Pathway raises $30M at $500M for post-Transformer AI models
Two stories about AI economics this morning: one lab just landed a hefty seed valuation for betting that reasoning doesn't have to be expensive, and a respected benchmarking outfit launched a tool for testing whether models actually fit your work.
Pathway, an AI research lab betting that reasoning doesn't have to be expensive, has raised a $30 million seed round at a $500 million valuation to scale its "post-Transformer" BDH architecture. The company's flagship result so far: a 150-million-parameter model called BDH-CQ scored 29.5% on the public ARC-AGI-1 evaluation set at a calculated cost of about $0.0007 per task — roughly 0.85 seconds of H200 GPU time — which Pathway says pushes past the benchmark's cost-versus-accuracy frontier.
The bet is architectural. BDH, short for "Dragon Hatchling," performs its reasoning in a continuous latent workspace rather than generating long chains of thought as text — and text-based reasoning is exactly what makes token-by-token inference so pricey. Pathway's argument is that some of AI's cost and memory problems are baked into the Transformer design itself, not something more compute will fix; it's a direct challenge to the brute-force scaling orthodoxy, and a familiar theme for readers of our Deep Dive — Power, not GPUs, now sets the pace of AI. Backers include Id4 Ventures, TQ Ventures, Red Bridge Ventures, Kadmos Capital, Wilson Sonsini's investment arm, and Databricks chief AI scientist Jonathan Frankle, with Transformer co-inventor Łukasz Kaiser on a newly formalized advisory group. The capital buys GB300 compute and bigger BDH models — early pretraining runs up to 600 billion parameters reportedly show "Transformer-like" scaling. The real test is whether latent reasoning stays cheap and accurate when the models stop being small.
Artificial Analysis has launched Optima, a platform that lets teams build custom benchmarks from their own data and compare models on quality, cost per task, and time per task. Users can upload evaluation datasets or agent traces, or simply describe a use case and let Optima generate the test cases; scoring is rubric-based or pairwise. It targets a known failure of general-purpose benchmarks — scores that swing with implementation details and rarely predict real-world fit, the same gap behind our coverage of PerceptionBench: no frontier model breaks 60% on basic vision. Early testers built benchmarks for finance and accounting agents to find models that cut costs by a factor of ten without major quality loss, according to Artificial Analysis. It's a pragmatic fix — though a custom benchmark is only as sound as the person who designs it, and the field's deeper methodological problems don't disappear.
Would a model that reasons for a fraction of a cent per task change how you build agents? Tell us in the comments.
