The AI log, frontier · open weights · science · image, video, world
11 Nov 2026
Gemini Decision 1, Praxis-1, AlphaGenome 2, Raon-1 nano
Google's decision endpoint, Runway's open world action model, DeepMind's cell-level DNA model, and a small Apache 2.0 model from KRAFTON.
By Ali Asif, published
What shipped
- Google
Gemini Decisions API
A Gemini API endpoint powered by gemini-decision-1 that answers up to 1,000 yes/no, multiple-choice or rubric questions per request with confidence scores, at $0.10 per million input tokens and no billed output.
- Runway
Praxis-1
An open-weights 27B mixture-of-experts world action model under Apache 2.0 that scores 46.2% on NVIDIA's RoboLab-120 against 42.9% for FLUX 3 Action.
- Google DeepMind
AlphaGenome 2
A conditional DNA model that takes a cell type plus a 1-million-letter sequence, trained on over 100 million cell profiles across 68 tissues, available for non-commercial research through the API.
- KRAFTON AI
Raon-1 nano
An Apache 2.0 Korean and English model with 1.5B effective parameters, tool calling and a 262,144 token context, leading its comparison set on IFBench at 75.9.
Four releases today, and the most useful ones share a theme: models built for one narrow job rather than a general chat window. I read each primary page and did not run any of them.
Gemini Decisions API
Google put a decision endpoint into the Gemini API, powered by a model called gemini-decision-1. You send one shared input, which can be text, images, audio, video or PDFs up to 256K tokens, and attach up to 1,000 independent questions. Each question is yes/no, a multiple choice with 2 to 255 options, or an ordered rubric score. Every answer comes back with a confidence score between 0 and 1.
The pricing is the interesting part. Input is $0.10 per million tokens, output is $0, and the shared input is billed once per request no matter how many questions ride on it. Google quotes 150 to 250 ms end to end, with latency staying nearly flat from 1 to 100 questions on contexts up to 16K tokens. It says the model leads its benchmark suite on agent routing, reasoning, ranking and multimodal understanding, with the lowest calibration error, and that it is nearly 5x cheaper than Gemini 3.5 Flash-Lite with the Interactions API. That suite is Google's own. It is in the Gemini API and AI Studio, with a no-cost tier for prototyping.
Praxis-1
Runway released Praxis-1, a world action model that predicts video and actions together. It is a 27B mixture-of-experts diffusion transformer, and the weights, inference code and fine-tuning code are all Apache 2.0, commercial use included.
On NVIDIA's RoboLab-120 leaderboard, across 1,200 episodes, Praxis-1 scores 46.2%, ahead of FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8% and pi0.5 at 28.0%. Fine-tuned on Google's RT-1 robot, which was not in its training data, it reaches 80.0% on SimplerEnv against 52.4% for the original RT-1 policy. Runway also says web video works about as well as teleoperated robot video for pretraining. It runs on H100, B200, RTX 5090, RTX PRO 6000 and Jetson Thor, and launch partners include Noble Machines, Standard Bots and Ultra.
AlphaGenome 2
DeepMind's new DNA model is conditional: you give it a tissue, cell type or state plus a 1-million-letter sequence, and it predicts gene expression, histone modifications, chromatin accessibility and DNA methylation for that condition. It was trained on more than 100 million individual cell profiles across 68 tissues.
On TenK10K, which covers 28 primary immune cell types, it found causal variants at 30% versus 16% and directional gene effects at 43% versus 23%, compared with the original AlphaGenome fine-tuned on single-cell data and other specialized tools. It is available for non-commercial research through the AlphaGenome API, and DeepMind states it is not validated for clinical or diagnostic use.
Raon-1 nano
KRAFTON AI released Raon-1 nano, a dense model with 1.5B effective parameters (3.1B total including a per-layer embedding table), under Apache 2.0. It handles Korean and English, supports tool calling and takes up to 262,144 tokens of context. The model card lists IFBench at 75.9, with 73.7 for the next model, and says it does not lead everywhere: BFCL v4 is 56.3 against 61.0 for MiniCPM5 2B. Two of its benchmarks are internal and unreleased.
What I'd watch next
I want to see someone other than Runway reproduce the RoboLab-120 ranking, and independent calibration numbers for Gemini Decision 1 on questions that are not Google's. For AlphaGenome 2, the next signal is whether outside labs report results on their own single-cell data.
Sources
Topics: Google DeepMind, Google, Gemini, DeepMind, World Models, Robotics, Runway, Decision Models