New frontier AI models got faster at spatial biology, but accuracy stayed flat

In SpatialBench results, GPT-5.5 and Opus 4.7 got faster, but accuracy remained nearly unchanged. GPT-5.5 cut average runtime almost in half compared to GPT-5.4, but accuracy stayed at 57.65% versus 57.44%, and Opus 4.7 came in at 52.41%, effectively similar to Opus 4.6's 52.83%.

The benchmark consists of 159 spatial biology analysis tasks, covering real platforms such as Xenium, Visium FFPE, MERFISH, TakaraBio Seeker, and AtlasXomics DBiT-seq. Each task starts from a real analysis state and requires recovering a specific biological outcome, with answers compared against expert-reviewed ground truth.

By platform, GPT-5.5 improved on Visium, Xenium, and MERFISH, but fell behind on TakaraBio and AtlasXomics. Opus 4.7 also led by 11.1 percentage points on Xenium, but was similar or lower on other platforms.

Expert-reviewed analysis paths revealed recurring failure patterns.

• Treating barcodes, beads, and spots as independent replicates, causing pseudoreplication
• Applying default scRNA-seq normalization directly to targeted spatial panels
• Clustering without donor/timepoint integration, misreading batch structure as biological structure
• Miscounting beads or markers as cells, misinterpreting spatial units
• Failing to recover de novo niches and tissue states

The examples are clear-cut. In AtlasXomics's SPATIAL10_genome_wide_de_pct, where the expected value is about 1.2%, the models judged 92-94% of genes as sex-differential. In MERFISH's norm_02_myelin_gene_coexpression_normalization, where the expected Spearman correlation between Mbp and Plp1 is about 0.308, GPT-5.5 and the Opus series produced about -0.16. The cause was applying sc.pp.normalize_total(adata, target_sum=1e4) and sc.pp.log1p(adata) directly to a 374-gene targeted panel. In TakaraBio Seeker's oocyte_count_per_timepoint, where the expected immature oocyte count is 850, the models overestimated at 1,510-3,463. In spatial biology, more than speed, judgment that reflects platform-specific context remains the bottleneck.