◆Own the stack instead of renting it: $15,388 stays with you over three years.
Cumulative cost of the same five-person workload over 36 months.
ModelledOwnershipModelled. Arithmetic on stated inputs. The inputs are named on the card.Method 36 months, five people. Hardware amortised at zero residual value; resold at 30% the gap widens by $720. Inputs [C].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
$15,388Gap at month 36
4.5×Rent multiple
8.4 moBreak-even
02
◆Price rise, outage, shutdown, ban — OwnX survives all five. Hosted AI survives none.
Whether the workload still runs and the history is still yours, under five failure modes.
Our framework, our scoringOwnershipOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method Five failure modes, five stack types. 'Survives' means the workload still runs and the history is still yours. Our definition, published.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
5 of 5OwnX
0 of 5Hosted AI
5Failure modes tested
03
◆Ownership, scored: OwnX 5.0. The best alternative manages 3.8.
Our ownership framework applied to five stack types.
Our framework, our scoringOwnershipOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method Our criteria, scored by us. This is not a measurement and is not presented as one.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
5.0OwnX average
1.3Hosted AI
3.8×Gap
04
Five vendors quietly cut what your money buys. Our terms moved 0%.
Change in what five vendors gave you, against the terms in force eighteen months earlier.
From published sourcesCandourFrom published sources. Read from public pricing pages, contracts or documentation.Method Our own accounts and invoices, Feb 2025 – Aug 2026. Vendors anonymised pending legal review; dated screenshots are in the data room [C].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
5 of 5Vendors that changed terms
−67%Median change
0Changes to OwnX terms
05
Nobody else owns all five layers — price index, cloud, models, OS, device.
Which layers of the stack each kind of vendor actually owns.
Our framework, our scoringOwnershipOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method Layer definitions are ours and published. Category comparison, not a vendor comparison.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
5 of 5OwnX layers
2.5 of 5Best competitor
6Categories
06
◆Leaving OwnX takes half a day. Leaving an enterprise contract takes 210.
Days from the decision to leave until the same workload runs somewhere else.
From published sourcesOwnershipFrom published sources. Read from public pricing pages, contracts or documentation.Method Timed on our own accounts where an exit was available; otherwise read from standard contract terms [C]. 18 exit attempts in total.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
0.5 dOwnX worst case
210 dEnterprise median
18Exits timed
07
Consumer AI got 60% dearer in a year. Our price didn't move.
Cumulative change in published list price by vendor category, against Q3 2025.
From published sourcesCostFrom published sources. Read from public pricing pages, contracts or documentation.Method Published list prices captured quarterly from public pricing pages [C].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
+60%Worst consumer rise
−38%Best marketplace fall
0%OwnX
08
◆Renting costs $19,800 over three years. Owning costs $4,412.
Where three years of money goes, costed line by line for both stacks.
ModelledCostModelled. Arithmetic on stated inputs. The inputs are named on the card.Method Overage taken from our own invoices as a proportion of base. Electricity measured at 75 W sustained, $0.15/kWh, 40% duty cycle.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
$19,800Rent total
$4,412Own total
78%Less
02 · Earth Compute · Live
Compute has a price now.
Oil has a price. Gold has a price. Compute didn't — so we built the index, made it public, and kept the history.
44Provider feeds tracked27 pull live from the provider's own API
120,389SKUs trackedBare metal and cloud, all feeds
The same GPU server costs 5.2× more if you don't shop. We shop 34 providers every hour.
The same 8×H100 configuration, priced across every provider we index, in one hour.
From the live indexCostFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Live index snapshot, on-demand list pricing, spot excluded, captured within a single hour on 12 Aug 2026 [A].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
5.2×Cheapest to dearest
$7.22Spread per hour
34Providers
02
◆We can see 62% of the world's compute. The target is 90%, and the gap is published.
Share of identified global compute capacity present in the index.
From the live indexCandourFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Indexed capacity over best available estimate of total addressable capacity [A]. An estimate over an estimate, and labelled as one.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
62%Coverage today
90%Target
28 ptDistance to target
03
6,621 data centres mapped — 31× more than the biggest number anyone else has.
Data-centre facilities counted, against the only two public numbers that exist.
