Last Look in 2026: What Symmetric Means, What Asymmetric Hides, and Why It Still Matters

Last look is the part of FX market structure that everyone agreed should die a decade ago and that, in 2026, is still everywhere. The premise is simple: when a liquidity provider streams a quote and a taker hits it, the LP gets a brief window — typically tens of milliseconds — to decide whether to honour the price. The premise has been defended, deplored, regulated, refined, and renamed. It has not been removed.

This post is not an argument for or against. It is an argument that the question 'does this venue use last look' has been the wrong question since at least 2018. The right questions are about hold window length, rejection symmetry, and information use.

Why last look exists in the first place

A streaming LP commits to a price for a notional size at a moment in time, knowing the price might be stale before the taker hits it. The window between price publication and taker arrival can be tens of milliseconds — long enough for a real move on a fast tape. Without some recourse, the LP either has to (a) post wider quotes to absorb that latency risk, or (b) lose money systematically to takers who arrive only when the LP's price has gone stale in their favour.

Option (a) makes the screen uglier. Option (b) is unsustainable. Last look is option (c): the LP keeps the right to check, in a defined window, whether the price is still consistent with current market data before honouring the trade. The compromise is real even if the implementation has been routinely abused.

Symmetric last look

Symmetric last look means the LP applies the same tolerance band, in the same direction, to its check, regardless of which way the market moved during the hold. If the price moves against the LP by more than X, the LP rejects. If the price moves IN FAVOUR OF the LP by more than X, the LP also rejects — passing the surprise back to the taker.

This is the model the FX Global Code endorses. It is also the model that, by 2026, every credible institutional LP claims to operate. The differentiator is no longer 'do you do symmetric' — the differentiator is whether the hold window itself is short and whether the symmetry holds under stress conditions, not just calm ones.

Asymmetric last look (and its quieter cousins)

Asymmetric last look rejects when the price moves against the LP and fills when the price moves in the LP's favour. Anyone reading this knows that is not a thing they want to trade against. It is also extremely rare in 2026, in name.

What is less rare are the quieter cousins of asymmetric last look that achieve the same economic outcome through other mechanisms. Variable hold-window length, conditional on direction. Pre-trade credit checks that happen to take longer when the move was against the LP. Multi-stage routing where a parent LP's last look is propagated through to a child LP, and each stage has its own checks. The taker experiences something that looks superficially like a fill-rate problem; the underlying behaviour is asymmetric in everything but its label.

If you want a deeper read on why this kind of structural information leak is precisely the cost component that retail-grade TCA reports never surface, see Decomposing Execution Cost.

Hold window length

The hold window is the time the LP has between receiving the taker's request and either filling or rejecting. In 2018 some institutional LPs were operating with hold windows of 100 ms or more. In 2026 a hold window of more than 30 ms on a major pair is, with rare exceptions, a tell that the LP either has poor pricing infrastructure or is using the window for something other than a price check.

Drovix operates with a fixed hold window of under 5 ms on majors, under 15 ms on crosses, and zero on the wholesale book for known institutional counterparties. We publish the per-pair distribution. A taker who cares can ask for the histogram; it does not change in private.

What the venue does with the information

This is the question that is rarely asked because it is rarely answerable. When an LP rejects a trade after a 30-ms hold, it has just learned something: someone, at this venue, wanted to buy or sell this size of this pair at this moment. Even if the trade did not happen, the information about the attempted trade can be re-used by the LP, by an affiliate, or by the LP's pricing model. The taker's intent has leaked, the LP's book is now informed, and the next price the taker sees may not be coincidentally worse.

The Code addresses this with explicit prohibitions on pre-hedging during the hold window. The enforcement is operationally hard. The honest answer for an institutional taker is to evaluate it by outcome: track the move in the mid in the 30 seconds after a reject, in the direction of the attempted trade, and average it across hundreds of rejects. If the number is meaningfully non-zero, the information is leaking, regardless of what any policy document says.

Practical evaluation for an institutional taker

• Ask for the hold-window distribution per pair, not the headline average. The tail of the distribution is where the bad behaviour lives.
• Ask for the rejection rate split by which side moved during the hold. A symmetric venue has roughly equal rejection rates on both sides; an asymmetric one does not.
• Ask for the post-reject mid-move statistic in your direction. If the venue will not provide it, compute it yourself from your own reject logs.
• Treat 'we don't do last look' as a marketing statement, not a structural one. Many venues that say this route through LPs that do.

Drovix's stance

Drovix uses last look on the streaming book under the symmetric model the Code defines. We publish hold-window distributions per pair, rejection-rate split by direction, and post-reject mid statistics on demand. We do not pre-hedge during the hold window, and our wholesale book — used by known institutional counterparties — operates without last look on standard sizes.

That stance is not unique. What is increasingly unique is the willingness to publish the numbers that let a counterparty verify it.

Where to go next

→ FIX Tags That Decide Fill vs Re-Quote — the protocol-level mechanics of how a last-look rejection actually reaches your OMS, and which tags to read.

→ Signed Flow and Information Half-Life — why the cost of information leak depends on how fast your specific strategy ages, and how to measure your own decay.