Base's July 2025 switch to 200 ms flashblocks cut competitive arbitrage bots' priority fees by 59% relative to activity-matched controls. The savings came chiefly from opportunities with several searchers. Winning an uncontested opportunity cost almost as much as before.
What changed in Base's auction?
Base's sequencer orders transactions by priority fee per gas. Before July 2025, a 2 s block effectively gave searchers one sealed-bid first-price auction for position. A searcher spotting a pool dislocation submitted an arbitrage transaction with a fee bid. The highest bid landed first; losing transactions reverted and still paid for gas burned before failure. Flashblocks place ten 200 ms preconfirmation sub-blocks within the same 2 s canonical block, locking each sub-block's order as it is produced. Searchers now face ten sequential auctions. Winning the top position means beating arrivals in one 200 ms window rather than across the block.
The change matters most to searchers, who pay most of Base's priority fees. Decentralized-exchange transactions made up one fifth of Base's transaction count, yet paid 70% of priority fees in June 2025 and 58% in August. Anchuri, Felten and Mamageishvili relate the expected fee response to the value at stake. In their diffusion-plus-jump price model, an opportunity is worth κσ²B + J, where B denotes the window length. Shortening B reduces the diffusion component. A discontinuous price move can still create an arbitrage opportunity in a short window, leaving the jump term J and a fee floor.
The authors use on-chain data from a Base archive node for June 2025 (pre) and August 2025 (post), excluding July. Their searcher group contains 3,032 addresses above the 99.9th-percentile activity threshold of 405 transactions, each with a revert rate of at least 0.20 and a maximum declared bid of at least 0.011 gwei. The 8,053 controls pass the same activity threshold and fail both searcher markers. Since flashblocks have no on-chain marker, the authors infer sub-block boundaries from upward fee jumps between different senders in a block. They estimate the headline effect with a two-way fixed-effects difference-in-differences on an address-day panel.
How firm is the 0.187 gwei cut?
The question is what happens to searcher bids when Base changes from one 2 s auction round to ten 200 ms rounds while keeping the fee mechanism itself. Searchers' effective priority fee falls 0.187 gwei relative to controls (SE 0.038), or 59% of its 0.317 gwei pre-period mean. The raw difference in changes is 0.168. The authors attribute the gap to fixed-effects weighting in an unbalanced panel. Declared bids fall 0.228 gwei, compared with 0.183 before weighting. Across the chain, mean fees move from 0.038 to 0.025 gwei. That chain-wide change also contains the July-to-August market shift and cannot identify the flashblock effect.
Position in the block gives a clearer signature. C measures the share of a group's priority-fee value paid in the first tenth of block gas. Searcher C falls from 0.98 to 0.28; control C moves from 0.07 to 0.09. The resulting DiD is -0.72. June's pre-trend is flat (p = 0.45), although the authors say their coarse group-hour panel has low power. One auction piles high bids near the start of the block. Ten auctions distribute them through it.
The fee pre-trend is also flat (p = 0.785), and the event study shows a step down at launch. June supplies only one month of pre-period data. More awkwardly, the main screen uses both months, allowing post-launch behavior to determine whether an address qualifies. Screening on June alone yields 0.120 gwei for the fee and 0.165 for the bid, respectively 64% and 72% of the main estimates. The authors say this removes post-period selection. Mean reversion remains possible because high June bids help select treated addresses and low ones help select controls.
Control fees present another limitation: they average 0.001 gwei across both months, against 0.127 for searchers. An equal-percentage shock would be hard to see through additive day effects. The authors point to comparable annualized volatility in June and August, at 66% and 70%, and to fees stepping down at the July launch without earlier drift between the groups. They acknowledge that an equal-percentage trend remains difficult to check.
I take the direction as settled and the size as 0.120 to 0.187 gwei. The revert-rate estimate (+0.021) fails its pre-trend at p < 0.001; the authors present it descriptively.
Six-searcher fees: 11.4 to 1.0 gwei, with uncontested wins near 0.17 gwei
For a desk, the split between contested and uncontested wins matters more than the average. An uncontested winning fee moves from 0.177 gwei at 2 s to 0.164 at 200 ms. With six or more searchers on a pool, it drops from 11.446 to 1.027 gwei. Most opportunities had little competition even before the change: 81.4% of June pool-block opportunities involved one searcher, compared with 78.0% in August.
The authors read the split similarly. Their abstract separates an approximately invariant floor from a competitive premium, and their conclusion describes compression at 200 ms as first-order, with diminishing gains below that window length. They qualify the evidence. The 7.9% difference between 0.177 and 0.164 has no test, and β(c) partly reflects opportunity value: larger dislocations attract more searchers. Their value-controlled check uses gas used as a proxy and covers only the 200 ms data.
Speed enters the ordering
In August, 766,510 captures had a linked reverting attempt. A higher-fee attempt reverted one reconstructed slice after a lower-fee winner in 5.4 to 6.6% of those cases. Weighting by the winner's fee value reduces the figure to 1.9 to 2.7%, placing these late-arrival losses mainly among lower-fee wins. The authors treat both ranges as upper bounds because slippage and stale state also cause reverts. Their pool attribution covers 96% of core searchers' reverted fee value in August 2025, though they say that coverage does not bound misassignment.
The inferred timing has its own limit. Reconstruction produces 19.5 slices per block against ten actual flashblocks, so a revert one reconstructed slice behind a winner might belong to the same window. Among higher-fee losers within two slices of the winner, representing 9.8 to 11.9% of contested captures, the median bid edge was 0.02 to 0.03 gwei. Even on these upper-bound counts, a few hundredths of a gwei in extra bid can lose to earlier arrival. The value of latency remains unpriced.
Retries multiplied.
Across all senders, reverts per block rose from 17.6 to 25.1. Reverts per inferred contest fell from 17.6 to 1.3. That per-contest measure depends on inferred slices, which overcount flashblocks, and the revert response is not identified.
Below 200 ms?
Arbitrum One voted in September 2026 to adopt a PGA with a 125 ms round, adjustable from 25 to 250 ms. The question of fees below 200 ms therefore has a prospective user. The authors offer two projections. Their measured-competition model takes the winning fee from 0.181 gwei at 200 ms to 0.169 gwei at 50 ms, toward the 0.164 floor; it also reproduces 0.316 gwei at 2 s. A simpler thinning model projects a 33 to 68% reduction in the bid premium at 50 ms. The two Poisson-thinning shapes producing the 68% end require an expected 6 to 92 competitors per opportunity in a 2 s window. The measured figure is 1.56. The authors say measured competition supports only the power shape, which gives the 33% end.
They call both projections rough bounds. The measured-competition model holds a searcher's bid against c rivals fixed across window lengths. The authors themselves flag that their 2 s versus 200 ms table contradicts this assumption, making the projection illustrative rather than exact. Still, with 0.14 expected rivals per 200 ms window, there is little competitive premium left to squeeze. I favor the flat projection. A PGA chain operating below 200 ms with a falling uncontested floor would change my mind.
We did not run this ourselves. Reproducing it requires Base transaction-level ordering, gas-auction fields and decoded pool swaps. Our data stops at one-minute price bars.