From the live indexCoverageFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Ours: PeeringDB + OpenStreetMap facilities, deduplicated by location, 24 Aug 2026 [A]. Google: ABI Research 2025 campus estimate. AWS: leaked 2023 data.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
The four physical layers indexed alongside the machines.
From the live indexCoverageFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Snapshot of 24 Aug 2026 [A]. Each layer carries its own open licence — OpenStreetMap ODbL, PeeringDB CC-BY, Epoch AI CC-BY — and its own refresh cadence.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
10,674Fibre routes
1,326Exchange points
12 GWIT power capacity
05
Twenty-one live catalogues in one query — from Azure's 70,276 products to Hostkey's 42.
Live inventory carried per provider — 900× between the largest and the smallest.
From the live indexCoverageFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Live feed table, 24 Aug 2026 [A]. Counting conventions differ by provider; each is counted as it publishes itself, so these compare catalogue size, not capacity.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
70,276Largest catalogue
42Smallest live
21Live feeds
06
◆Every feed shows its age in public — freshest one day, stalest thirty-nine.
Days since each live feed's price content changed — the twelve oldest shown.
From the live indexCandourFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Live feed table, 24 Aug 2026 [A]. Every feed is re-checked at least daily; 'data age' is the time since the price content last changed. Cherry Servers at 39 days is the worst on the board and it is published, not hidden.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
1 dFreshest data
39 dStalest, and shown
4 of 21Older than two weeks
07
Search one provider's portal, see one provider. Search here, see thirty-four.
How much of the market is visible from where you're standing.
From the live indexCoverageFrom the live index. Counted or captured by the index. Not a hardware measurement.Method A provider's own portal can only ever show that provider. One Earth Compute search spans every indexed provider with live pricing at once [A].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
34Providers in one search
1In a provider portal
34×Field of view
08
◆2,837 prices, each linked to its evidence. Disagreements get published, not averaged.
The five feeds behind 2,837 priced cells — and what happens when they disagree.
From the live indexCandourFrom the live index. Counted or captured by the index. Not a hardware measurement.Method 2,837 priced cells across 5 feed types, each stored with the date it was seen and a link to its evidence [A]. The vendor's own API is treated as the authoritative source for what that vendor charges.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
2,837Priced, evidence-linked cells
27Direct provider APIs
0Disagreements averaged away
09
The same kilowatt-hour costs $0.070 in Indonesia and $0.442 in Britain.
Business electricity by country, weighted into the blend by data-centre count.
From the live indexCostFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Industrial tariff — the rate a data centre buys on, not the residential one. 29 countries in the blend, 197 observations, 2010–2026 [A].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
$0.070Cheapest, Indonesia
$0.442Priciest, the UK
$0.190Blended per kWh
10
Norway's grid is 8× cleaner than Britain's. Where you run is the biggest carbon lever.
Grid carbon intensity where the data centres are.
From the live indexCostFrom the live index. Counted or captured by the index. Not a hardware measurement.Method gCO₂ per kWh, OWID / Ember [A]. Norway's grid at 28 g against the UK's at 217 g means geography moves emissions nearly 8× before any hardware choice is made.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
28 gNorway
217 gUnited Kingdom
7.8×Spread
11
Equinix alone runs 215 facilities. Two operators hold a third of the listed world.
PeeringDB-listed facilities by operator.
From the live indexCoverageFrom the live index. Counted or captured by the index. Not a hardware measurement.Method 5,265 PeeringDB-listed facilities across 2,342 operators, snapshot 24 Aug 2026 [A].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
215Equinix facilities
2,342Operators listed
148Countries with a facility
12
Two in five of the world's listed data centres sit in one country.
Facilities by country, PeeringDB top five.
From the live indexCoverageFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Share of PeeringDB-listed facilities by country, 24 Aug 2026 [A].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
39.4%United States
30.2%Everywhere else
148Countries in the index
13
GPU prices fell 56% in a year. Hyperscaler list prices fell 4%. We kept the receipts.
Monthly median price for an 8×H100 equivalent, against hyperscaler list.
From the live indexCostFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Captured daily, reduced to monthly medians [A]. Providers entering or leaving change the denominator, so the count per month matters.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
−56%Median, 12 months
−4%Hyperscaler list
14×Divergence
14
We carry live prices for 34 providers. The next best source carries nine.
