Small Model Upset: Pathway's BDH-CQ Architecture Nears GPT-5 on ARC-AGI at 1/11th the Cost
ahuja_priyank · x · 2026-08-13
Pathway's 150M-parameter model, BDH-CQ, has achieved a breakthrough on the ARC-AGI-1 benchmark. It reached a 29.5% pass@2 accuracy at a cost of $0.0007 per task, making it roughly 11 times cheaper than GPT-5.6 Luna (Low), which scores 34.2%.
Core Architectural Innovation:
- Instead of generating long chain-of-thought traces, BDH-CQ reasons in latent space using recurrent memory.
- It allows adjustable reasoning effort: LOW (21%), MEDIUM (27%), and HIGH (29.5%).
This progress suggests that the next major AI breakthrough might come from architectural innovation rather than just scaling up models, drastically improving 'intelligence per dollar'.
Related event: Pathway's 150M BDH-CQ Model Sets New Cost-Efficiency Frontier on ARC-AGI(7 posts)→
More from Models
- Users Report GPT-4o Acting 'Dumb' Recently, Suspecting Massive Nerf — imjustnewatai · 2026-08-13
- Qwen3.8-2.4T-A95B Available on Modal with 1M Token Context — AAAzzam · 2026-08-13
- Hands-on with Grok 4.6: Sonnet-class Performance at a Fraction of the Cost — bindureddy · 2026-08-13
- Users Complain About Qwen 3.8 Open Weights: No Vision, Crippled vs API — entsnack · 2026-08-13
- From Intelligence to Cost-Efficiency: The Evolving Core Metric for LLMs — beffjezos · 2026-08-13
- DeepSeek and Grok Updates Trigger AI's Jevons Paradox with Plunging Costs — R_D · 2026-08-13