that recur
across the whole market
Patterns visible at several providers at once are more useful than any single-vendor grievance, and more reliable — a theme across six companies is unlikely to be one unhappy cohort.
Price complaints are universal and independent of price level
Cost is the first or second negative theme at all six providers with meaningful review volume, including the cheap ones. The conclusion is not that the market is overpriced but that headline $/GB is not the number customers react to: feature multipliers, minimum purchases, expiring versus non-expiring traffic and fair-use thresholds decide the effective rate.
Small customers report worse economics than enterprises
Reviewers at Oxylabs, Bright Data and Decodo independently describe entry tiers as disproportionately expensive. The ladders on pages 19 to 23 corroborate the structure: the spread between entry and best published rate exceeds 2:1 at three of the seven providers.
Support sentiment is bimodal everywhere
Every provider with review volume has support as both a leading strength and a recurring criticism. At IPRoyal the split is starkest — 216 G2 reviews tag support positively against 30 tagging it negatively. The shape of the negative half is near-identical across vendors: scripted replies, long chat queues, and reviewers reporting that responsiveness drops once a billing dispute begins.
Verification and account-suspension friction rose across the industry
Identity-document requests, account restrictions and signup rejections appear in recent reviews at four providers. Decodo has stated publicly that it blocks forwarding email domains as "standard risk-management practice". Read as an industry compliance shift rather than a service failure this is explicable: it follows sustained enforcement pressure on the sector, described on page 12.
"Is this IP what I bought?" recurs at four of the six
Reviewers across several providers report mismatches between the advertised type or location of an address and what detection tooling reports. We publish this as buyer due diligence and not as a claim that any provider misrepresents its product: verify IP type and geolocation independently, on your own targets, before committing to a term. It is a cheap check and the one most often skipped.
Individual reviewer allegations of billing impropriety or product misrepresentation were reviewed during research and are deliberately not reproduced. They are unverified single-source claims from pseudonymous accounts, several concern conduct that would be unlawful, and the providers concerned respond to most such reviews with a contrary account. The neutral form of the underlying signal is carried in themes 01 and 05.
Provider notes
Positioning in each vendor's own terms, plus corporate and product developments on the public record. Facts and dates only; nothing here is an assessment of quality.
Oxylabs
Announced $130m from Warburg Pincus at a $3.6bn valuation on 9 July 2026, its first outside investment. Owns Webshare, acquired September 2022. Named as a co-defendant in Reddit v. Perplexity et al., filed October 2025 and proceeding, with nothing decided on the merits. Publishes the widest product range in this set, including both per-IP and per-GB datacenter variants.
Bright Data
Reports over $300m ARR growing 50% year over year, unaudited. Won summary judgment against Meta in January 2024, and saw X Corp.'s complaint dismissed with leave to amend in May 2024 — the two most-cited US scraping decisions, both in its favour. Its mobile product page was present as recently as March 2026 and is gone now, and Proxyway dates the discontinuation to April 2026; Bright Data itself has published nothing on the change.
IPRoyal
The only provider here that declines to sell shared datacenter proxies on stated principle. Prices ISP, datacenter and dedicated mobile by rental duration rather than quantity — structurally hard to compare and unusually easy to trial, since a 24-hour ISP rental is the cheapest way into that category anywhere in this report. Residential traffic never expires. Carries the widest G2-to-Trustpilot gap of any provider covered.
Decodo
States that legacy Smartproxy endpoints, credentials and subscriptions remain functional with backward compatibility maintained indefinitely. Sells shared and dedicated variants of both ISP and datacenter — the broadest structural coverage here — but publishes ladders for only some of them. Publishes billing terms most explicitly of the seven: monthly recurring, plus VAT.
Where litigation appears above it is because the docket is a public record and, in Bright Data's case, because the outcomes were in its favour and are frequently cited by others. The inclusion of a pending case in Oxylabs' entry carries no implication about its merits; a defendant in a proceeding that has not been decided is exactly that.
Provider notes,
continued
Webshare
The most datacenter-weighted provider here and the cheapest per IP by a wide margin, with an unusual free tier of 10 proxies and 1 GB of monthly traffic stated as available "forever" with no card required. The only provider in the set publishing an annual billing option alongside monthly. Its five G2 reviews are too few to read; its Trustpilot sample of 1,094 is substantial.
Evomi
Publishes a Core residential rate of $0.49/GB at a $49.99 monthly commitment, the lowest in this report by a factor of five, alongside a separate Premium residential product starting at $3.98/GB — the two are different products in the vendor's own framing and should not be read as one price. Bills datacenter per GB rather than per IP. Public review coverage is thin at roughly 60 reviews in total.
DoubleData Proxy Publisher
Sells all five product categories with published entry rates in each. Residential runs $4.20/GB pay-as-you-go to $2.93/GB at a $1,999 commitment, placing it mid-pack in this set; datacenter starts at $0.75 per IP from five IPs, at the cheaper end but well above Webshare. Claims over 125 million residential IPs, 700,000 ISP and 770,000 datacenter addresses across 195 countries.
Publishes no billing period for its per-IP products, no sticky-session maximum, and no tiers between its entry and enterprise rates — three gaps this report criticises in others and records here for the same reason.