Distinct providers whose live pricing each source actually carries.
From published sourcesCoverageFrom published sources. Read from public pricing pages, contracts or documentation.Method Counted by hand from each source's public coverage list, August 2026 [C].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
34Earth Compute
9Best alternative
3.8×Coverage lead
15
◆Our price data is public, free, and needs no account. Everyone else's is a funnel.
What you can actually do with each source's price data.
Our framework, our scoringOwnershipOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method Read from each source's public terms and product pages [C]. Our categories.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
6 of 6Earth Compute
3 of 6Best alternative
$0Cost to use
16
Price transparency, scored: Earth Compute 5.0 against a best-rival 2.5.
Our transparency framework applied to five kinds of price source.
Our framework, our scoringCoverageOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method Our criteria, our scores, published so they can be disputed.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
5.0Earth Compute
2.5Best alternative
4Criteria
17
Same server, different continent: $5.60 an hour in Hong Kong, $11.80 in Australia.
Indexed median $/hr by region and configuration.
From the live indexCostFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Median across all indexed providers with capacity in that region, August 2026 [A].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
$5.60Cheapest region
$11.80Dearest
2.1×Regional spread
18
Our prices are fifteen minutes old. The best alternative's are a day old.
Median age of the price each source shows you.
From the live indexSpeedFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Timestamp on the served price against the underlying capture, sampled 200 times per source [A].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
15 minOur median age
24 hBest alternative
96×Fresher
03 · OwnCloud · Beta
Twelve minutes, not six weeks.
Host, train and run models for up to 80% less — and leave whenever you want, for nothing.
The right infrastructure in 12 minutes — roughly 5× faster than a major cloud.
Time to reach suitable infrastructure, not time for a server to boot.
ModelledSpeedModelled. Arithmetic on stated inputs. The inputs are named on the card.Method This models a workflow. It is not a boot-time measurement and we do not present it as one. 'Major cloud' is the mean of the three hyperscalers marked, each timed separately on our own accounts.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
12 minOwnCloud
~5×Faster than a major cloud
14×vs manual
02
Cheapest of eleven, and it isn't close: we save 80%, the runner-up 41%.
Saving on the same workload, priced across every provider we index.
From the live indexCostFrom the live index. Counted or captured by the index. Not a hardware measurement.Method The same workload priced across all indexed providers in one month [A]. Published pricing throughout; no negotiated rates.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
80%OwnCloud saving
41%Best of the rest
2×Lead over second
03
◆The same four-GPU year: $16,469 here, $31,438 at the market average.
Twelve-month total for the same four-vGPU workload.
ModelledCostModelled. Arithmetic on stated inputs. The inputs are named on the card.Method Market average is the mean across indexed providers carrying the equivalent configuration for the full twelve months [A].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
$16,469OwnCloud
$31,438Market average
$14,970Annual saving
04
◆Leaving costs $0. Everywhere else it's egress fees, notice periods and stranded data.
What it costs to walk away, four ways.
From published sourcesOwnershipFrom published sources. Read from public pricing pages, contracts or documentation.Method Egress from published per-GB rates; term and notice from standard contracts [C]. Export timed by us on both sides, three runs each.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
$0OwnCloud exit cost
$184Egress alone, elsewhere
30 minOur full export
05
◆The same $200 job for $10 — one better decision at a time.
One batch inference job, with one more decision applied at each step.
Measured on hardwareCostMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method Quality held constant and verified against the same eval at each step — the model-size step only counts if output quality holds.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
$200 → $10End to end
95%Reduction
5Runs averaged
06
Describe the workload; 32,324 offers become five qualified ones in under a minute.
A live sample of five qualified offers, drawn from 32,324 in under sixty seconds.
From the live indexSpeedFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Live sample of 5 of 32,324 offers [A]. These are different hardware classes, so the prices are not comparable with each other — the point is that one requirement returns a qualified shortlist across providers, regions and hardware types at once.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
32,324Offers searched
5Qualified in the sample
under 60 sFeed freshness
07
Airbnb made the world's places bookable. OwnCloud makes its compute deployable.
The same market structure, eighteen years apart.
