OpenAI Announces 'Presence' for Running Enterprise Agents in Production

OpenAI has launched Presence, a new enterprise offering built to help companies run AI agents in production. Whereas the existing Workspace Agents targeted internal workflows, Presence is designed with external deployments such as customer support in mind. For complex use cases, OpenAI engineers will step in directly to help with the build — a move that lays the groundwork for lifting agents from 'experiment' to 'real operation.'

A Wave of Newcomers Hits the Open-Model Front

A roundup of open models introduced a batch of new releases all at once. On top of Kimi K3 and DeepSeek-V4-Flash, which we've already covered, the ranks of organizations open-sourcing strong models keep growing.

Thinking Machines' Inkling**: The company's first in-house model, a 975B-A41B multimodal MoE that takes text, image, and audio input. It can't claim to be the strongest in its size class, but it looks promising as the base for Tinker, their commercial fine-tuning service. A smaller variant, Inkling-Small (276B-A12B), was released alongside it.

Tencent's Hy3**: A 295B-A21B MoE. The biggest story here is the license change — a switch from their previous custom restrictions to Apache 2.0. The company claims it solved a 50-year-old unsolved math problem using a dedicated proof harness.

Meituan's LongCat-2.0**: A massive 1.6T-parameter MoE. What stands out is less the performance than the training environment — it's the first 'non-Huawei, non-trial' model trained entirely on Huawei's Ascend 910. As a real-world demonstration of how far training can go on Chinese-made AI chips, it's a milestone.

Poolside's Laguna-S-2.1**: An 118B-A8B MoE from a company best known for coding. It drew attention for fitting on a single DGX Spark, and it adopts OpenMDW, an Apache 2.0-style license for AI models.

Also in the lineup: Motif-3-Beta (a 314B-A13B preview) from Korea's Motif, Instella-MoE-16B trained by AMD on its own Instinct hardware, and Swiss-AI's Apertus-v1.5-70B, among others.

There was a time when some predicted that soaring training costs would push the industry toward consolidation, but reality has moved the other way. Players keep multiplying, and the battle at the Pareto frontier of open models is only getting fiercer.

Only 8.9% of Sites Block AI Crawlers — 94.8% Never Get Cited at All

An analysis of AI search engines has been making the rounds in tech communities. Two numbers stand out:

Only **8.9%** of sites block AI crawlers.

Meanwhile, **94.8%** of sites are **never cited even once** in AI-generated answers.

In other words, blocked or not, the overwhelming majority of sites simply never get picked up by AI. For site owners, 'discoverability' in the AI era looks set to become a challenge quite distinct from search SEO — and one that is still largely uncharted.

Arrakis Raises $8 Million for Agent Runtime Security

As agents move into production, attention is turning to runtime control. Startup Arrakis has raised $8 million with agent runtime control at its core. As agentic adoption spreads, demand for guardrails that keep agents from behaving in unintended ways keeps widening.

Other Moves

Australia** is planning to broaden the scope of its News Bargaining Initiative — which pushes Big Tech to pay media outlets — and to strengthen how it charges them. A bill is expected to be introduced within the year, with Meta, Google, and TikTok among the likely targets.

An analysis from the St. Louis Fed catalogued AI- and productivity-related statements made on corporate earnings calls. A clearer picture is emerging of how companies describe the returns on their AI investments.