Fastest EVM swap quote API
Wall clock p50 latency of GET /quote per provider, measured every minute on a 6 pair basket across 5 chains.
TL;DR. As of , ParaSwap posts the lowest quote latency at 205 ms (p50, 24h) across 8 ranked providers. Source: OpenChainBench, https://openchainbench.com/benchmarks/evm-quote-latency.
EVM swap aggregator APIs all promise sub second quotes. This benchmark measures how often they actually deliver one. The harness rotates a 5 pair basket across Ethereum, Base, Arbitrum, BSC and Robinhood Chain, fires a quote request at every supported provider in parallel every 60 seconds, and records the wall clock round trip. The leaderboard sorts by p50 latency over the last 24 hours, lower is better. A geographic split is also available, with the same probe running independently from us east, eu west and Singapore so a wallet integrator can pick the provider whose edge gateway is closest to their backend.
Methodology
Each provider's public quote endpoint is hit on six fixed pairs (1 ETH to USDT on Ethereum, 1000 USDC to USDT on Ethereum, 0.5 ETH to USDC on Base, 100 USDC to ETH on Arbitrum, 1 BNB to USDT on BSC, 0.1 ETH to USDG on Robinhood Chain), rotating one pair per 60 second tick. Per (provider, chain) we land roughly 240 samples in 24 hours, enough to put p90 and p99 buckets on a solid base. Quote quality (output amount, net of gas) is not scored, only latency.
- Six fixed pairs across five chains. Polygon and Optimism will be added in v2 once they reach a comparable sample volume.
- 60 second tick, one pair per tick, round robin. Every (provider, chain) sees a quote every 6 minutes.
- Three regions (us east, eu west, sgp). Each is a separate Railway service writing the same metric family with a region label, so the page can filter to a single edge or aggregate across all three.
- Latency is wall clock from request dispatch to last byte received, observed only on the happy path (HTTP 2xx and a parseable output amount). Failures land on dedicated counters (auth, throttle, no route, other) and pull the success gauge to 0.
- Native asset sentinel is the canonical 0xEeeeeEeee form. Uniswap uses 0x0000... and Bebop substitutes the wrapped equivalent on the sell side. The adapter normalises per provider.
- Provider auth: three providers need a key (Mobula in an Authorization header, Route as a Bearer token, Uniswap in a vendor header), and KyberSwap, Bebop, LI.FI, CoW, ParaSwap, Nordstern and OpenOcean answer a bare request. The key is not what buys the low end of this table.
Results: 9 providers measured, 8 ranked by Quote latency (p50, 24h)
| № | Providers | p50 | p90 | p99 | Success | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | ParaSwapTrades as Velora, no key | 205 ms | 350 ms | 854 ms | 100.00% (n=4320) | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 2 | KyberSwapNo auth, per chain aggregator engine | 222 ms | 359 ms | 888 ms | 100.00% (n=4320) | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 3 | NordsternNo key, no registration, 121 chains | 248 ms | 420 ms | 791 ms | 99.94% (n=3598) | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 4 | UniswapTrading API, self-service key | 408 ms | 669 ms | 924 ms | 100.00% (n=3600) | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 5 | MobulaAPI key required, dedicated edge | 502 ms | 784 ms | 1.13 s | 100.00% (n=3600) | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 6 | LI.FI30+ chain coverage, same chain via fromChain equal toChain | 703 ms | 1.25 s | 2.91 s | 62.03% (n=2677) | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 7 | BebopRFQ gasless, EIP 55 checksum required | 755 ms | 1.14 s | 1.56 s | 100.00% (n=3600) | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| 8 | CoW ProtocolNo auth, RFQ batch auction, no BSC | 1.78 s | 2.57 s | 3.48 s | 100.00% (n=2880) | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Measured, not ranked | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| not ranked | RouteRobinhood onlyRobinhood Chain only, routes through Kyber and Nordstern | 408 ms | 538 ms | 693 ms | 99.9% (n=719) | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Frequently asked
Which EVM swap quote API has the lowest latency right now?
ParaSwap currently leads at 205 ms (p50 over the last 24 hours) across all chains and regions, across 9 measured providers. The leaderboard re sorts every minute on fresh Prometheus samples, so the answer reflects measured latency on the live basket, not a marketing claim.
Does this measure quote quality or just speed?
Speed only. Latency is wall clock round trip from request to last byte, recorded only on a successful parseable response. Output amount in USD is not scored on this bench because providers measure it differently (Bebop is net of fees on a gasless RFQ, others are gross), and converting all outputs to a comparable USD value introduces an oracle bias we do not want to bake into a latency leaderboard.
Why are some providers not on every chain?
Each provider declares its supported chain set. The harness skips probes for chains a provider does not claim to cover so the failure counter is not inflated by structural non coverage.
What happens when a provider rate limits the harness?
Each adapter classifies the response. 429 lands on a throttle counter and the success gauge drops to 0. 401 and 403 land on an auth counter. Empty or zero output amounts on a 200 land on a no route counter. None of those failure modes contribute to the latency histogram, so the p50 number stays representative of healthy quotes. Read the Success column with that in mind: it is this harness's success rate, not the provider's reliability. On 2026-10-19 LI.FI read 62% because 1,602 of its 4,320 daily probes were throttled by the free-tier budget we probe it on, which says what our access buys and nothing about what a paying integrator sees.
How is the geographic split implemented?
Three identical Railway services run the harness with different MONITOR_REGION env vars (us east, eu west, sgp). Each writes the same metric family with a region label. The bench page filter on region selects which slice of the histogram is queried.
Source code github.com/ChainBench/OpenChainBench/tree/main/harnesses/evm-swap-quoting