Our framework, our scoringOwnershipOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method Our framing. It is an argument about market structure and there is no measurement here.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
2008Airbnb's fragmented market
2026Ours
1Deployment layer
08
Seven hyperscalers and seven data-centre firms behind one routing layer.
What sits on each side of the routing layer today.
From the live indexCoverageFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Current integrated network [A]. Names are the integrations live in the routing layer, not partnerships or endorsements.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
32Providers integrated
14Named on both sides
1-clickDeploy
09
One deployment layer, 482 cities, 60 countries.
What the routing layer actually reaches.
From the live indexCoverageFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Counted from live provider integrations, August 2026 [A].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
482Cities and regions
60Countries
32Providers
10
900 petabytes of bandwidth and 26 of memory, searchable from one place.
Capacity present in the index across all integrated providers.
From the live indexCoverageFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Aggregated from provider catalogues [A]. Availability at any given moment is a different number and we do not claim it.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
900 PBBandwidth
26 PBMemory
32,324Compute offers
11
Run the same model for a year and pay 80% less than the hyperscaler bill.
Cumulative twelve-month cost of one 8B model at 40% duty cycle.
From the live indexCostFrom the live index. Counted or captured by the index. Not a hardware measurement.Method Same model, same token volume, run concurrently for twelve months [A].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
ModelledCandourModelled. Arithmetic on stated inputs. The inputs are named on the card.Method One 8B model, twelve months. Cost as a function of duty cycle with all other variables held constant.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
13%Crossover duty cycle
47%Our median customer
$2,700Saving at 40%
13
◆Taking your data out costs $0 here — and up to $4,600 elsewhere.
Published egress cost by data volume.
From published sourcesOwnershipFrom published sources. Read from public pricing pages, contracts or documentation.Method Rates read from public pricing pages, August 2026 [C]. Ours is zero because there is no egress charge.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
$0Any volume
$4,60050 TB elsewhere
$0.092Per GB, typical
14
Buying compute, scored on price, choice, speed and freedom: 5.0 against 3.5.
Our buying framework applied to five ways of getting compute.
Our framework, our scoringOwnershipOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method Our criteria, our scores, published.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
5.0OwnCloud
3.5Best alternative
4Criteria
04 · OwnAI · Beta
No meter. No cap.
Every other assistant stops you somewhere. We published where we lose, too.
10Capability domainsOne workspace
$980Twelve-month stackAgainst $3,600 for proprietary APIs
Ten kinds of AI work — text to video to agents — in one workspace.
The ten capability domains available in a single workspace.
Our framework, our scoringCoverageOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method A capability map, not a benchmark. The blind quality test below is where the quality claim is actually tested.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
10Domains
1Workspace
2Domains we don't claim
02
A year of AI for $980. The proprietary route runs $3,600.
Blended annual cost of a mid-size AI workload.
ModelledCostModelled. Arithmetic on stated inputs. The inputs are named on the card.Method Inputs are published list prices [C]; the workload profile is ours and is stated on the method page.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
$980OwnAI + OwnCloud
$3,600Proprietary APIs
73%Less
03
◆Five paid assistants stopped us inside six hours. OwnAI has no stop to hit.
Hours of sustained normal use before hitting a stop.
Measured on hardwareCostMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method Paid tiers at published rates, identical prompt sequence, five runs each, August 2026. Vendors anonymised; run log in the data room.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
0OwnAI stops
4 hMedian time to a stop
5Assistants tested
04
Five subscriptions cost $110 a month. One $25 subscription does all five jobs.
The common five-tool stack against a single subscription.
From published sourcesCostFrom published sources. Read from public pricing pages, contracts or documentation.Method Prices from public pricing pages [C]. The comparison only holds if OwnAI covers all five jobs — tested in the quality benchmark below.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
$110Current stack
$25OwnAI
77%Less
05
◆The 'cheap' model really costs 2.7× its sticker once retries are billed.
Headline price against true cost once retries are billed.
Measured on hardwareCandourMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method Our blind eval set, run against each model in the same month. Retries are billed and counted. The eval run count is pending publication — it is not the Own 1 figure.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
+167%Cheapest model's real cost
+9%OwnAI routing
5Models compared
06
◆Read the terms: of six providers, only one leaves your prompts yours.
What each provider's terms of service permit.
