150M BDH-CQ hits 29.5% on ARC-AGI-1 for $0.0007 a task

A 150-million-parameter reasoning model called BDH-CQ has claimed a new cost-efficiency point on the public ARC-AGI-1 benchmark, reaching 29.5% pass@2 for a computed $0.00070 per task, less than a tenth of a cent. The paper posted to Hugging Face reports that inference takes roughly 0.85 H200 GPU-seconds per task at an assumed $3 per H200-hour, which is what pulls the price so low.

What sets BDH-CQ apart is that it combines in-context learning with recurrent latent reasoning. Rather than emitting a long chain-of-thought in tokens, the model updates a recurrent memory as demonstrations are ingested, then iterates in a high-dimensional latent workspace before decoding an answer. Intermediate reasoning states are never verbalized. That design is what enables the sub-cent cost, since the system is not paying for a long stream of generated tokens on each puzzle. It sits in a broader run of latent-reasoning work that our generative-AI tracker has been logging steadily this quarter.

Against the current leaderboard, the authors report BDH-CQ is roughly 57x cheaper than GPT 5.6 Luna (Low), which scores 34.2% at $0.040 per task in ARC Prize data as of July 2026. After OpenAI's 80% public API price cut for GPT 5.6 Luna on July 30, 2026, that gap shrinks to about 11x. Note the direction of the comparison: BDH-CQ is cheaper, not more accurate; it sits below Luna on the score axis but far to the left on the cost axis, which is why the paper frames it as a new point on the Pareto frontier rather than a new headline score. An independent black-box audit by co-authors from Bielik and New York University reproduced the 29.5% pass@2 result without access to the model weights. For context on where the accuracy frontier is moving, Anthropic reported yesterday that an unreleased Claude pushes the Zeta bound to 67.2%.

The caveats are worth spelling out. The training mixture blends "privately curated examples" with the public ARC-AGI-1 training set, RE-ARC, ConceptARC, ARC-Heavy and ARC-GEN100K, and the exact update rules for the recurrent memory and latent workspace are held as proprietary. So the deployed numbers can be audited but the recipe cannot be reproduced from the paper alone. The controlled experiments also show sharp limits: strict-task accuracy on ConceptARC is 59.38%, ordering falls off past sequence length six, and forcing the model to select between two demonstrated rules using a marker cue drops accuracy by 43.3 points.

The interesting bet is scale. The authors report that early runs confirm Transformer-like pretraining scaling laws from 1B to 600B parameters while preserving BDH-CQ's latent reasoning behavior. If that holds outside a 150M ARC configuration, cheap latent inference becomes the regime other reasoning teams have to answer to.