Faster, verified, private: measured
Stock anon-rpc against Atmos, from the cloud. The same wallet-shaped workload runs in each setup: plain RPC, the stock anon-rpc Nym worker, and Atmos on top of that same worker, with and without verified reads and a constant-rate schedule. The ideal is a truly anonymous RPC: as fast as plain RPC, every answer proven, and nothing learned about who, when or what.
Faster
Speed details ↓5.5–6.6× faster page load than stock anon-rpc
Ten parallel reads, median per region: Atmos 2.7–3.6 s, stock anon-rpc 14.7–22.0 s, plain RPC 0.19–0.60 s.
- Against plain RPC, a call costs about 0.5–0.6 s more: the mixnet round trip.
- The slowest 5% of steady calls are 3–5× faster than stock.
- A single typical call is only 1.2–1.7× faster; the gain is in parallel reads.
10-read burst, Atmos v1 in the browser harness · Cloud Run, 3 regions × 4 sessions · 2026-11-06 · bench:v9 (94f519b)
Verified
Verification details ↓7 methods proven against a light client; stock anon-rpc proves none
For a proven method, a lying provider or exit can cause a failure, never a wrong answer.
eth_blockNumbereth_getBalanceeth_getTransactionCounteth_getCodeeth_getStorageAteth_calleth_getTransactionReceipteth_chainIdeth_sendRawTransactioneth_estimateGaseth_getBlockByNumbereth_getBlockByHasheverything else
proven checked, not proven: a local constant, and a transaction hash that must match partly: the gas figure and a block's transaction list are the provider's sent plainly, and counted
- A verified balance costs about 0.5–0.9 s more than an unverified one.
- The latest block number takes 0 ms: it is answered from the verified head.
The SDK and the host · Sepolia sync-committee light client · in the benchmark, eth_call, eth_estimateGas and receipts passed through unchecked
Private
Privacy details ↓52% chance is 50%: an observer of outgoing requests can't tell busy from idle
With shaping on, from the request count. From bytes it was 60%, not significant.
- Who Hidden from the providerIt sees a Nym exit, not your IP. Only the Nym directory and your entry gateway see it.
- When Hidden only with shaping, which is opt-inOff by default; turn it on with shape: true or ATMOS_SHAPE=1. Outgoing requests then leave on a fixed clock. Replies are not padded.
- What Not yetProviders still see which accounts you query. The access-list mode (experimental) hides repeat reads while it keeps pace.
atmos-host, local sessions (2–4 hosts), 12 s windows, best single-threshold observer · outgoing requests only, replies not measured · not the cloud run
The first answer is slower. Atmos took 17–71 s to its first answer across setups and regions, against 16–33 s for stock anon-rpc. Why
Faster
Four sessions per arm on publicnode and mevblocker, one image for every arm, with the arm order rotated per task. Measured 2026-11-06.
Results, europe-west1
| Arm | Sessions | 10-read burst | Steady p50 | Steady p95 | First answer | Success |
|---|---|---|---|---|---|---|
| Direct fetch (not anonymous) | 4 | 185 ms 104 ms – 244 ms | 36 ms | 56 ms | 122 ms | 100% |
| Stock anon-rpc, Nym worker | 4 | 17.9 s 12.0 s – 24.0 s | 915 ms | 3.2 s | 15.9 s | 100% |
| Atmos v1, browser harness | 4 | 2.7 s 2.2 s – 3.3 s | 538 ms | 1.0 s | 26.7 s | 100% |
| Atmos v1, Node harness | 4 | 2.7 s 1.2 s – 2.7 s | 554 ms | 889 ms | 19.3 s | 100% |
| Atmos with verified reads | 4 | 2.7 s 2.0 s – 4.0 s | 0 ms | 1.3 s | 36.2 s | 100% |
| Verified reads, as the self-hosted endpoint runs them | 4 | 2.9 s 2.5 s – 3.8 s | 0 ms | 1.5 s | 29.3 s | 100% |
| Atmos with a constant-rate schedule | 4 | 7.7 s 6.8 s – 8.5 s | 2.0 s | 2.5 s | 16.9 s | 100% |
Burst values are the median across sessions, with a bootstrap 95% interval underneath. Steady latency is per call during the polling phase. "First answer" runs from creating the worker to its first successful reply, boot included.