From published sourcesPrivacyFrom published sources. Read from public pricing pages, contracts or documentation.Method Read from published terms of service, August 2026 [C].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
6 of 6OwnAI
3 of 6Best alternative
5Providers read
07
◆We beat the frontier baseline in five task classes — and publish the two we lose.
Score against the frontier baseline, by task class.
Measured on hardwareCandourMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method Blind pairwise eval across seven task classes. The score is a keyword heuristic, not a human rating — that limitation matters and is stated. Run count pending publication.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
5 of 7Classes we claim
2 of 7Classes we lose
58%Our worst
08
◆Our retry rate is 8.5%. The frontier's 4.8% is better — at four times the price.
Share of runs needing at least one retry to pass the eval.
Measured on hardwareCandourMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method From the same blind eval set. Frontier's 4.8% is genuinely better than our 8.5% — at roughly four times the cost per job.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
8.5%OwnAI
4.8%Frontier, and better
62.4%Cheapest model
09
First token in 0.9 seconds, and our slowest moment beats their best by 5×.
Time to first token, 10th to 99th percentile, 2,000 requests per provider.
Measured on hardwareSpeedMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method 2,000 requests per provider spread across a week, same prompt, same region.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
0.9 sOur p50
2.2 sOur p99
5.2×Best rival p99
10
44 models across seven modalities — every one actually benchmarked.
Models actually benchmarked, by modality.
Measured on hardwareCoverageMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method Every model here has at least twelve completed runs. Models with fewer were excluded rather than reported thinly.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
44Models run
7Modalities
12Minimum runs per model
11
Six platforms, six kinds of output: only one answers yes to all six.
Native capability to produce each kind of output.
From published sourcesCoverageFrom published sources. Read from public pricing pages, contracts or documentation.Method Read from public product documentation, August 2026 [C]. Native capability only.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
6 of 6OwnAI native
4 of 6Best rival
6Platforms
12
◆Model, provider, region, cost visibility — we score 8 of 8. The best rival, 3.5.
What you control, rather than what the model can make.
From published sourcesOwnershipFrom published sources. Read from public pricing pages, contracts or documentation.Method Read from public documentation, August 2026 [C]. Enterprise-only capability counts as partial.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
8 of 8OwnAI
3.5 of 8Best rival
8Control questions
13
◆Run it on your desk, your servers, or our cloud. The desk needs no network at all.
Three execution zones, and what changes between them.
Our framework, our scoringPrivacyOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method Product architecture as published [C]. Execution zone, model, provider, region, infrastructure type and retention policy are all user-selected rather than fixed — this is a description of the control surface, not a measurement of it.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
3Execution zones
1 of 3Needs no network
7Settings the user controls
14
◆We scored ourselves 10, 9.3 and 9.2 — and put the 8.5 on the chart too.
Four self-assessed scores, ordered so the lowest is not buried.
Our framework, our scoringCandourOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method Self-assessed against published criteria. User savings is our lowest score and is shown last deliberately.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
10.0Highest
8.5Lowest, and shown
9.3Mean
05 · OwnOS · Beta
An agent with a wallet.
The first operating system where an agent has to ask before it spends your money — enforced by the system, not by a policy document.
13.2 sPower-on to first tokenDemo data, not measured
Models, agents, wallet, rollback — built into the OS, not bolted on.
The six capabilities built into the system rather than bolted on.
Our framework, our scoringOwnershipOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method A description of the architecture. The capability matrix is where these claims are tested against rivals.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
6System primitives
100%Open source
12+Built-in models
02
◆Told to spend $80, every other setup paid. OwnOS declined, ten out of ten.
Money the same scripted agent managed to spend without asking.
Measured on hardwareOwnershipMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method Live sandbox transaction of $80. OwnOS blocked all ten attempts; the other four completed all ten.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
0 of 10OwnOS completions
40 of 40Elsewhere
$3,200Spent testing
03
◆Fresh install, no config: other systems let agents spend, delete and email. OwnOS allows none.
What an agent can do unasked, by operating system.
From published sourcesOwnershipFrom published sources. Read from public pricing pages, contracts or documentation.Method Read from public documentation and default installs, August 2026 [C].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
0 of 6OwnOS allows unasked
6 of 6Desktop + agent
6Capabilities
04
◆Power button to first AI token in 13.2 seconds — 3× ahead of Windows. Demo data.
