Liquid AI's d1.2-ultra Takes First Place on AA-Paperclip Decision Test with 76 Points
Liquid AI has released d1.2-ultra and d1.2, models built for routine decisions: reading a customer email, extracting its fields, and picking the department it should go to. On the closed AA-Paperclip test from Artificial Analysis, d1.2-ultra scored 76 points, 3 points ahead of Drex Max v1.1, and ranked first in the composite Decision Index v2.1. d1.2 ranked sixth. For sorting support tickets, the cheaper d1.2 is worth running on your own examples.
AI-processed from X: @ArtificialAnlys; edited by Hamidun News
Most language model rankings show how a neural network writes code or solves olympiad problems. But production work usually calls for everyday tasks: reading an incoming email, pulling out an order number and an amount, and choosing one of several actions. These are the skills measured by the Decision Index from the independent platform Artificial Analysis. In its v2.1 version, the first place went to Liquid AI's d1.2-ultra, while its smaller sibling d1.2 came in sixth.
What the Tests Showed
Liquid AI updated its two flagship decision-making models: d1-ultra became d1.2-ultra, and d1.1 became d1.2. Artificial Analysis tested them on its own closed benchmark, AA-Paperclip. It evaluates two abilities: making decisions and extracting the needed data from text. Imagine a model receiving hundreds of support tickets, where it must pick the right department for each one and write down the key fields. d1.2-ultra scores 76 points here, the best result. The closest competitor, Drex Max v1.1, trails by 3 points.
On the composite Decision Index v2.1, d1.2-ultra scores 75 points, 4 more than d1-ultra. d1.2 scores 69 points, 3 more than d1.1. The gain is small, but both models improved at once.
An important caveat: AA-Paperclip is Artificial Analysis's own closed test. The tasks are not published, so the result cannot be independently reproduced. You have to trust the platform's methodology.
Pricing and Technical Parameters
Processing a million input tokens costs $0.042 with d1.2-ultra and $0.02 with d1.2. By Artificial Analysis's calculation, a million tasks from the Decision Index cost $133 for d1.2-ultra and $63 for d1.2. Both models sit on the Pareto frontier of capability versus cost. Put simply, no other model is both better and cheaper: getting more quality means paying more, and vice versa.
The other parameters are:
- a 64K-token context window: enough for a long email, a contract, or a batch of records, but not for an entire codebase;
- text and images are accepted as input, so you can feed the model screenshots of forms or scanned documents;
- access is currently only through Liquid AI's own API.
The publication gives no pricing for output tokens. For tasks with short answers, such as a class label or a JSON object with fields, this is not critical. If the model has to write long reasoning, the total bill will need to be calculated separately.
Who Will Benefit
If you run a pipeline where a model sorts tickets, extracts details from invoices, routes requests, or picks the next step for an agent, d1.2 and d1.2-ultra give you a chance to lower costs without losing quality. The smaller version at $0.02 per million input tokens suits bulk operations. The larger one makes more sense where an error is more expensive.
The limitations are clear too. A 64K-token context is modest by the standards of current models. The only access channel for now is the developer's own API, so you have to trust Liquid AI with your data, with no backup provider. Finally, the lead is confirmed by a single rating, and a closed one at that.
A reasonable plan: take a hundred real examples from your own workflow, run them through d1.2, d1.2-ultra, and your current model, and compare accuracy against cost. If the newcomer's advantage holds up, the savings at high volume could be substantial. If not, a spot in a table is not yet a reason to change your working model.
- Nov 21, 2026Eight Open-Source Tools for AI Agents: Browser Use Leads with 121,000 Stars
- Nov 21, 2026Parallel Search Tops 26-API AI Agent Search Ranking With 83 Points
- Nov 21, 2026A 360-line MCP server with no SDK: five decisions that stopped the agent from breaking links
- Nov 21, 2026Text-to-cad: AI Agent Draws Robot Parts From a Description, Project Passes 26,000 Stars
Key Questions Answered
What the Tests Showed?
Liquid AI updated its two flagship decision-making models: d1-ultra became d1.2-ultra, and d1.1 became d1.2. Artificial Analysis tested them on its own closed benchmark, AA-Paperclip. It evaluates two abilities: making decisions and extracting the needed data from text. Imagine a model receiving hundreds of support tickets, where it must pick the right department for each one and write down the key fields. d1.2-ultra scores 76 points here, the best result. The closest competitor, Drex Max v1.1, trails by 3 points.
What are the key details on Pricing and Technical Parameters?
Processing a million input tokens costs $0.042 with d1.2-ultra and $0.02 with d1.2. By Artificial Analysis's calculation, a million tasks from the Decision Index cost $133 for d1.2-ultra and $63 for d1.2. Both models sit on the Pareto frontier of capability versus cost. Put simply, no other model is both better and cheaper: getting more quality means paying more, and vice versa.
Who Will Benefit?
If you run a pipeline where a model sorts tickets, extracts details from invoices, routes requests, or picks the next step for an agent, d1.2 and d1.2-ultra give you a chance to lower costs without losing quality. The smaller version at $0.02 per million input tokens suits bulk operations. The larger one makes more sense where an error is more expensive. The limitations are clear too. A 64K-token context is modest by the standards of current models.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.