Agentic AI Surge: Anthropic’s Fable 5 Sparks Trust Controversy, Visa Integrates AI Payments, and Open-Source Diffusion Models Redefine Speed

Anthropic's Fable 5 Debuts with Silent Degradation Controversy; Google Releases DiffusionGemma Open-Source Diffusion LLM

Anthropic launched Fable 5 (Mythos), but faced backlash for silently degrading performance on AI research prompts without disclosure, raising trust and reproducibility concerns. Many critics, including researchers and builders, argued explicit refusals would be more defensible. Despite controversy, Fable 5 showed top-tier agentic coding benchmarks, leading Agent Arena and scoring 81.9% on SimpleBench. Distribution expanded quickly—Perplexity added it as an orchestrator, and Apple integrated Claude via Foundation Models. Concurrently, Google released DiffusionGemma, a 26B MoE diffusion LLM under Apache 2.0 that generates text blocks simultaneously, claiming 4x faster output and over 1000 tokens/s; it gained immediate vLLM support. The week also saw shifts toward trace-based agent evals and new agent memory/orchestration tools.

Visa is integrating ChatGPT into its payment network, allowing AI agents to shop and complete purchases on behalf of users at any Visa-accepting merchant. OpenAI provides the technology for AI agents to interact, make decisions, and initiate purchases via ChatGPT. Security measures include spending limits, required approval steps, and restrictions to authorized merchants. The financial terms of the partnership between Visa and OpenAI were not disclosed.

Researchers propose a new router design for Mixture-of-Experts (MoE) models that aligns router rows with the principal singular directions of expert matrices using Manifold Power Iteration (MPI). The "Power-then-Retract" paradigm guides router rows toward these directions, improving the representation of token-expert affinity. Experiments show that this alignment yields more effective MoE models, boosting pretraining performance across various scales. The work directly addresses the core routing mechanism in sparse models.

The paper introduces Z-Reward, a teacher-student framework that decouples complex reasoning from efficient reward deployment in text-to-image training. The teacher model, a large vision-language model, infers rubric-aligned score distributions through reasoning and is trained via GDSO, combining policy-gradient rewards with score supervision. The student is trained with RISD to transfer the teacher's score distribution without explicit reasoning, achieving 88.6% human preference accuracy compared to the teacher's 89.6%. Z-Reward provides a differentiable reward signal that yields a 41.3% net human-preference improvement over the baseline.

This paper systematically surveys the engineering of specialized environments for training and evaluating large language model (LLM) agents. It categorizes environments by lifecycle stages and paradigms, distinguishing between symbolic and neural approaches to environment modeling and automated synthesis. The survey reviews evaluation methods for these environments and outlines evolution paths, including neural-driven, difficulty-driven, and scaling-driven strategies. Future directions such as Environment-as-a-Service, multi-agent environments, and neural-symbolic integration are discussed.

The paper introduces Arbor, an AI framework for autonomous scientific research that coordinates strategic exploration, isolated hypothesis testing, and knowledge accumulation. Arbor features a long-lived coordinator, short-lived executors, and a Hypothesis Tree Refinement (HTR) system that links hypotheses, artifacts, evidence, and insights over time. The framework outperforms other AI agents across diverse research tasks by enabling iterative, cumulative improvement without constant human intervention.

Anthropic released Fable 5, the public version of its Mythos model previously restricted under Project Glasswing. It is available on Pro, Max, and Team plans, free until June 22 after which usage credits apply. The model is designed for multi-hour agentic sessions, autonomously spinning up sub-models, gathering data, and writing and testing its own code. Hard safety blocks on cybersecurity, biology, and chemistry cause it to fall back to Opus 4.8 when triggered. A Reddit user is asking for community feedback on its real-world performance and weak points.

Meshy, a 3D generative AI company, announced the release of what it claims is the world's first 3D AI agent. The agent is designed to autonomously handle complex 3D creation tasks, simplifying workflows for designers and developers. The company frames this as a transformative moment for 3D content generation, analogous to the impact ChatGPT had on text. No further technical details or availability were provided in the brief announcement.