Power-on through desktop and model load to the first generated token.
Demo data, not measuredCandourDemo data, not measured. Illustrative only. Has not been run as a controlled benchmark.Method This is demo data and is labelled as such on the chart and in the legend. It has not been run as a controlled benchmark. Until it has, it is illustrative only.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
13.2 sOwnOS target
42.3 sWindows 11
3.2×Claimed gap
05
◆Nine agent-era capabilities: OwnOS has all nine. The best rival scores 4.5.
Built for agents, or built for desktop apps.
From published sourcesOwnershipFrom published sources. Read from public pricing pages, contracts or documentation.Method Read from public documentation and default installs, August 2026 [C].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
9 of 9OwnOS
4.5 of 9Best rival
9Capabilities
06
◆Thirty days unattended: 12,888 actions, 41 asked permission, zero went rogue.
Thirty days of continuous agent operation, by outcome.
Measured on hardwareOwnershipMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method One agent, thirty days, real workload. The audit log is downloadable.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
12,888Total actions
0.37%Needed a human
7Hard blocks
07
27 GB free after boot — enough headroom for a model the others can't fit.
Free memory on a 32 GB machine five minutes after boot.
Measured on hardwareCostMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method Measured with each system's own reporting after a five-minute settle. Three boots each, median reported.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
27.1 GBOwnOS free
+4.7 GBvs next best
13BExtra model that fits
08
The whole OS takes under 2 GB. Windows takes 9.4.
Memory the operating system holds after boot, before any model loads.
ModelledCostModelled. Arithmetic on stated inputs. The inputs are named on the card.Method OwnOS figure is the published boot footprint [A]. Rival figures are derived from the post-boot measurements above and are indicative rather than directly measured.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
under 2 GBOwnOS
9.4 GBWindows 11
4.7×Lighter
09
Permission prompts land in 0.8 seconds — fast enough that nobody turns them off.
Agent request to approval prompt on screen, p10 to p99.
Measured on hardwareSpeedMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method Sampled across the thirty-day run — 41 real approval events for OwnOS, synthetic load for the others.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
0.8 sOur median
4.6 sOur p99
41Real approval events
10
A scoped, budgeted agent in five minutes. Building the same yourself: forty hours.
Clean machine to an agent running with spend, file and network scopes set.
Measured on hardwareSpeedMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method Three timed attempts per environment. A first-timer took materially longer everywhere.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
5 minOwnOS
9×Faster than next
40 hBuild it yourself
11
◆One request, four steps, $1.24 spent — every step inside limits the owner set.
A real four-step run: what moved, what it cost, and where approval sat.
Our framework, our scoringOwnershipOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method The published one-request walkthrough [C]: budget $3.00, approval threshold $2.00, requested amount $1.24. A description of the mechanism, not a benchmark of it — the thirty-day audit card above is where the mechanism is measured.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
$3.00Budget the owner set
$1.24What the agent spent
0Files uploaded in step one
12
◆Four security layers shipped. The fifth says 'planned', in public.
The security ladder, with its one unfinished rung shown.
Our framework, our scoringCandourOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method Implementation status as published [C]. Remote attestation is listed as planned rather than implied as done — shipping-state candour is the benchmark here.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
4 of 5Implemented
1Publicly marked planned
TPM2Hardware-bound keys
13
Agent safety, scored: 5.0. The nearest alternative manages 2.8.
Our agent safety framework applied to five environments.
Our framework, our scoringOwnershipOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method Our criteria, our scores, published.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
5.0OwnOS
2.8Best alternative
5Criteria
14
◆The only audit log a stranger can verify: signed, exportable, complete.
What each environment's audit log actually contains.
From published sourcesCandourFrom published sources. Read from public pricing pages, contracts or documentation.Method Read from public documentation [C]. The signed row is the only one an auditor cares about.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
6 of 6OwnOS
2.5 of 6Best alternative
12,888Entries in sample
06 · Own 1 · Beta
Thirty-one days. Eight hundred and twelve runs.
A machine on your desk that got 138% faster while it sat there, and never sent a byte anywhere.