A comparison that excluded the company paying for it would be easier to defend and less useful to read. Including DoubleData Proxy means every criticism in this report had to survive being applied to its own product first. Weigh the row accordingly; it is marked in every table for exactly that purpose.
What the industry is
doing about itself
Three things a buyer can check: who joined the sector's voluntary body, what the European regulator proposed in July, and which technical standards are real.
| Provider | Listed as a member |
|---|---|
| Oxylabs | Yes |
| Decodo | Yes |
| Evomi | Yes |
| Webshare | Not listed; owner Oxylabs is |
| Bright Data | Not listed |
| IPRoyal | Not listed |
| DoubleData Proxy | Not listed |
The European Data Protection Board published draft guidelines on scraping for generative AI on 8 July 2026. They treat legitimate interest as the workable legal basis but require a documented balancing test; expect a complete list of scraped sources where feasible, plus the crawler's characteristics; require exclusion of sites likely to hold vulnerable people's data; and treat robots.txt and CAPTCHAs as anti-scraping signals to be respected.
Consultation closes 30 October 2026. These are drafts, not law, and this is an inventory of a public document rather than legal advice.
Web Bot Auth. Cloudflare's signed-agents scheme launched in August 2025 with OpenAI's ChatGPT agent, Block's Goose and Browserbase. It works in production, but remains an IETF draft rather than a ratified standard.
Content Signals. Cloudflare's robots.txt extension for expressing search, AI-input and training preferences. Google's John Mueller said in July 2026 that it has "no effects whatsoever for any crawler or LLM".
llms.txt. Ahrefs checked 137,210 domains in May 2026: 28% publish the file and 97% of those received zero requests that month.
Governance in this sector currently consists of voluntary membership, draft regulation and one working authentication scheme. None of it substitutes for asking a provider where its addresses come from and getting the answer in the contract — which is why that question opens the checklist at the end of this report.
Matching the product
to the workload
The proxy type is usually determined by the target's defences and the session behaviour your workload needs, not by budget. Budget determines which provider, and that is the easier question once the type is settled.
| Workload | Typical fit | Billing model that suits it | What decides the provider |
|---|---|---|---|
| AI training data collection | Residential, datacenter for undefended sources | Per GB | Volume tiers above a terabyte, which most of this set does not publish. Sourcing documentation matters most here. |
| Real-time retrieval for agents | Residential or ISP | Per IP where volume is steady | Latency and session stability rather than pool size. Per-IP billing removes the cost volatility. |
| Price and availability monitoring | Residential, geo-targeted | Per GB | City and ZIP-level depth in your own markets. Country counts are close to meaningless. |
| Account-bound, session-heavy work | ISP / static residential | Per IP | Session duration limits and bandwidth allowance per address — the variable that differs most at similar headline rates. |
| SERP and rank tracking | Datacenter, escalating to residential | Per IP or per GB | Cost per successful request at your own success rate, not the headline. The category where cheap capacity most often wins. |
| Bulk collection, undefended sources | Shared datacenter | Per IP | Price and concurrency. Six of seven sell it; the spread mostly reflects exclusivity and bandwidth. |
Most cost overruns come from buying one product class for every workload. Splitting collection so undefended sources run on datacenter capacity and only the defended remainder touches residential is worth more than any negotiated discount — compared per gigabyte, the gap between the two classes here runs from roughly 6:1 to 12:1.
Access is also becoming permissioned: Cloudflare blocks by default for new domains and has set 15 September 2026 for mixed-use crawler blocking on ad-monetised pages. A strategy built entirely on being indistinguishable from a browser is exposed to a visible direction of travel.
Ten questions worth
asking any provider
None can be answered from a pricing page, which is the point. Each addresses something this report found either unpublished across the market or inconsistently published within a single vendor's own site — the publisher's included.
Where do the residential addresses come from, and what is the consent mechanism?
Ask for the actual answer, not the compliance page.
Do you resell or white-label capacity you do not operate?
Google's threat intelligence team describes this as routine in the sector.
How is success rate defined in the contract?
An HTTP 200 returning an empty page is a failure. Confirm the definition agrees.
What is the fair-use threshold, and what happens at it?
Throttling, overage billing and hard cut-offs are three very different answers.
What is the billing period and the exit term?
Two of the seven publish no billing period for at least one product — IPRoyal and the publisher of this report.
What is the maximum sticky session duration, in writing?
Three of seven do not publish it and one contradicts itself by a factor of 48.
How many other customers route through the same address?
Nobody publishes this, and it explains most of the price spread in shared datacenter.
What does the tier ladder look like above the published tiers?
Several providers publish nothing above 100 GB.
Is the rate I am being quoted promotional, and for how long?
Two vendors were running promotions on our access date and one gives no expiry.
Can I trial this on my own targets before committing?
Six of seven offer a trial or money-back window. It is the only test that counts.
Published by DoubleData Proxy in August 2026. Pricing and product specifications were read directly from each vendor's public pages on 24 August 2026 and carry that date throughout. Market context is sourced to court records, regulator publications, CDN and security-vendor telemetry, and company announcements, each dated in the text.
This report describes products that change weekly. If a figure attributed to your company is wrong or has moved, we would rather correct it than defend it — write to us and we will amend the document and note the change. DoubleData Proxy publishes this report and is one of the seven providers compared in it; its row is marked in every table so you can weigh it accordingly.
