TLDR AI 2025-03-26

AI companies face a unique problem when it comes to pricing: On one hand, users expect a frictionless experience that allows experimentation without barriers. On the other, every product interaction comes at a very real price. Training models can cost millions, and expenses for API calls quickly add up.

That tension can make profitability a delicate balancing act – but it doesn't have to. As the billing engine powering OpenAI, Metronome wrote the pricing and billing playbook that AI startups need to get their pricing right. The playbook dives deep into the pricing strategies that enabled OpenAI's soaring revenue, including:

• Token-based pricing metrics to bridge resource consumption + user engagement
• Redefining billing as a product experience
• The nuances of OpenAI's PayGo + prepaid credits model

Read the OpenAI playbook

OpenAI's GPT-4o introduces improved image generation with precise text rendering, instruction following, and multi-turn editing.

DeepSeek has released its new V3-0324 model, which outperforms GPT 4.5 in most benchmarks and features major improvements in performance.

Qwen has released a strong vision language model that is open and can run reasonably well on consumer hardware.

Midjourney has released some work to improve diversity in creative writing models. It was able to post-train a small 7B model that outperforms much larger open and closed models in creative writing.

FFaceNeRF enhances 3D face editing by overcoming fixed mask limitations in NeRF-based methods.

Test time compute for video that uses a guidance model to reject frame paths that don't conform to physics or the user specified prompt. Using test time compute for this task dramatically improves performance on benchmarks.

Dereflection Any Image (DAI) introduces a new diffusion-based approach for reflection removal, leveraging a high-quality dataset and progressive training.

FastCuRL-1.5B-Preview advances slow-thinking reasoning models with curriculum reinforcement learning, achieving state-of-the-art results in fewer training steps.

VGGT is a feed-forward neural network that directly infers all key 3D attributes of a scene, including extrinsic and intrinsic camera parameters, point maps, depth maps, and 3D point tracks, from one, a few, or hundreds of its views, within seconds.

SISO is an inference-time optimization method that personalizes images from a single subject image without training. It can personalize the subject of a given image or generate new images with the personal subject.

Pete Florence, previously a senior research scientist at DeepMind, has left to form Generalist AI, a robotics startup backed by Nvidia's VC arm, NVentures. The company aims to create general-purpose robots, with Florence highlighting the vision of reducing the marginal cost of physical labor to zero. Florence joins other former DeepMind scientists in launching tech ventures amidst Nvidia's growing influence in AI.

Outreach founder Manny Medina has launched Paid, a platform designed to help AI agent startups manage pricing and profitability. With backing from top investors, Paid aims to redefine how AI agents are billed and valued in the emerging agent economy.

Claude now features web search functionality for real-time insights and up-to-date responses, complete with source citations.

OpenAI has introduced new audio models, including Voice Engine, capable of generating realistic voices from short samples.

AI search startup Perplexity is in talks to raise $1 billion at a valuation of $18 billion.