Own 1 · beta unit
812Measured runsPlus 276 video runs, held separately
◆Text generation got 138% faster in 31 days. The hardware never changed.
All 492 text, code and agent runs on the device, with the best-so-far line.
Measured on hardwareSpeedMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method 492 runs across chat, coding and agent workloads, 23 May – 17 June 2026 [B]. Grouped by model, quantisation and engine and never averaged across them. Power limits dominate: the same machine at 28 W measures roughly half the throughput it does at 65 W.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
+138%Since day one
91.7Best tok/s
492Runs plotted
02
◆An image took 7.6 minutes. It now takes 9.9 seconds — 46× faster.
All 226 image runs on a log axis — the only axis where a 46× drop is visible.
Measured on hardwareSpeedMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method 226 runs, 20 May – 20 June 2026 [B]. Log scale because on a linear axis a 46× improvement collapses into the baseline and cannot be read at all.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
−98%Time per image
9.9 sBest
46×Faster
03
◆Voice began slower than real speech. It now runs 12.5× faster than realtime.
Realtime speedup over 31 days — audio seconds produced per wall second, measured on the Own 1 itself. Above the dashed line is faster than speech.
Measured on hardwareSpeedMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method 94 runs across five voices and four harnesses, 30 May – 12 June 2026 [B]. Day one ran below realtime, which is why the reference line is on the chart.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
+1,931%Since day one
12.50×Best realtime
0.62×Day one
04
◆Eighteen models raced on speed and quality. Eight are worth picking — and the fastest scores 9% on quality.
Every benchmarked model plotted on generation speed against answer quality.
Measured on hardwareCandourMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method 18 models scored on the same prompts [B]. A model is on the frontier when nothing else is both faster and better. The fastest model scores 9.4% on quality — speed alone is not a result.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
8 of 18On the frontier
87.1%Best quality
66.7Fastest, at 9.4% quality
05
◆Text, image, voice, agents: every workload got dramatically faster in a month.
Day one against best so far, each row scaled to its own range.
Measured on hardwareCandourMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method From the 812-run programme, 20 May – 27 June 2026 [B]. Gains come from optimisation work, not hardware changes.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
46×Best gain, image
+1,931%Voice speedup
31Days
06
◆812 measured runs: 492 text and code, 226 image, 94 voice.
What the 812 measured runs actually cover.
Measured on hardwareCoverageMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method 812 measured runs by modality, from the published export [B]. Video is excluded from this count because it carries no throughput metric, not because it failed.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
492Text, code, agents
226Image
94Voice
07
◆880 of 1,088 runs on one engine — and engines are never averaged together.
Runs by execution harness across the whole programme.
Measured on hardwareCandourMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method 1,088 records across 13 harnesses [B]. Results are grouped by harness and never averaged across them, which is why the tail is shown rather than folded in.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
13Harnesses
880Largest, llama.cpp Vulkan
3Smallest
08
Thirty-five models tested. Not one cherry-picked winner.
The twelve most-run models in the programme.
Measured on hardwareCandourMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method 35 distinct models across 1,088 records [B]. The two best performers are highlighted so they can be checked against the full distribution rather than quoted alone.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
35Models run
162Most-run model
12Shown here
09
◆One hour of AI work: cloud assistants sent up to 486 MB. Own 1 sent zero bytes.
Bytes leaving the network interface during one hour of identical work.
Measured on hardwarePrivacyMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method tcpdump at the interface, one hour, same prompts, three repetitions. The capture file is downloadable from the method page.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
0 KBOwn 1
486 MBWorst cloud case
3Captures per environment
10
75 watts flat out — an eighth of a gaming PC doing the same work.
Sustained wall power during identical inference work.
Measured on hardwareCostMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method Plug meter at the wall, one hour sustained load per machine, sampled every second, median of three runs.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
75 WOwn 1 sustained
14 WIdle
8×Gaming PC draw
11
◆Shared by ten people it costs $400 a head — against $1,320 a year in subscriptions.
Hardware cost divided by the number of people sharing it.
ModelledCostModelled. Arithmetic on stated inputs. The inputs are named on the card.Method Hardware at $4,000 including the replacement fund. Ten concurrent users is the tested ceiling [A]; five is the configuration we recommend.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
$400Per person at ten
$4,000At one person
10×Sharing advantage
12
◆Pays for itself in 6.8 months at five people. At one person it's 39, and we say so.
