Pathway's 150M model just hit 29.

A 150M-parameter model scoring 29.5% on ARC-AGI-1 while costing only $0.0007 per task is a massive wake-up call for anyone obsessed with just scaling parameter counts. Pathway just dropped BDH-CQ, and the efficiency numbers are frankly absurd. We've spent the last few years believing that the only way to solve complex reasoning tasks was through massive LLM agents or grueling, token-heavy chains of thought that eat up compute by the second. BDH-CQ flips that on its head by moving away from the standard Transformer constraints.

The technical shift here is the move toward recurrent memory and latent reasoning. Instead of the model outputting a long string of intermediate "thinking" tokens to reach a conclusion—which is where most of the cost and latency in current reasoning models come from—this architecture handles the logic in a latent space. It's essentially doing the heavy lifting internally rather than spelling everything out in the token stream. This is a huge deal for anyone building a real-world AI workflow where latency and API costs actually matter.

There's also some interesting industry tea here. Andrew Curran had been hinting at a memory-efficiency breakthrough coming from a non-major lab, and the dots are lining up. Lukasz Kaiser, who is literally a co-author of the original Transformer paper and an OpenAI researcher, is an investor and adviser for Pathway. It feels like the people who built the Transformer are now the ones leading the charge to evolve past it.

When you look at the cost-efficiency frontier, the gap between this and the giants is staggering. Most "reasoning" models require massive compute clusters to hit these kinds of benchmarks, but seeing a 150M model punch this far above its weight class suggests that architectural elegance beats raw scale. If we can get ARC-AGI performance out of models this small, the deployment phase for local, on-device reasoning is going to happen much faster than predicted.

This looks like a practical tutorial in how to optimize for intelligence per parameter. By utilizing recurrent memory, the model maintains a state that allows it to iterate on a problem without the linear cost increase of a growing context window. It's a lean, mean reasoning machine that makes the current "brute force" approach to LLMs look incredibly wasteful. I'm curious to see if this latent reasoning approach can be scaled up or if it's specifically tuned for the types of abstract puzzles found in ARC. Either way, the "bigger is always better" narrative just took a hit.