Catch up on Venture Beat articles in minutes

Enterprise AI orchestration is consolidating on model-provider platforms, with Anthropic's Claude leading. However, a gap exists between ambition and reality, as most deployed 'agents' are chatbot wrappers, not true multi-step workflows. Enterprises favor hybrid control planes to avoid vendor lock-in and prioritize workflow tooling and security, with real-time fiscal control remaining an exception.

Thinking Machines released Inkling, an open-source multimodal language model designed for cost-efficiency and censorship resistance. It performs well on benchmarks, though some Chinese models and closed-source alternatives lead in specific areas. Inkling's unique 'controllable thinking effort' allows developers to balance cost and performance.

Cohere VP Rachad Alao emphasized enterprise AI sovereignty requires control over the full agent stack, not just data or infrastructure. He argued that increasing agent complexity drives token utilization, making efficient model routing crucial. Cohere's approach focuses on solving problems privately and securely, using the right model for each task to reduce unnecessary usage and vendor lock-in.

Canva launched Code 2.0, an AI website builder for all users, including free accounts. It focuses on design integration and ease of use for non-technical users, differentiating from competitors by prioritizing visual output and editing. The tool aims to make interactive web elements accessible within the existing Canva ecosystem.

1Password launched AI Spend and Consumption Management to track AI token costs. The tool provides visibility into AI service usage from vendors like OpenAI and Anthropic, aiming to help IT and finance teams manage escalating expenses.

ACRouter, an open-source framework, dynamically routes AI prompts to optimize speed and cost. It uses a Context-Action-Feedback loop to learn from model successes and failures, outperforming static routers and expensive single-model strategies. The system is particularly effective for verifiable tasks like coding and data retrieval, offering significant cost savings.

Kubernetes-native workspace delivery solves the split infrastructure problem by integrating desktop management into existing cloud-native workflows. This approach offers consistent tooling, enhanced security through container isolation, and efficient scaling. Kasm Workspaces is presented as a solution that leverages Kubernetes for orchestration and delivery, with features like declarative configuration and horizontal scaling.

DeepSeek's 75% price cut on its V4-Pro model highlights the '100x problem' in AI agents: token consumption outpaces inference cost reduction. Agentic workflows, unlike simple chatbots, involve numerous model calls per user request, drastically increasing operational costs. This challenges traditional SaaS pricing models, as power users can incur higher inference costs than their subscription fees. Solutions involve cost-aware routing, prompt caching, and disciplined context management. Companies must prioritize inference cost as a key metric and budget accordingly to survive the evolving AI infrastructure landscape.

Slopsquatting, a new supply chain threat, exploits AI coding tool hallucinations to inject malware. Attackers register AI-generated fictitious package names, which developers then unknowingly incorporate into their code. This bypasses traditional typosquatting defenses and poses a significant, persistent risk.

57% of enterprises report AI agents providing incorrect information due to poor context. A governed context layer is the proposed solution, with 75% of enterprises lacking one. Vendors are developing diverse approaches to this layer, and enterprises that have experienced AI errors are more likely to invest in solutions.

AI agents often provide incorrect information due to missing or inconsistent business context, a problem affecting 57% of enterprises. A governed context layer is presented as the solution, though vendor approaches vary. Enterprises that have experienced AI errors are more likely to invest in context solutions.

OpenAI launched ChatGPT Work, an AI agent for task management across email, Slack, and calendars, powered by GPT-5.6. It aims to transform ChatGPT into an autonomous work platform, moving beyond text generation to execute multi-step tasks independently. The product emphasizes cloud-based persistence and mobile accessibility, competing with similar offerings from Anthropic and Microsoft.

A VentureBeat Research survey reveals that 86% of enterprises run GPUs at 50% capacity or less. Most deployed AI agents perform single-prompt tasks, not multi-step ones, and enterprises are retrofitting controls for identity, evaluation, and cost. Security incidents are common due to credential sharing and inadequate automated evaluations.

Enterprise AI teams face an evaluation gap as AI agents gain autonomy faster than companies can verify them. A survey reveals many deployments fail in production despite passing internal tests, highlighting a mismatch between automated evaluations and real-world outcomes. The article emphasizes the need for repeatability and robust regression testing over deployment speed to ensure dependable AI.

Google's TabFM is a new foundation model for tabular data that uses in-context learning, eliminating the need for per-dataset training. It overcomes LLM limitations by preserving table structure and offers faster deployment, though inference is more computationally intensive. While currently non-commercial, integration into Google BigQuery aims to broaden accessibility.

Shared API keys expose AI agents in 69% of enterprises, leading to significant security risks. Major security firms are investing billions in solutions for this emerging threat. Despite high confidence in tooling, many enterprises lag behind AI-enabled attackers, with larger organizations facing greater exposure.

AI code generation offers speed but faces enterprise integration, compliance, and maintenance hurdles. Operationalizing AI requires robust infrastructure, data readiness, and governance, not just code quality. Developers' roles shift to oversight and architectural judgment, with competitive advantage lying in encoded domain knowledge.