Months to payback by team size.
ModelledCandourModelled. Arithmetic on stated inputs. The inputs are named on the card.Method Hardware plus measured electricity against the subscription stack it replaces, at that team's usage. No residual value assumed. The 39-month single-user figure is published because it is the weak one.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
6.8 moAt five people
39 moAt one person
3Minimum we recommend
13
◆Text, code, audio and embeddings fly. Image and video don't — both are on the chart.
Throughput by modality, p10 to p90.
Measured on hardwareCandourMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method From the 812-run programme, later runs only so the optimisation is settled. Minimum 40 runs per modality [B].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
91.7Text tok/s median
2 of 6We don't claim
40Minimum runs
14
◆72% less energy and carbon than the cloud — which is still cheaper per task, and we show it.
Energy, carbon and cost for a thousand tasks.
ModelledCandourModelled. Arithmetic on stated inputs. The inputs are named on the card.Method Electricity at $0.15/kWh, Victorian grid intensity 0.79 kg CO₂e/kWh. Cloud figures from published PUE and regional intensity [C].Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
−72%Energy
−72%Carbon
+2,500%Cost — we lose
15
◆Nine capabilities built in. On a Mac Studio you assemble seven of them yourself.
Nine capabilities, and what the obvious alternative actually offers instead.
From published sourcesOwnershipFrom published sources. Read from public pricing pages, contracts or documentation.Method Read from public documentation [C]. 'Install it yourself', 'Manual orchestration' and 'Time Machine restore' are real capabilities — they are stated rather than scored as absent, because the difference is whether the OS provides it, not whether it is possible.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
9 of 9Built into Own 1
2 of 9Built into macOS
34Providers it can burst to
16
◆One desk machine: ten AI processes, 99 TOPS, 8 TB — and $0 per token.
The hardware and the runtime, on one card.
Measured on hardwareOwnershipMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method On-device specification for the beta unit [A]. Ten people per machine is the tested sharing ceiling, not the recommended configuration — five is what we recommend, and the payback benchmark uses five.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
10AI processes at once
99TOPS
$0Per-token cost
17
One box: 99 TOPS, eight agents, eight terabytes, ten people.
The capacity of a single machine.
Measured on hardwareCostMeasured on hardware. We ran it on a machine and the raw runs are in the export.Method On-device figures from the benchmark programme [A]. Ten concurrent users is the tested ceiling; five is the recommended configuration.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
99TOPS
2 + 8Models and agents
$0Per-token cost
18
Two models and eight agents cover twelve job roles, around the clock.
What two models and eight agents add up to.
Our framework, our scoringCostOur framework, our scoring. Our criteria, scored by us, published so it can be argued with.Method The two-plus-eight capacity is measured [A]; the twelve-role figure is our framing and is presented as positioning.Sources [A] Supplied provider and coverage data · [B] Recovered from the published O1-BETA benchmark export · [C] Public pricing pages, contracts and documentation, read August 2026.Reading it Hover any bar, dot or cell on the chart for its exact value. ◆ marks a benchmark that measures ownership rather than performance.
12Job roles
24/7Availability
$0Per-token cost
How to read this. Every benchmark states which kind of evidence it rests on,
and gives its sample size, window and what was held constant. The five kinds are not
interchangeable and are never mixed on one card.
Measured on hardware
We ran it on a machine and the raw runs are in the export.
From the live index
Counted or captured by the index. Not a hardware measurement.
From published sources
Read from public pricing pages, contracts or documentation.
Modelled
Arithmetic on stated inputs. The inputs are named on the card.
Demo data, not measured
Illustrative only. Has not been run as a controlled benchmark.
Our framework, our scoring
Our criteria, scored by us, published so it can be argued with.
Benchmarks marked ◆ measure ownership: exit cost, candour, bytes
leaving the device, and what survives when a vendor changes its mind. Competitor names are
withheld where we have not sought permission; every anonymised row is identified in the data
room. Where we lose, the chart says so — the two task classes below our claim threshold, the
13% duty-cycle crossover, the frontier model that beats our retry rate, the 39-month
single-user payback, and the cost line on Own 1 energy are all published rather than
omitted.
