Today's AI news: Stripe buys OpenRouter for $7B, Liquid AI launches LFM2.5-VL-DSpark, and new robotics research pushes the field forward.
Stripe Acquires OpenRouter for $7 Billion
The biggest industry news today is a landmark acquisition: Stripe has purchased OpenRouter — the well-known AI model routing platform — for $7 billion. The deal was discussed in depth on the Latent Space podcast, featuring OpenRouter co-founder Alex Atallah and AMP partner Anjney Midha. The acquisition underscores just how dramatically the AI landscape has shifted since 2023, when many observers doubted there could ever be more than one or two frontier model labs. Today there are dozens of competitive model providers, and OpenRouter built its reputation by giving developers a unified interface to access them all. The $7 billion price tag signals that AI infrastructure — not just the models themselves — is now among the most valuable territory in tech.
Liquid AI Launches LFM2.5-VL-DSpark Vision-Language Model
Liquid AI has released a new model called LFM2.5-VL-DSpark, detailed in a post on Hugging Face. The model is designed to accelerate vision-language workloads, combining visual and language understanding in a single architecture. While full technical benchmarks were not provided in the summary, the release represents a continued push by Liquid AI to expand its model family beyond text into multimodal territory — a space that has seen intense competition in 2026.
NVIDIA Warp and MjWarp: Accelerating Robotics Simulation
NVIDIA has published a practical guide on Hugging Face explaining how to use its Warp framework alongside MjWarp to speed up robotics simulation and reinforcement learning workflows. The post targets developers and researchers looking to reduce the computational cost of training robot policies in simulation — a critical bottleneck as the robotics AI field scales up. NVIDIA's continued investment in simulation tooling reflects the broader industry view that high-fidelity, fast simulation is essential for training capable real-world robots.
AD-WM: Making Robot Planning Smarter with Action-Discriminative World Models
Researchers have introduced AD-WM — an Action-Discriminative World Model designed specifically for counterfactual model predictive control (MPC). The key insight behind the work is that standard world models are trained to predict what will happen next (factual transitions), but robot planning actually requires comparing what would happen under different possible actions from the same state. A model can have low prediction error and still be poor at distinguishing between candidate actions.
AD-WM addresses this by combining residual latent dynamics with an action-recovery regularisation objective, encouraging the model to preserve action-relevant information during planning. The results are striking: on the OGBench-Cube benchmark, hard-start task success jumped from 3.7% to 52.0% compared to a matched baseline. In real-world testing on a Franka robot arm, basic pick-and-place success improved from 42.2% to 71.1% with zero lab-specific adaptation, using a frozen V-JEPA 2 encoder. The authors argue that world models for planning should be evaluated on whether they support good counterfactual decisions — not just whether they predict the future accurately.
RAPID: Teaching Robots to Program Themselves from a Single Demonstration
A separate piece of robotics research introduces RAPID — Robot Agentic Programming from Demonstrations. The system uses a coding agent loop to automatically generate, verify, and refine robot programs from a single visual human demonstration, with no manual specification of task goals or action primitives required.
RAPID represents tasks using an object-centric relational program structure, meaning it captures the logic of what a person demonstrated rather than simply replaying their exact motions. This allows the resulting programs to generalise across variations in object pose, shape, material, and environment. Tested on eight contact-rich manipulation tasks in simulation and deployed on a real Franka arm, RAPID delivered strong performance across the board. The work points toward a future where robots can be taught new skills through observation alone, with AI agents handling the programming required to make those skills reliable and reusable.
Frequently Asked Questions
Why did Stripe acquire OpenRouter, and how much did it pay?
Stripe acquired OpenRouter — a platform that gives developers unified access to dozens of AI models — for $7 billion. The deal reflects the growing value of AI infrastructure as the number of frontier model providers has expanded dramatically since 2023.
What is Liquid AI's LFM2.5-VL-DSpark model?
LFM2.5-VL-DSpark is a new vision-language model released by Liquid AI, published on Hugging Face. It is designed to accelerate workloads that combine visual and language understanding in a single model.
What does the AD-WM robotics research achieve?
AD-WM is an action-discriminative world model for robot planning. It improved hard-start task success from 3.7% to 52.0% on a simulation benchmark, and boosted pick-and-place success on a real Franka robot arm from 42.2% to 71.1% without any lab-specific adaptation.
What is RAPID and how does it help with robot programming?
RAPID — Robot Agentic Programming from Demonstrations — is a system that automatically generates, verifies, and refines robot programs from a single visual human demonstration. It uses an object-centric relational program structure so the resulting programs generalise across variations in object pose, shape, and material, and has been tested successfully on a real Franka robot arm.