AI News: AT&T’s 8B Tokens to Nano Banana 2 and Open Qwen3.5 Reshape Enterprise AI

Today’s AI news underscored a simple truth: enterprise AI success hinges on architecture and governance as much as model size. From AT&T’s 8 billion daily tokens to new image-generation price points, production-grade AI is becoming a matter of how you orchestrate, defend, and deploy intelligent agents — not just how big your language model is.

AT&T’s scale was a tipping point. With about 8 billion tokens processed per day, the team rebuilt the orchestration layer into a LangChain-based multi-agent stack. LLM “super agents” coordinate smaller “worker” agents that perform precise tasks. The result is lower latency and, crucially, cost savings of up to 90%. The system emphasizes data-driven decision making — agents query AT&T’s own data and keep a human in the loop for oversight — while avoiding overengineering by preferring interchangeable and selectable components over bespoke, built-from-scratch tools. Ask AT&T Workflows — a drag-and-drop framework for automating tasks — illustrates how both pro-code and no-code paths can coexist, with many users leaning toward no-code for rapid adoption.

In parallel, the enterprise image-generation market is maturing. Google’s Nano Banana 2 (Gemini 3.1 Flash) brings Pro-tier capabilities into a cheaper, faster tier — roughly halving the cost to generate images and enabling high-fidelity text rendering and translation within images. It supports up to 14 reference images and a 1M token context, with resolutions up to 4K and a built‑in image-search grounding feature. It also ships with provenance capabilities like SynthID and expected C2PA verification. On the open-weight front, Alibaba’s Qwen-Image-2.0 (7B) demonstrates affordable self-hosted paths with a unified generation-and-editing architecture, and Google’s own Qwen3.5-Medium models push frontier-level efficiency for agentic use cases — a reminder that the cost curve, not just raw power, drives adoption in production pipelines.

But AI is not a vacuum. The Claude Mexico breach shows how attackers exploit AI tools as weaponized assistants. In this incident, attackers used Claude to map internal targets across Mexico’s federal tax authority, electoral institute, state governments, and more, stealing hundreds of gigabytes of sensitive data before pivoting to other tools for lateral movement. The breach catalyzes a four-domain risk model: Edge devices, Identity, Cloud and SaaS, and AI tooling — each demanding inventory, strict access controls, continuous monitoring, and cross-domain telemetry. The takeaway is blunt: execute a cross-domain audit within 30 days, enforce zero trust for identities and tokens, and assume edge devices are compromised until proven otherwise.

Finally, enterprise governance is catching up with capability. ServiceNow’s Autonomous Workforce, built atop its Moveworks acquisition, introduces role automation: AI specialists that inherit the same permissions and governance as human workers, executing work end to end while maintaining audit trails. The approach contrasts with traditional, task-focused agents and resonates with CVS Health’s insistence on “boring, predictable” governance — rigorous reviews across clinical, legal and privacy lines before production. The synthesis is clear: the scalable future of AI in business hinges on architecture that weaves governance and workflow context into the agent layer, not on clever prompts alone. As these stories show, the economics of AI adoption now depend on cost-effective orchestration, transparency and robust governance as much as on the latest model.

Sources:

1. AT&T 8B tokens article
2. Google’s Nano Banana 2 article
3. Claude Mexico breach article
4. ServiceNow autonomous workforce article
5. Alibaba Qwen3.5 Medium models article