Google Cuts Enterprise AI Inference Costs 70% as Workflow Agents Hit General Availability

Google Undercuts OpenAI on High-Volume Inference

Google launched Gemini 3.5 Flash at $1.50 per million input tokens and $9.00 per million output tokens on May 18, pricing the model 25% below Gemini 3.1 Pro and 3.3× cheaper on input than OpenAI's GPT-5.5. For enterprises running high-throughput workloads—document processing, internal search, customer service assistants—the cost delta changes the budget conversation from "can we afford this?" to "which workflow do we automate first?"

The same week, Google reported AI Mode crossed 1 billion monthly active users, expanding the distribution surface for its enterprise AI stack and increasing pressure on OpenAI and Anthropic to match token economics or defend on differentiation.

Microsoft Ships Computer-Use Agents to Every Power Platform Tenant

Microsoft moved computer-use agents in Copilot Studio to general availability on May 13, making the capability standard in every Power Platform license. The agents can navigate browser-based workflows and legacy UIs without custom RPA tooling, which directly challenges UiPath, ServiceNow, and AWS automation products.

For buyers already paying for Microsoft 365 or Power Platform, this lowers integration friction and may eliminate the need for separate automation software. The competitive impact is immediate: enterprises evaluating automation stacks now have a native Microsoft option that requires no additional procurement cycle.

Anthropic Targets Regulated Buyers with Security Primitives

Anthropic shipped what the company described as "the security primitives enterprises had been waiting for" in mid-May, focusing on audit controls, policy enforcement, and governance rather than raw benchmark performance. The move positions Anthropic against OpenAI and Google in finance, healthcare, and other regulated sectors where compliance risk and auditability often matter more than output quality.

For procurement teams, the shift means shorter security review cycles and lower compliance risk if the primitives reduce manual control requirements. Anthropic is betting that differentiation on trust will matter more than differentiation on speed for a meaningful segment of enterprise buyers.

Enterprise Demand Moves from Assistants to Multi-Agent Workflows

The pattern across Google, Microsoft, and Anthropic is the same: enterprises are no longer buying standalone chatbots. They are deploying agents that orchestrate multi-step workflows, call external tools, and execute tasks without human supervision. Google's enterprise documentation frames the shift as moving "from assistants to agentic teams," where specialized agents handle distinct parts of a business process.

This broadens competition beyond model providers to workflow platforms and systems integrators that can operationalize multi-agent orchestration. Technology buyers will increasingly ask vendors whether they support tool use, agent handoffs, and workflow governance, not just summarization or Q&A.

OpenAI Adds $4 Billion Services Layer

OpenAI created a new OpenAI Deployment Company backed by more than $4 billion on May 15 to accelerate enterprise adoption through embedded engineering teams and consulting services. The move puts OpenAI in direct competition with Microsoft, Accenture, and other integrators that sell implementation services, not just model access.

For large buyers, the services layer can reduce implementation risk, but it also signals that serious deployments now come with heavier integration and professional services spend, not just API fees. Buyers evaluating OpenAI should budget for both.

Two-Thirds of Enterprises Hedge Model Strategy

A July enterprise AI brief found that two-thirds of organizations are already hedging their AI model strategy, blending closed models and open-weight alternatives to reduce concentration risk. The finding favors vendors that support multi-model workflows over single-model lock-in strategies and weakens pure platform exclusivity plays.

Procurement teams are more likely to demand portability, fallback models, and multi-vendor governance because dependency risk is now a mainstream management concern. Buyers should ask vendors how their platforms handle model switching and whether agent workflows can run across providers.

What to Watch

Agent security is becoming a runtime issue, not just a policy issue. A recent agentjacking disclosure showed that a crafted error event could hijack coding agents in controlled testing, demonstrating that agent containment, authorization, and observability layers matter before production deployment. Security teams will increasingly require tool-permission controls and audit logs before approving autonomous workflows.

The pricing war between Google, OpenAI, and Anthropic will continue to compress inference costs, but the real competitive battle is shifting to workflow orchestration, multi-agent governance, and enterprise trust. Buyers evaluating platforms should focus on integration friction, security primitives, and multi-model portability rather than token cost alone.