How the storage layer changes agent performance.

The same agent finished 74% faster and cost 81% less on Fastio than on Dropbox.

Fastio was tested head to head against Box, Dropbox, Google Drive and OneDrive: an identical copy of the same 2,163 files stored natively in each, the agent working through each provider's own connector. At the medians of 3 runs, Fastio reached a complete answer in 9:10 for $10.14, against 16:34 to 35:00 and $23.76 to $54.48 through the other four. It can also sit over the storage you already run, as an accelerator for those four rather than a replacement.

We asked one question: does the connector between an agent and its files change how the agent performs? To find out we ran the same agent, Claude Opus 5 at medium effort in Cowork, on the same prompt against an identical 2,163-file corpus stored in Fastio, Box, Dropbox, Google Drive and OneDrive, and varied nothing else. Each provider ran the task 3 times.

What we found. On the contract renewal review, Fastio reached a complete answer 45% faster than Box, 57% faster than Google Drive, 71% faster than OneDrive and 74% faster than Dropbox, at 57% to 81% lower cost than the same four. Every run was scored against an independently validated answer key, so what the page measures is the time and cost of reaching a checked answer.

The same agent ran the same prompt against the same 2,163 files in all five. Only the storage connector varied.

What these figures are. Every figure on this page is the median of 3 runs per provider, taken 13 to 15 September 2026, and every median is published with the range those 3 runs covered. The range is part of the figure: on some providers it is wide, and the tables and charts show it.

Why the connector matters. An agent working on files spends most of its time finding, opening and reading documents. Each step is a call through the connector, and each call adds latency and tokens. A connector that returns search results and extracted text gives the agent a shorter path to the answer than one that returns a file list and raw bytes.

One agent, five connectors, three runs

The contract renewal review asks the agent to build a renewal calendar from the whole corpus: every renewing agreement, and for each one its renewal date, its notice deadline and whether that deadline has passed, plus the status of the NDA. Each provider ran it 3 times. Each row below is one provider: a small dot for each run, the filled dot on the median.

Time. Fastio's median run finished in 9:10, with its 3 runs falling between 8:10 and 9:14. Box's median was 16:34, Google Drive's 21:30, OneDrive's 31:44 and Dropbox's 35:00.

Cost. Fastio's median task cost $10.14 against $23.76 for Box, $27.63 for OneDrive, $45.83 for Google Drive and $54.48 for Dropbox.

What sits underneath. Tool calls and documents read are where the work differs most. At the median Fastio made 38 tool calls and read 13 documents; the other four made 497 to 782 calls and read 156 to 511 documents, which is 13x to 21x the calls Fastio needed.

At the median Fastio made 38 tool calls and read 13 documents. The other four made 497 to 782 calls and read 156 to 511 documents.

Medians with their ranges, every provider:

Fastio relative to each provider. The same gap can be read from either side, so both are given. Against Box, Fastio was 45% faster and 57% cheaper, or Box was 81% slower and cost 134% more. Against Google Drive: 57% faster and 78% cheaper, or 135% slower and 352% more. Against OneDrive: 71% and 63%, or 246% and 172%. Against Dropbox: 74% and 81%, or 282% and 437%.

How we tested

Harness. Every session ran with Claude Opus 5 at medium effort in Cowork as the main agent, and the provider's connector as the only permitted tool. Each provider ran the task 3 times, 13 to 15 September 2026, giving the fifteen runs behind this page. The prompt text was identical except for the sentence naming the storage location. Only the storage connector varied.

Corpus. The corpus, calloway_v2_messy, is 2,163 files for a fictional industrial-materials company, with an identical copy loaded into each provider. It carries the structure the task turns on: agreements that were later restated, so an agent that stops at the first document it finds will put the wrong date in the calendar.

Fastio's setup. Fastio ran on MCP v0.48 with workspace intelligence enabled, so each document was processed once when it was added: a search index, extracted text and extracted metadata, with OCR for the scanned files. Processing a set of documents carries a one-time ingestion fee, paid once for that set.

The task. The agent was asked for a renewal calendar covering the whole corpus:

Every agreement that renews, with its renewal date, its notice deadline and whether that deadline has passed, and the status of the NDA, working only from the connected storage.

Scoring. The answer key was validated independently of the runs and covers all 9 renewing agreements in the corpus. Each run is scored on the listing, the renewal date, the notice deadline and its passed flag, and the NDA status, and an agreement listed that should not be counts against the run. Speed is only useful if the answer is right, so the time and cost on this page are the time and cost of reaching a checked answer.

Counting tool calls. A tool call is every tool invocation by the agent and by any subagent it spawned, connector and non-connector alike. That includes bookkeeping calls such as tool lookups and task-list updates, and it includes calls that failed. Fastio's median of 38 includes about ten bookkeeping calls; Box's 782 includes 764 connector calls spread across seven subagents.

Fastio's connector accepts several requests in one call, while the other four take one request per call. This page counts calls. Counted as individual requests instead, Fastio's three runs were 24, 52 and 32, a median of 32, and the comparison lands in the same place: 92% to 96% fewer either way.

Failed calls are part of the count, and provider rate limiting pushes the competitor totals up: the median run failed 29 calls on Box, 68 on Dropbox, 38 on OneDrive, none on Google Drive and 2 on Fastio. Excluding failures moves no median by more than 10% and changes no provider's rank.

Metrics. Time to answer is wall clock from prompt to a complete answer, end to end: harness duration from prompt submitted to the answer finished, including model generation, connector latency and waiting. Tool calls are the count of tool-use blocks in the turn, main agent plus any subagents. Documents read are the distinct documents the agent opened through the connector. Input tokens count uncached, cache-write and cache-read tokens across all models. Cost is the harness total in US dollars at list rates, which is model spend only: it carries no connector, ingestion or indexing charge, and no provider's subscription or seat licenses. Published figures are the median of the 3 runs, and every median is given with the range those runs covered.

What this means for your files

What the test establishes. One agent, one prompt, one 2,163-file corpus, 3 runs per provider. The connector was the only thing that changed, so the gap that opened is the connector's: at the medians of those 3 runs, 45% faster than Box, 57% faster than Google Drive, 71% faster than OneDrive and 74% faster than Dropbox.

The storage layer is not neutral. It sets how much work the agent has to do.

Why the gap is that wide. Through Fastio the agent asked for search results and extracted text and received them, because each document had been processed once when it was added. The other four connectors returned a file list and raw bytes, so the agent read the corpus itself, on every run. That is what separates 38 tool calls at the median from 497 to 782, and it is why the same task cost 57% less than through Box and 81% less than through Dropbox, again on medians of 3 runs.

What that is in practice. For anyone running agents against a company's files, this is the same task finished sooner and for less money, on the documents they already have, because the agent takes a shorter path to the answer.

It does not ask you to move. Box, Dropbox, Google Drive and OneDrive stay where they are and keep doing what they do. What changes is the connector the agent reaches them through. This is a layer added over the storage you run, not a replacement for it.

What we would measure next. These figures are one task on one corpus. The question worth answering is what the same change does to your own work, and that is a far shorter test than this one.