How the storage layer changes agent performance.

Fastio makes the files you already keep in Box, Dropbox, Google Drive or OneDrive faster for agents. With the same agent, the same prompt and the same 211 files, only the connector changed. Time to a complete answer fell by as much as 64%.

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 in Claude, on the same prompt against an identical 211-file corpus stored in Fastio, Box, Dropbox, Google Drive and OneDrive, and varied nothing else.

What we found. On the multi-document customer relationship audit, Fastio reached a complete answer 35% faster than Dropbox, 50% faster than Box, 54% faster than Google Drive and 64% faster than OneDrive. It made 52 to 83% fewer tool calls, used 30 to 54% fewer input tokens and cost 19 to 37% less per task. Fastio reported 11 of the 12 ground-truth facts and handled all 5 traps; Google Drive reported all 12 and handled 4.

What these figures are. Every figure on this page is from a single run per provider on 9 September 2026: 15 fresh sessions, one per provider per test, each wave fired within about fifteen seconds per provider. Repeat runs and larger corpora are scheduled and will be added as they complete.

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, extracted text or single pages gives the agent a shorter path to the answer than one that returns a file list and raw bytes.

One agent, five connectors, one run

The multi-document customer relationship audit asked the agent to prepare a renewal brief for one customer from about 20 documents, using only the storage connector. The chart below shows the same run four ways. Pick a metric to compare providers on it.

Time. Fastio finished in 2m 50s. Dropbox took 4m 24s, Box 5m 43s, Google Drive 6m 10s and OneDrive 7m 48s.

Tool calls and files. Fastio made 29 tool calls and opened 18 files. The other four made 61 to 167 calls and opened 47 to 109 files. This is where the largest differences sit: 52 to 83% fewer calls and 62 to 83% fewer files opened.

Tokens and cost. Fewer calls and fewer files mean less text through the model. Fastio used 2.37M input tokens against 3.38M to 5.11M, and cost $3.06 against $3.75 to $4.83.

Here is the full run, every metric, every provider:

Fastio relative to each provider. Against Box: 50% faster, 83% fewer tool calls, 30% fewer input tokens, 19% lower cost. Against Dropbox: 35%, 75%, 48% and 29%. Against Google Drive: 54%, 52%, 35% and 19%. Against OneDrive: 64%, 76%, 54% and 37%.

Corpus size. One size, 211 files, has been measured in the runs to date (9 September 2026). Larger corpora are scheduled; the table below gains a row as each run completes, and it cannot show a trend until it has more than one.

Single-document lookup. The single-document contract lookup asks for two facts from one contract. All five providers ran it on 9 September 2026. Fastio answered in 31.2 s with 4 tool calls; OneDrive took 45.8 s and 7 calls, Dropbox 54.9 s and 15, Google Drive 55.2 s and 10, Box 86.7 s and 23. That is 64% faster than Box, 43% faster than Dropbox, 43% faster than Google Drive and 32% faster than OneDrive, with 83%, 73%, 60% and 43% fewer tool calls and 75%, 64%, 54% and 40% fewer input tokens. Cost was $0.88 for Fastio: 34% lower than Box at $1.33, 22% lower than Dropbox at $1.13, 7% lower than Google Drive at $0.95, and 10% higher than OneDrive at $0.80. All five answers were correct, all five caught the draft-MSA trap, and there were no connector errors and no fabrications.

Accuracy across the five connectors

Speed is only useful if the answer is right, so every brief was scored against the answer key. Accuracy was close to parity across the five connectors. Fastio reported 11 of the 12 ground-truth facts and handled all 5 traps. Google Drive reported all 12 facts and handled 4 traps; it got the 61-plus aging comparison wrong. Box reported 11 facts and handled all 5 traps, and made one fabricated claim, ten years of bank statements. Dropbox reported 10 facts and handled 4 traps; it could not read the scanned credit memo through its connector. OneDrive reported 11 facts and handled 3 traps; it could not read the scanned credit memo either.

