Google Cloud CEO Thomas Kurian used his keynote at Gemini at Work 2026 to say something no model maker had said quite that plainly. "Gemini is the agent, and the model underneath it is a separate choice," he wrote in the announcement post on October 8. The new Gemini agent sends each job to whichever model fits, and on day one that list includes Anthropic's Claude models alongside Google's own.
The launch puts a rival's software inside Google's flagship product, and it does it at the moment enterprise AI budgets are under real pressure. Google says per-token prices have fallen 98% since 2024, but that customers' token volume has grown faster than the savings. Running every simple loop through a premium model breaks budgets, Kurian argued, so Google built Smart Routing to triage workloads automatically and real-time spend caps that pause a project's agent when it hits a hard limit.
The agent is the product
The Gemini agent does one thing its predecessors did not: it plans a task, picks tools, spins up temporary sub-agents with their own identities, and comes back with finished work inside the apps people already use. It runs in Gmail, Drive, Docs, Sheets, Chat, and Calendar with the same memory it has on the web and the command line. It also works through Microsoft 365 and Slack.
Every action gets logged under the agent's own identity rather than a person's, and traffic passes through a firewall Google calls Agent Gateway. "Gemini gives you identity, discovery, governance, and security by default," the post says. That is the part that makes the launch newsworthy, since audit trails and permissions are what stalled a lot of agent pilots last year.
Google claims scale behind the product. Nearly 500 Cloud customers each processed more than a trillion tokens in the past year, per the company. Roughly 80% of all Google Cloud customers now use its AI products, and nearly 90% of the Fortune 100 use Gemini Enterprise. Those are Google's own figures and nobody outside the company has audited them, but the customer list is long enough to take seriously.
When a rival's model becomes a feature
The multi-model routing is where this story gets interesting. Google's post says the agent orchestrates "across our Gemini family of models and Claude models from Anthropic today, and other leading private and open models in the future." It also admits the obvious implication out loud: "The best model for the task is not always the largest one."
PayPal already routes 10 million multi-model requests a week, according to Google. Shopify blends frontier models for millions of merchants. Both are listed as evidence that customers want choice, not as customers switching wholesale to a single lab. Google's reasoning is that the leading model changes every few months, so an agent that can swap models keeps the context, skills, and data in place when the rankings move.
Anthropic spent the same week pushing hard on the low end of the market. On October 7 it released Claude Haiku 5.5 and said the model costs around 75% less to run than Haiku 4.5 on average, with pricing of $0.10 per million input tokens for prompts under 100,000 tokens. It scored 72.4% on the OSWorld 2.1 computer-use benchmark against 48.9% for OpenAI's GPT-6 Luna and 15.7% for Haiku 4.5.
Asana's staff software engineer Aaron Vinh said his team saw "over a 30% reduction in latency for task completions and up to 2.5x faster inference per agent turn." HubSpot's Ze'ev Klapow reported a 92.8% average score on the company's CRM eval suite, the best it had seen. AlphaSense runs about 8 million calls a week through its document Q&A feature and scored the new model at 0.84 against Haiku 4.5's 0.76.
What the numbers add up to
The two announcements fit together. Cheap small models are the reason an agent can afford to route a hundred classification calls to Haiku while sending one hard reasoning job to a frontier model. Google is not being generous to a competitor. It is buying the cheapest reliable route for the 90% of its customers' calls that don't need Argon.
The enterprise numbers tell the same story from the other side. Bradesco cut document review from one hour to five minutes and reported a 60% drop in risk inconsistencies. BNP Paribas is putting Gemini Enterprise in front of more than 65,000 employees. Japanese insurer SOMPO built over 10,000 custom agents across its 34,000 staff and cut new model development from a week to a day. None of those wins come from a single model being 20% better on a benchmark. They come from wiring the agent into the workflow and controlling what it costs.
The part nobody has tested yet
Google did not publish pricing for the Gemini agent or a general availability date. It did not say what share of jobs runs on Claude. Those are the two questions a procurement team will ask first, and the answers are missing.
There is also a structural question. If the model really is a setting, then the switching story cuts both ways, and Google knows it. An agent layer that treats Claude as one option among five is also an agent layer where Claude is one pricing update away from being the cheapest option for most of the traffic. Google is trading model lock-in for agent lock-in, and it's not clear that trade will hold when Anthropic or OpenAI start selling their own routing layers at enterprises directly.
What to watch is the next two quarters. Google promised pricing and availability details, and Anthropic is rolling out monthly API credits of $100 to $500 for Max and Team subscribers this week to pull developers onto the platform. Whichever vendor moves first on the enterprise contract, the direction is already set: the model is becoming a component, and the thing companies actually pay for is the thing that decides which model to call.
More coverage: AI desk and the archive. Google's full announcement is on the Google Cloud blog, Anthropic's Haiku 5.5 details are on its site, and TechCrunch has its take on the launch.