ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
TopicDisconfirmAI EditorialUpdated 2026-11-21 00:46

Liquid AI d1.2 Series

4 steps1 pieces4 Sources

US-based AI firm Liquid AI, an MIT spinout headquartered in Cambridge, Massachusetts, released its new omnimodal decision models, d1.2 and d1.2-ultra, on November 16, 2026. The launch marks a significant update to the company's "System One" architecture, which answers structured questions with values and probabilities rather than generating text tokens. According to third-party benchmarking by Artificial Analysis (AA), d1.2-ultra achieved a score of 75 on the Decision Index v2.1, ranking first among decision models and surpassing the previous leader, Drex Max v1.1, by three points. Both models now support audio input alongside text and images, expanding their utility for tasks like identifying languages or detecting background sounds.

The core value proposition lies in calibration and cost-efficiency. Liquid AI reports that d1.2-ultra has a calibration error of 0.8 across 13 benchmarks, significantly lower than competitors like Jev 1.14 (2.7) and GPT-6 Luna (7.2). Lower calibration error means the model's stated confidence levels more accurately reflect its actual correctness, allowing developers to automate high-confidence decisions while routing uncertain cases to human review. Pricing remains unchanged from the previous generation: $0.02 per million input tokens for d1.2 and $0.042 for d1.2-ultra, with no charges for output tokens. This positions both models on the Pareto frontier of capability versus cost, according to AA data.

While the performance metrics are compelling, key details remain unsettled. Liquid AI has announced plans to open-weight the models but has not provided a specific date. Furthermore, the calibration figures cited are based on internal tests conducted by Liquid AI; independent verification of these specific probability gaps against other models is pending. Readers should watch for the official release of the open weights, the integration of these models into third-party platforms like OpenRouter and Vercel, and any subsequent third-party audits of the calibration claims.

How it got here

  1. 2026-10-15

    October 15, 2026: Liquid AI releases d1-ultra, claiming it outperforms all other decision models on the Artificial Analysis Decision Index by at least 8 points.

Sign in / Sign up
ReadingResearchRadarInvestment framework
liquid.ai ↗
  • 2026-11-16

    November 16, 2026: Liquid AI launches d1.2 and d1.2-ultra, adding audio input and five new task types (value reading, span extraction, object localization, number estimation, multi-choice selection).

    liquid.ai ↗
  • 2026-11-16

    November 16, 2026: Artificial Analysis confirms d1.2-ultra ranks #1 on the Decision Index v2.1 with a score of 75, leading Drex Max v1.1 by 3 points.

    cellcog.ai ↗
  • 2026-11-18

    November 18, 2026: Media coverage highlights the low API pricing ($0.02/$0.042 per million input tokens) and the company's statement that open weights will be released 'at some point' without a fixed date.

    aiweekly.co ↗
  • What is still unsettled

    • When will Liquid AI actually release the open weights for d1.2 and d1.2-ultra?
    • Will independent researchers replicate the reported calibration error advantages over Jev and GPT-6 Luna?
    • How will the addition of audio input affect latency and cost structures in real-world enterprise deployments?

    Sources

    • liquid.aiLiquid AI Blog↗
    • linkedin.comArtificial Analysis / LinkedIn↗
    • cellcog.aiCellCog Blog↗
    • aiweekly.coAI Weekly↗

    Coverage in this topic

    • US AI Firm Liquid AI Releases d1.2 Series Decision Models, Tops Third-Party Benchmark and Plans to Open Weights2026-11-18News

      Liquid AI launched full-modality decision models d1.2 and d1.2-ultra, with the latter ranking #1 on the AA index and showing lower calibration error than competitor Jev.