Google Cuts Gemini 3.5 Flash Pricing 25%, Pressures Enterprise Token Budgets

Token Cost Becomes a Vendor-Selection Criterion

Google brought Gemini 3.5 Flash to general availability at $1.50 per million input tokens and $9.00 per million output tokens, pricing it 25% below Gemini 3.1 Pro and 3.3 times cheaper on input than GPT-5.5. For buyers building high-volume workflow automation — document summarization, routing, agent task execution — this shifts the economics of model choice. When a single workflow runs thousands of inference calls per day, token cost per workflow becomes as important as raw capability. Google is betting that enterprises will trade slight performance differences for cost certainty in production.

The direct competitors are OpenAI's GPT-5.5 pricing, Microsoft's Azure model access, and Anthropic's Claude tiers. All three now face pressure to defend margin in the agent automation layer or cede share in high-throughput use cases. For buyers, this means vendor-agnostic architectures matter more than they did six months ago. Locking into a single model stack exposes you to pricing risk; building with swappable backends gives you leverage in annual renewals.

Microsoft Brings Computer-Use Automation to Every Tenant

Microsoft shipped Copilot Studio computer-use agents to general availability on May 13, 2026, making browser and UI automation a standard capability for every Power Platform tenant. This is not a pilot feature or premium SKU — it is now in the platform. The competitive target is obvious: UiPath, Automation Anywhere, and other RPA vendors that have spent the past 18 months adding LLM intelligence to their automation stacks.

What changes for buyers is the default position. Instead of asking whether to buy automation tooling, the question becomes whether Microsoft's native agent layer is good enough to replace specialized RPA vendors or whether you still need both. That decision depends on governance maturity. Microsoft owns identity, workflow control, and permission boundaries across the stack, which simplifies deployment but increases reliance on a single vendor for enforcement. If your compliance team already struggles with Microsoft's rate of change, adding agent-layer automation to that stack accelerates the governance backlog.

Anthropic Adds the Security Layer Enterprises Were Waiting For

Anthropic released new enterprise security primitives for Claude-based deployments, addressing the auditability and containment problems that have kept agents out of production in regulated environments. The specifics are not public, but the timing is deliberate: enterprises need proof that agent tool execution can be logged, reviewed, and stopped before they will fund large-scale rollouts.

The competitive dynamic here is clear. OpenAI's enterprise controls, Microsoft's Copilot governance stack, and the platform security features from Google Cloud and AWS are all racing to prove they can protect enterprises from runaway agents. Anthropic is positioning itself as the vendor that built containment into the model layer rather than bolting it on afterward. For buyers, this means security requirements now drive model selection in ways they did not a year ago. If your compliance function cannot audit an agent's decision path or limit its API access, that agent does not go into production, regardless of accuracy.

The Vendor Landscape Gets Crowded at the Workflow Layer

OpenAI launched ChatGPT Work for professional automation, joining Microsoft Copilot, Google's Gemini enterprise offerings, and Anthropic's Claude tools in the race to own task execution. Google's AI Mode reportedly crossed 1 billion monthly active users, and Google paired that scale with an agent-first product direction. The pattern is consistent: every major model vendor is building workflow automation into their stack rather than leaving that layer to third parties.

This creates a platform consolidation decision for buyers. Funding separate search, knowledge retrieval, and assistant layers is harder to justify when vendors bundle those capabilities into their core offering. The risk is lock-in. Consumer-scale usage drives faster feature rollout into business products, but it also means product direction is set by the consumer roadmap, not enterprise requirements. Buyers evaluating multi-year commitments need to model the switching cost if a vendor changes pricing, bundles features into higher tiers, or deprecates APIs that custom workflows depend on.

What to Watch

The shift from single-model assistants to agentic teams orchestrating workflows end-to-end is accelerating budget allocation from experimentation to platform engineering, security review, and workflow redesign. That reframes buying decisions. Instead of asking which copilot to pilot, buyers are now deciding whether to fund reusable agent infrastructure across operations. The vendors winning that decision are the ones who can prove cost predictability, governance depth, and safe tool execution at scale — not the ones with the flashiest demo.