Spatial transcriptomics has gone from a technique you read about to one your committee expects you to have considered. The problem is that "which platform should I use?" has no context-free answer, and most of the comparisons you will find are written by people selling one of them.
Here is the version that starts from your experiment rather than from the instrument.
First, be sure you need spatial at all
Spatial data is expensive, slow, and considerably harder to analyse than the alternatives. It earns its cost when the position of a cell is part of your hypothesis: tumour margins, niches, layered tissue, cell-cell interactions that only happen at a boundary.
If your question is which cell types are present and what they express, dissociated single-cell RNA-seq answers it with better sensitivity per cell, mature analysis tooling, and a fraction of the headache. Plenty of projects buy spatial data and then answer their actual question with the cell-type proportions, which scRNA-seq would have given them more cheaply. Be honest about which one you are.
The real axis of choice
Platforms are usually presented as a resolution ranking. That is the wrong frame. The genuine trade-off is three-way, and you cannot currently have all three at once:
- Resolution. Are you measuring a spot containing several cells, or individual transcripts inside one cell?
- Plex. How many genes? A few hundred to a few thousand in a targeted panel, or the whole transcriptome?
- Cost and throughput. How many samples can you afford, which usually decides whether the study is powered at all.
Every platform is a different answer to that trade-off. Choosing well means knowing which of the three you can least afford to give up.
Sequencing-based: whole transcriptome, binned space
Visium HD covers the whole transcriptome, roughly 18,000 human genes, across a continuous lawn of 2 x 2 µm barcoded squares with no gaps, about 11 million of them per capture area (10x Genomics). It is usually described as single-cell scale, and that phrase deserves unpacking before you budget around it.
At 2 µm, most squares contain very few reads at normal sequencing depth, so that output is realistically for visualisation. The recommended unit for clustering and differential expression is the 8 µm bin, which aggregates 4 x 4 adjacent squares. Recovering actual cells from those bins is an open analysis problem with its own dedicated methods, such as bin2cell. So you get unbiased whole-transcriptome coverage and near-cellular spatial precision, but cell boundaries are inferred rather than observed.
One practical constraint that catches people out: validated whole-transcriptome probe sets are human and mouse only. If you work in another organism, this decision is already made for you.
Imaging-based: real cells, chosen genes
Xenium, CosMx and MERSCOPE image transcripts directly in intact tissue, so you observe subcellular locations and genuine cell morphology rather than inferring them. The cost is that you measure only the genes on your panel, and you have to choose that panel before you see any data.
They differ in chemistry, which is why their failure modes differ. Xenium uses padlock probe ligation with rolling circle amplification. CosMx uses cyclic hybridisation of barcoded probes with no enzymatic amplification. MERSCOPE commercialises MERFISH, using combinatorial error-correcting barcodes.
Choosing a panel is the underrated hard part. A panel that omits a marker you need later cannot be rescued without running the samples again, and you will not know what you needed until analysis. Budget thinking time for this, and talk to someone who has run the panel on your tissue type.
What the benchmarks actually say
There is now real head-to-head data rather than vendor claims. A comparison of imaging platforms on FFPE tumour samples found Xenium produced higher transcript counts per matched gene with much lower background, while CosMx offered a larger standard panel and detected genes that the smaller Xenium panels missed entirely.
The more important finding is the one that resists summarising: results were specific to the assay and the tissue tested, not a universal ranking. Sample age and fixation mattered. A platform that wins on fresh tissue may not win on a decade-old archival block.
This is genuinely useful, because it tells you what to do: ignore league tables and ask whether anyone has published on tissue like yours, preferably with a similar fixation history.
What it costs, and why nobody will tell you
Neither 10x nor its resellers publish list pricing for spatial reagents or instruments. The only real numbers are the pass-through rates academic cores publish, which vary widely by institution and by whether you prepare samples yourself. As one public anchor, the University of Missouri genomics core lists Visium HD at just over $6,000 per slide.
Treat any single number as an anchor, not a quote. Ask your own core for their current rate, and ask specifically what is excluded: sequencing is often billed separately, and analysis almost always is.
A shorter version
- Position is not part of your hypothesis: use scRNA-seq and spend the difference on more samples.
- You do not know which genes matter yet, and you work in human or mouse: sequencing-based, whole transcriptome, analyse at 8 µm bins.
- You know your genes and you need true single-cell boundaries: an imaging platform, and spend real effort on the panel.
- Either way, find published data on tissue resembling yours before you commit, because that predicts your result better than any benchmark.
The most common expensive mistake in this field is not picking the wrong platform. It is running three samples on a beautiful platform when the question needed twelve.