Model Development - The Batch

Illustration of agent processes: observe, reason, and act, with memory stored for visual inspections.

Model Development

Context Engineering for Images: MIT researchers build VISTA, an agent harness that keeps images in memory for faster retrieval

Managing an agent’s context by storing all the input text it receives and giving it tools to search the text boosts its performance.

Two robotic hands play checkers on a table, observed by a standing person in a workshop.

Model Development

World Models Become Robot Playgrounds: Google Genie 4 combines with Gemini models to build better video game worlds, robotics training

Google says that Genie 4, its latest world model, will be used in tandem with Gemini models to create digital practice spaces for robots, giving them endless worlds to train in

Image illustrates improvement and investment loops crucial for Claude's company growth and development.

Business

Anthropic Opens Its Books: Anthropic's S-1 outlines company's strong financials, methods, potential risks after IPO

Claude’s maker will sell company shares to the public for the first time.

Bar chart ranks AI models by score; Gemini 4 Carbon leads with 54.8%, highlighting its top performance.

Model Development

Gemini Leads the Pack Again: Google Gemini 4 Carbon tops Intelligence Index, cybersecurity benchmarks

Google’s Gemini 4 Carbon puts the company on top of Artificial Analysis’ Intelligence Index, edging past Anthropic’s Claude Opus 5.5, while costing less per task.

Machine Learning Research

Nvidia Courts A New Partner: Reflection releases Beam, an open weight model from an independent U.S.-based lab

Reflection AI, a closely-watched, Nvidia-backed U.S. startup, announced its first model.

Machine Learning Research

Meta Releases Wonder, Opens Spark: All about Muse Wonder (API), Muse Wonder Full (subscription), and Muse Spark Open (weights available)

Meta marched all the way back to the AI frontier to offer a Muse model for everyone.

Machine Learning Research

Mistral’s Sovereign Appeal: Mistral Large 4, the leading European AI model, outperforms all open models outside of China

Mistral’s latest model is its largest and most capable, and the most powerful yet released by a European company.

A chart of Terminal-Bench 4.0 results for Claude Haiku 5.5, Haiku 4.5, and Sonnet 5.5 on two different cache read settings, all at different effort settings

Machine Learning Research

Claude Fills Out Its Model Family: Claude Sonnet 5.5 and Haiku 5.5 give developers fast, highly capable alternatives to Opus 5.5

Claude Haiku, Anthropic’s fastest and least expensive model gets a long-awaited update.

Colored blocks on a grid under a robotic arm, labeled left and right, reflect robotics advancements.

Machine Learning Research

Video Models Steer Robots: Black Forest Labs’ Flux 3 Action applies video expertise to a World-Action Model

Many companies that build image and video generators have pivoted to robotics, extending their expertise in world models to action.

Benchmark chart shows Gemini 4 Argon scoring highest on Vals Index and Vibe Code Bench categories.

Machine Learning Research

Argon, A Low-Hallucination Model Build for Knowledge Work: Google's first Gemini 4 Model, Argon, avoids hallucinations and shines at knowledge work

Google took over half a year before it replaced its flagship AI model, Gemini 3.1 Pro Preview.

Pink-themed chat window mentions deployment issues, paired with a sleek iOS beta release page.

Machine Learning Research

Sol Nearly Eclipses Astra: OpenAI’s GPT-6.1 Sol is DevDay’s star, joined by Dots and new ChatGPT integrations

OpenAI’s latest mid-tier model lands within a point of its flagship at less than a quarter of the cost per task on Artificial Analysis’ Intelligence Index.

Graph illustrates decreasing memory use per token in DeepSeek models, highlighting V4.1-Flash efficiency.

Machine Learning Research

DeepSeek’s Flash Leapfrogs Pro Again: How V4.1’s architecture reduces cache reads

When agents make a tool call, they send the same context back through the model every time.