- Direct fetch (not anonymous)
- Stock anon-rpc, Nym worker
- Atmos v1, browser harness
Show as a table
| Read | Direct fetch (not anonymous) | Stock anon-rpc, Nym worker | Atmos v1, browser harness |
|---|---|---|---|
| 1 | 165 ms | 4.81 s | 2.67 s |
| 2 | 165 ms | 10.31 s | 2.67 s |
| 3 | 166 ms | 10.78 s | 2.67 s |
| 4 | 166 ms | 11.36 s | 2.68 s |
| 5 | 166 ms | 12.30 s | 2.68 s |
| 6 | 168 ms | 12.40 s | 2.68 s |
| 7 | 170 ms | 13.82 s | 2.68 s |
| 8 | 170 ms | 14.06 s | 2.68 s |
| 9 | 172 ms | 15.23 s | 2.68 s |
| 10 | 174 ms | 17.13 s | 2.68 s |
A call with no answer after 24 s is cut off. In this run the stock worker's median session answered all ten burst reads before the cut-off, one after another, over 17 s; in earlier runs most of them hit it.
The staircase is the cost of parallel calls through one stock worker. Its transport keeps a single pooled connection per origin, so every extra concurrent call opens a fresh TCP and TLS connection through the mixnet, one after another. Atmos sends the same ten reads as one JSON-RPC batch on a connection that is already warm.
Three regions
The final run: every arm on one image, Cloud Run in three regions, 4 sessions per arm per region,
10-minute sessions. Each cell is the burst median, then steady p50 and the share of calls
answered. A steady p50 of 0 ms for verified reads is eth_blockNumber answered from the
verified head.
| Region | Direct | Stock anon-rpc | Atmos v1 | Atmos, Node harness | Verified, self-hosted |
|---|---|---|---|---|---|
| europe-west1 | 185 ms 36 ms · 100% | 17.9 s 915 ms · 100% | 2.7 s 538 ms · 100% | 2.7 s 554 ms · 100% | 2.9 s 0 ms · 100% |
| us-east1 | 271 ms 85 ms · 100% | 14.7 s 709 ms · 100% | 2.7 s 603 ms · 100% | 1.5 s 579 ms · 100% | 3.9 s 0 ms · 99% |
| asia-south1 | 596 ms 210 ms · 100% | 22.0 s 1.2 s · 100% | 3.6 s 841 ms · 100% | 2.1 s 874 ms · 100% | 5.8 s 0 ms · 100% |
When a route stalls, Atmos switched to its standby 14.7–18.3 s later, and each session lost only the one call in flight (run 4).
Against direct RPC, per method
The parity run: steady calls only, 4 sessions per arm per region. Each cell is p50 / p95. Atmos adds about 0.5–0.6 s per call, which is the mixnet's round trip, and a verified balance adds about another 0.5–0.9 s for its proof. The latest block number is answered from the verified head.
| Region | Method | Direct | Atmos v1 | Verified |
|---|---|---|---|---|
| europe-west1 | eth_blockNumber | 37 ms / 64 ms | 542 ms / 1.6 s | 0 ms / 1 ms |
| europe-west1 | eth_getBalance | 37 ms / 57 ms | 584 ms / 1.6 s | 1.2 s / 3.9 s |
| us-east1 | eth_blockNumber | 121 ms / 152 ms | 672 ms / 2.0 s | 0 ms / 1 ms |
| us-east1 | eth_getBalance | 124 ms / 175 ms | 674 ms / 1.6 s | 1.1 s / 3.4 s |
| asia-south1 | eth_blockNumber | 161 ms / 270 ms | 769 ms / 2.5 s | 0 ms / 1 ms |
| asia-south1 | eth_getBalance | 162 ms / 271 ms | 767 ms / 2.7 s | 1.7 s / 4.9 s |
In the benchmark, eth_call, eth_estimateGas and receipts pass through
the verified arm unchecked; the self-hosted endpoint runs eth_call locally on proven
state, which is measured separately.
Time to the first answer
This is where Atmos is slower than the stock worker. Before its first answer, Atmos loads the Nym directory, tests a gateway, and checks the route with a real call; verified reads also sync a light client. Medians per region (europe-west1, us-east1, asia-south1):
| Stock anon-rpc | 15.9 s · 16.8 s · 32.8 s |
| Atmos v1, Node harness | 19.3 s · 35.9 s · 54.2 s |
| Verified reads, as the self-hosted endpoint runs them | 29.3 s · 52.2 s · 71 s |
Across all Atmos arms and regions, the first answer took 17–71 s, against 16–33 s for the stock worker.
