Google Cloud makes Gemini 1.5 Pro with 2-million-token context generally available on Vertex AI
Enzo Reno ·
Google on Tuesday made Gemini 1.5 Pro with a two million token context window generally available on Vertex AI, expanding regional coverage and enterprise controls for customers building retrieval and multimodal applications. The company said the model is available in the United States and European Union regions with default data isolation, and keeps previously published pricing of 7 dollars per million input tokens and 21 dollars per million output tokens on pay as you go, as listed on Google Cloud’s pricing page in 2024.
The move matters because context length has become a primary differentiator for enterprise buyers that need to process large documents, codebases, and long video. Anthropic’s Claude 3.5 Sonnet exposes a 200,000 token context window, and OpenAI’s GPT‑4o supported 128,000 tokens in 2024, so the two million token tier positions Google to capture RFPs in legal, finance, and media where chunking adds latency and cost.
Why it matters now
Google is bundling the model with Vertex AI’s grounding via text embeddings and BigQuery federation, which reduces integration effort for regulated workloads. The company is also enabling batch inference and parallel function calling for 1.5 Pro, features that CIOs have requested to manage cost at scale.
Google said internal evaluations show higher recall for multi document question answering at million token scale compared with Gemini 1.0 and 1.5 releases earlier in 2024. It cited performance on MMLU at or above 80 percent depending on prompt format, which is in line with public benchmarks from Google DeepMind for 1.5 Pro. The company added that Native Inference on Vertex AI cuts median latency for 512,000 token prompts compared with serverless endpoints that launched in early 2024, though it did not publish absolute numbers.
Pricing and availability
Google reiterated that pay as you go pricing remains 7 dollars per million input tokens and 21 dollars per million output tokens for Gemini 1.5 Pro, with volume discounts through committed use. For EU customers subject to the AI Act, Google pointed to updated system cards and SynthID watermarking for images and video, which DeepMind detailed in 2023 and extended across Gemini in 2024.
Context and competition
Enterprises have pushed for vendor diversification as Nvidia GPU supply remains tight while GB200 systems ramp in 2025 and 2026, according to Nvidia’s guidance at GTC 2024. Microsoft offers OpenAI models and small Phi models on Azure AI, and AWS Bedrock hosts Anthropic and Mistral, which makes Vertex AI’s breadth and long context a procurement lever.
For the broader market, general availability of two million token context with published per token pricing moves long context from research to standard infrastructure. That raises the bar for document heavy AI in sectors like insurance and public sector, and it pressures rivals to improve grounding, latency, and regional controls as large context becomes a default enterprise requirement.