Precision was 97.9% for Fastio, 93.8% for Box, 98.0% for Dropbox, 97.9% for Google Drive and 95.8% for OneDrive. Box was the only provider with a fabricated claim.

Connector behaviour. No provider made a failed connector call in the published runs, and no agent went outside its connector. OneDrive was the slowest provider on the audit, with 2 documents unreadable through its connector and 3 of 5 traps handled.

How we tested

Harness. Every session ran in Claude in Cowork, the desktop app, with claude-opus-5 as the main agent. The published figures come from 15 fresh sessions on 9 September 2026, one per provider per test. Each test was fired as one wave, with the five providers started within about fifteen seconds of each other. The prompt text was identical per test except for the sentence naming the storage location. Session event logs were pulled from the code-sessions API and scored against the corpus answer key. Only the storage connector varied between sessions.

Corpus. The corpus, calloway_synthetic_messy_v1, is 211 PDFs for a fictional industrial-materials company: legal and finance documents dated 2023 to 2026, with an identical copy uploaded to each provider. It is deliberately messy. It contains 10 exact duplicates, 5 stale draft versions with one term changed, 12 image-only scanned PDFs with no text layer, and 24 misfiled documents. Within it, 5 traps bear directly on the tasks: a draft MSA with a 30-day notice period instead of 60; a draft invoice with a different amount; the credit memo stored as an image-only scan; a duplicate copy of one SOW; and the amendment filed in the wrong folder.

Tasks. We ran 2 tasks. The single-document contract lookup asks for two facts in one contract:

What is the termination notice period in our MSA with Bexley Construction Partners, and how long is the initial term?

The multi-document customer relationship audit touches about 20 documents across legal and finance folders. The agent was preparing for a renewal call with a customer and was asked for:

A complete picture of the relationship: every agreement in force and its key terms, total SOW value, total invoiced and credited, and any payment problems, working only from the connected storage.

The audit also carried a hard tool-scope rule in the prompt:

Use only the storage connector. No memory, no other tools, no code. Say so if something is not in storage.

The answer key for the audit holds 12 ground-truth facts: the MSA term and dates, the 60-day notice, Net 30 payment terms, the amendment and what it changed, the NDA and its status, three SOW values, four invoice amounts, the credit memo, the invoice-to-PO-to-payment chain, and the certificate of insurance. All 5 providers ran the audit and the lookup on 9 September 2026.

Metrics. Time to a complete answer is wall clock, 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. Files opened are the distinct documents the agent read through the connector. Total tokens are input plus output across all models, where input counts uncached, cache-write and cache-read tokens. Cost is the harness total in US dollars at list rates. Coverage is how many of the 12 facts the answer reported. Traps are how many of the 5 planted traps were handled, where handled means the brief used the correct version or disclosed the problem. Precision verifies every checkable factual claim against the PDFs. Fabrications are confident invented claims, counted separately.

What the numbers cannot say yet

Single run per provider. Every figure on this page, for the audit and for the lookup, is from one run per provider, all on 9 September 2026 under the same conditions. Nothing is averaged. Repeat runs will be added as they complete. Until then, every figure on this page is a single observation.

Cost carries a cache-write effect. Fastio fired first in each wave and wrote the roughly 68k-token prompt prefix to cache; the other four providers read it, and reads bill below writes. That lowers their metered cost by about $0.20 to $0.30 each relative to Fastio. Token counts are unaffected.

The corpus is synthetic and small. 211 files in the runs to date (9 September 2026). Results on a real multi-gigabyte drive have not been measured. Larger corpora are scheduled.

Memory was excluded by prompt. Cowork memory persists across sessions and providers. The audit prompt told the agent not to read from or write to memory, and that instruction is what kept memory out of the measured runs. Earlier runs did touch memory.

One client. Every session ran in Claude; no other client has been measured.

Data and sessions

Each figure on this page comes from one named session. These are the ten published session ids behind the comparison table and the lookup. Earlier sessions and the OneDrive-only SOW revenue review are in the CSV as history and are not compared. The per-run data is available as a CSV.

Download the per-run data (CSV)