Browser or Node
The same client, with the anonymity worker in headless Chromium or in Atmos's Node harness (a V8 isolate, no browser). Run 7 adds the pinned worker hash and the worker's host allowlist: 8 of 8 boots passed under both.
| 10-read burst | Steady p50 | Steady p95 | Answered | |
|---|---|---|---|---|
| Chromium (atmos-run5) | 2.8 s | 522 ms | 746 ms | 100% |
| Node harness (atmos-run5) | 3.4 s | 572 ms | 759 ms | 100% |
| Chromium, pinned hash and allowlist (atmos-run7) | 2.5 s | 570 ms | 1.4 s | 100% |
| Node harness, pinned hash and allowlist (atmos-run7) | 4.0 s | 656 ms | 1.7 s | 100% |
Optional settings
Four sessions each. None beat the defaults by more than the spread between sessions, so all stay off in the client; the self-hosted endpoint turns on rotation (every 10 minutes) because it limits how long one route's timing lasts.
| 10-read burst | Steady p50 | Steady p95 | Answered | |
|---|---|---|---|---|
| Atmos v1, defaults | 2.6 s | 542 ms | 918 ms | 100% |
| Hedged reads | 2.5 s | 579 ms | 1.0 s | 100% |
| Route rotation every 2 minutes | 2.6 s | 631 ms | 1.1 s | 100% |
| More reply SURBs | 2.4 s | 616 ms | 1.1 s | 100% |
| 15 s exit connect timeout | 2.6 s | 687 ms | 1.1 s | 100% |
Earlier reference: run 2
Six sessions per arm per region, 10-minute sessions, sessions on flashbots left out. A stock burst of 24.0 s means the median session hit the 24 s cut-off.
| Region | Direct | Stock anon-rpc | Atmos v1 |
|---|---|---|---|
| europe-west1 | 195 ms 37 ms · 100% | 24.0 s 1.2 s · 96% | 2.3 s 746 ms · 100% |
| us-east1 | 326 ms 81 ms · 100% | 18.8 s 1.2 s · 100% | 3.3 s 826 ms · 100% |
| asia-south1 | 662 ms 154 ms · 100% | 24.0 s 2.1 s · 93% | 3.0 s 1.1 s · 99% |
Earlier measurements
Before this benchmark, the same effect was measured in a desktop browser with live Nym workers. Those runs used a home connection and few samples, so read them as direction, not as final figures.
| Pattern, 10 reads on one worker | Wall time (3 rounds) |
|---|---|
| 10 concurrent calls (how wallets issue them) | 8.5 / 9.1 / 9.9 s |
| 10 sequential calls | 6.7 / 6.8 / 7.0 s |
| One JSON-RPC batch | 1.06 / 1.20 / 1.45 s |
| Two 5-read batches on two workers | 0.83 / 0.89 / 0.99 s |
Verified
Verified reads
Every read checked against a light client, in europe-west1. Refreshing the verified head on a fixed clock, and reusing proofs within a block, made the burst about four times faster.
| 10-read burst | Steady p50 | Steady p95 | Answered | |
|---|---|---|---|---|
| Before (atmos-run3) | 13.6 s | 1.8 s | 5.9 s | 100% |
| After (atmos-run6) | 3.5 s | 0 ms | 1.4 s | 100% |
p50 is 0 ms because most steady calls are eth_blockNumber, answered from the verified head; verified reads alone are 0.9–1.4 s.
Private
What an observer can tell
An observer of the self-hosted endpoint's outgoing traffic tried to tell 12-second windows of wallet activity from idle ones, with the best threshold chosen with hindsight. 50% is chance. This covers outgoing requests only: replies are not padded, and incoming traffic has not been measured.
| From request count | From bytes | From head checks | |
|---|---|---|---|
| Shaping off | 98.6% | 98.6% | 51.4% |
| Shaping on | 52.1% | 60.4% | 50% |
Method
The workload
One session takes about ten minutes, on one RPC provider chosen per session:
- Boot. Create the worker; time to
ready, then to the first successful call. - Page-load burst. Ten reads at once: eight
eth_getBalance, oneeth_blockNumber, oneeth_getTransactionCount. - Steady polling.
eth_blockNumberevery 12 s and a balance every third tick. Three times per session, oneeth_call, oneeth_estimateGasand one receipt lookup. - Write. One unfunded Sepolia transaction from a fresh key. The expected answer
is
insufficient funds, so the write path is measured without moving value.
How sessions are compared
- The unit is the session, not the call, since calls within a session are correlated.
- Arms are interleaved in each region, with the order rotated per task, so drift over time and the cost of a cold start hit every arm equally.
- The direct arm is each region's own floor. Latency added by the anonymity network is reported against it, within the same region.
- Only keyless public RPCs are used, at a few requests per second per provider at most.
This run
- Run atmos-final, 2026-11-06; sessions ran 21:50–23:09 UTC. Executions: europe-west1
atmos-final-zpgmf, us-east1atmos-final-5m8dl, asia-south1atmos-final-bzqh9. - Image
bench:v9, built from commit94f519b. Cloud Run, 2 vCPU and 4 GiB per task. - 4 sessions per setup per region, 10-minute sessions, the setup order rotated per task. Providers publicnode and mevblocker (keyless); a call with no answer after 24 s is cut off.
- The unit is the session: medians, with bootstrap 95% intervals.
- Every session, 84 in all (3 regions × 7 setups × 4): CSV · JSON with the run details.