Google's AI year: how Gemini and Vertex caught up — and what that means for your stack

Eighteen months ago, recommending Google Cloud for a serious LLM workload was a hard sell. In April 2026, with Gemini 2 Pro and Flash shipping, Vertex AI Agent Builder hitting GA, and Code Assist quietly becoming the most-used Google developer product since Maps, the calculus has flipped. I have been migrating two clients off Bedrock-only stacks specifically because of latency and price on Gemini 2 Flash.

This is not a victory lap for Google. It is a status report on what is now actually viable.

Gemini 2: the model line that finally makes sense

The Gemini 1.x family was confusing. Pro, Ultra, Flash, 1.5, experimental variants — every quarter the SKU list shifted. Gemini 2, announced late 2025 and broadly rolled out by Q1 2026, collapsed it: Pro for reasoning, Flash for cheap latency-bound work, Nano for on-device. The 2M-token context on Pro is no longer a parlor trick — with the new caching pricing it is genuinely useful for whole-codebase analysis.

What actually moved the needle for my work:

• Native tool use that matches Anthropic's reliability. Earlier Gemini tool-calling was a coin flip; 2 Pro is at the point where I trust it for production agent loops.
• Multimodal that includes audio in and audio out at sane prices. Voice agents on Gemini 2 Flash are genuinely cheaper than Whisper-plus-GPT pipelines.
• Latency. p50 on Flash for short prompts is meaningfully under what GPT-4-class models deliver, and that compounds in chains.

How it stacks up

Numbers below are April 2026 list prices and self-reported context windows. Treat $/1M as directional — every provider has caching, batch, and committed-use discounts that move it.

| Model class | Context | $/1M in (approx) | $/1M out | Native tool use | Multimodal | p50 latency (short) | | ----------------- | ------- | ---------------- | -------- | --------------- | --------------- | ------------------- | | Gemini 2 Pro | 2M | $1.25 | $5.00 | Yes, reliable | Text/img/audio | ~700 ms | | Gemini 2 Flash | 1M | $0.10 | $0.40 | Yes | Text/img/audio | ~250 ms | | GPT-4 class | 200K | $2.50 | $10.00 | Yes, mature | Text/img | ~600 ms | | Claude 4 Sonnet | 200K | $3.00 | $15.00 | Yes, best-in-class | Text/img | ~750 ms |

Claude still wins for code-editing agents in my testing. Gemini wins on price-per-quality at the cheap tier and on long-context retrieval.

Vertex AI Agent Builder is the real story

The model is the headline; the platform is the bet. Vertex AI Agent Builder finally treats agent construction as a first-class product instead of a SageMaker-style pile of primitives. The pieces that matter:

• A managed retrieval layer that talks to BigQuery without me writing a vector-export pipeline. For analytics-heavy agents this is enormous.
• Reasoning engine deployment that is actually serverless. I do not pay for an idle GPU between requests, and cold start is sub-second for cached models.
• Tooling for evaluation built into the console, including LLM-as-judge with Gemini 2 Pro. Not as nice as Braintrust, but free and good enough.

The integration with Cloud Run is the part underrated by most reviews. Cloud Run now has a Gemini-aware sidecar mode where you can attach a model endpoint to a service and get streaming responses without managing the inference plane. For SaaS teams already on Cloud Run, this is the lowest-friction path to shipping LLM features I have seen on any cloud.

Code Assist, quietly

Code Assist is the product I was wrong about. I assumed Copilot's distribution lock-in was insurmountable. Inside three large clients, Code Assist usage is now higher than Copilot — driven entirely by the IDE-native code review and the fact that it can read across the entire BigQuery + Cloud Run + IAM context of a project. For a Google-shop team, that is a different product than Copilot, not a worse one.

Workspace and BigQuery

The Workspace integration moved from gimmick to default. Gemini in Sheets that runs SQL against BigQuery in the background is the kind of thing that quietly replaces a BI seat per analyst. I do not love the privacy review surface that opens up, but the productivity is undeniable.

The BigQuery side is the more interesting one for engineering teams. ML.GENERATE_TEXT calls against Gemini 2 from inside a SQL query, with the cost showing up on your BigQuery bill instead of a separate API contract, finally make in-warehouse LLM use practical at scale. I have one client running daily classification jobs over 40M rows entirely in BigQuery for less than a SageMaker batch job would cost.

Where it still loses

Three honest gaps:

• Region availability. Outside us-central1 and europe-west4, Gemini 2 Pro still has waitlists or higher latency. If your data has to live in Mumbai or Sao Paulo, check before you architect.
• Fine-tuning. Vertex tuning is improving but not at the level of Bedrock custom model import or Azure AI Foundry's tuning UX.
• The console. Vertex's console is still a Frankenstein of old AI Platform pages and new Agent Builder flows. Plan to live in the API.

Takeaways

• If you are price-sensitive and shipping Flash-class workloads, Gemini 2 Flash is the cheapest serious model in the market right now.
• For teams already on Google Cloud, Vertex Agent Builder + Cloud Run is the shortest path from prototype to production agents.
• BigQuery + Gemini for in-warehouse ML is no longer an experiment; it is the default for classification and extraction at scale.
• Do not migrate off Anthropic for code-editing agents. Do consider it for retrieval-heavy and voice workloads.
• Watch DeepMind's release cadence — the gap between research drops and Vertex availability has shrunk to weeks.

Google has not won. But for the first time, ignoring it on a serious AI architecture review is malpractice.