| Category | Function | Share | Rank | Negative rate | Labels | Quadrant | Since September 2026 |
|---|---|---|---|---|---|---|---|
| Object storage | Cloud and infrastructure | 19% | 2 of 51 | 17% | 47 | accepted challenger | ▲+1Since September 2026: 21% → 21%, +1 point. Inside the 11-point floor: within noise. Read over the models both editions asked. |
| Backup and disaster recovery | Cloud and infrastructure | 0% | 49 of 116 | 0% | 1 | under 10 labels · led by Veeam Data Platform at 38% | newNew since September 2026: not ranked then, 0% now. |
| Telemetry pipelines | IT operations and endpoint | 0% | 87 of 124 | 0% | 1 | under 10 labels · led by OpenTelemetry Collector at 37% |
| Category | September 2026 | Now | Change | Reading | Rank |
|---|---|---|---|---|---|
| Object storage | 21% | 21% | ▲+1Since September 2026: 21% → 21%, +1 point. Inside the 11-point floor: within noise. Read over the models both editions asked. | Within noise | Rank 2 of 51, unchanged |
| Backup and DR | 0% | 0% | newNew since September 2026: not ranked then, 0% now. | New this edition | Rank 49 of 116, unchanged |
| Telemetry pipelines | 0% | 0% | newNew since September 2026: not ranked then, 0% now. | New this edition | Rank 87 of 124, unchanged |
Shares here are read over the models both editions asked, so they can differ by a point or two from the standing above, which counts every model in this edition.
The floor is 11 points of share, measured: how far the models move a leader on their own when the same questions are asked twice with nothing changed. A larger change is movement; a smaller one is noise, and both are shown. Movement is read over the twelve models both editions asked; GPT-6 Luna, Muse Glimmer 30B joined this edition and are in the standing but not yet in the comparison. How the floor is measured · The editions
| Model | First choice | Alternative | Mention | Negative | Labels |
|---|---|---|---|---|---|
| Claude Haiku 4.5 | 0 | 0 | 0 | 1 | 1 |
| GPT-5.4 mini | 2 | 1 | 0 | 0 | 3 |
| Gemini 3.5 Flash | 5 | 1 | 1 | 0 | 7 |
| Perplexity Sonar | 0 | 0 | 0 | 1 | 1 |
| Grok 4.20 | 0 | 4 | 1 | 1 | 6 |
| Mistral Small | 0 | 0 | 0 | 1 | 1 |
| DeepSeek V4 Flash | 1 | 4 | 0 | 1 | 6 |
| Llama 4 Maverick | 1 | 0 | 0 | 0 | 1 |
| Qwen 3.7 Flash | 0 | 3 | 0 | 0 | 3 |
| Kimi K2 | 2 | 1 | 0 | 1 | 4 |
| GLM 4.7 FlashX | 0 | 4 | 0 | 0 | 4 |
| MiniMax M2.5 | 1 | 1 | 0 | 1 | 3 |
| GPT-6 Luna | 0 | 4 | 0 | 0 | 4 |
| Muse Glimmer 30B | 1 | 3 | 0 | 1 | 5 |
Verbatim evidence the judge attached to positive labels.
“Heavy Egress / Global Apps | Cloudflare R2 | Best for serving media, web assets, or SaaS app files directly to users globally with zero egress fees.” Gemini 3.5 Flash · Object storage · scale prompt · first choice
“Best for Active SaaS Applications & Global Read-Heavy Workloads: Cloudflare R2 ... is the premier choice.” Gemini 3.5 Flash · Object storage · direct prompt · first choice
“The best overall option in this category is Cloudflare R2, especially for read-heavy workloads” Muse Glimmer 30B · Object storage · comparative prompt · first choice
“Top Recommendation: Backblaze B2 or Cloudflare R2 (depending on your workload)” DeepSeek V4 Flash · Object storage · paraphrase prompt · first choice
Verbatim evidence attached to negative labels. A warning on a product with few labels is a warning; on a product with many, it is one voice among them.
“No versioning, no object lock, and no WORM compliance, so it is not suitable for immutable or regulatory backups.” Perplexity Sonar · Object storage · negative prompt · hard negative
“Multiple outages in 2026 (13 incidents in 8 days in Aug), including R2 failures, 503 errors, and unrepaired data” Grok 4.20 · Object storage · negative prompt · hard negative
“13 separate incidents in just 8 days... the frequent outages raise concerns about production reliability” Kimi K2 · Object storage · negative prompt · hard negative
“has no versioning and no object lock, making it unsuitable for immutable backups or regulatory compliance” Claude Haiku 4.5 · Object storage · comparative prompt · soft negative
Citations exist only for the models that return a source list, five of the fourteen in this edition, so these counts come from 131 of the 135 answers that named Cloudflare R2 and are not a share of its labels.
430 of the 461 domain citations in answers naming Cloudflare R2 came from somebody else's page.
Pages are listed as the models cited them.
Search figures are US estimates from DataForSEO, read September 28, 2026; AI search demand is its modeled, directional estimate, not a count of queries to any assistant. The answers are this edition's. Two measurements side by side: neither is read as the cause of the other.
An email the morning each edition publishes: where this product moved, where it held, and by how much against the noise floor. One address, confirmed by a click; a stop link in every email.
Already following? Everything you follow, with a stop for each.
What Cloudflare R2's own pages state, read September 28, 2026: blog.cloudflare.com/r2-ga, developers.cloudflare.com/r2, blog.cloudflare.com/r2-open-beta, developers.cloudflare.com/r2-sql, blog.cloudflare.com/r2-local-uploads. A claimed page can correct any of them.
Claiming is free and changes nothing in the data. A claimed page shows a verified contact who is told when each edition publishes and when Cloudflare R2's standing changes by more than the noise floor; the right to propose corrections to the vendor table, meaning names the judge wrote that should or should not read as Cloudflare R2, applied by version and listed in the change log; and a one-line description supplied by the vendor and marked as such.
A new claim receives the current edition's vendor brief for Cloudflare R2 by email, built from the raw record of the edition. It shows:
Cloudflare R2 is among the products AI models recommend first in Object storage this edition. Each badge says so in the buyer's words, names the edition, and links to the standing. The next edition issues a new badge; this one stays true as a record of October 2026.
Badges read on light and dark pages. Alt text carries the claim, the category, the edition and the index, so it stays a citation where the image does not load.