Google DeepMind Unveils Gemini 3 With Sub-Second Long-Context Reasoning Across 10m Tokens

Google DeepMind has formally launched Gemini 3, the latest generation of the company's flagship foundation-model family, with the headline performance claim of sub-second long-context reasoning across a 10-million-token context window — an architectural step-change that materially shifts the practical use-case envelope for a significant share of enterprise AI workloads.

The 10-million-token context window — equivalent to roughly 7,500 pages of dense business document — substantially extends what was previously possible across the proprietary frontier-model landscape. The more important architectural advance is the latency profile across that context window: what previously required tens of seconds for full-context reasoning has now been compressed to under a second on the production Gemini 3 stack, with the latency improvement coming from a combination of architectural changes in the attention mechanism, custom inference-stack optimisations on Google's TPU v6 hardware, and a substantially-improved retrieval-and-attention coordination layer.

The model family release covers four variants: Gemini 3 Pro (the flagship reasoning model), Gemini 3 Flash (a latency-optimised variant for agentic workloads), Gemini 3 Code (specialised for software-engineering tasks), and a multimodal Gemini 3 Vision variant that substantially extends the previous-generation visual-and-video reasoning capability. The benchmark performance places Gemini 3 Pro at the top of essentially every major industry-standard reasoning evaluation, with particularly notable leads on the practical-software-engineering and the scientific-reasoning evaluation suites.

The deployment infrastructure has been simultaneously upgraded. Google Cloud's Vertex AI platform now offers Gemini 3 Pro inference at $1.20 per million input tokens and $4.80 per million output tokens — pricing that meaningfully undercuts the equivalent OpenAI GPT-5 offering and that signals Google's continuing willingness to use its hyperscaler-scale TPU infrastructure to compete aggressively on the per-token economics of the inference market.

For the wider AI-applications landscape, the Gemini 3 release substantially shifts the practical envelope for enterprise deployment. Use cases that were previously gated by context-window or latency limitations — comprehensive financial-document analysis, full-codebase software-engineering tasks, longitudinal patient-record reasoning in healthcare, multi-document legal-discovery automation — are now operationally feasible at production-grade latency for the first time. The competitive response from OpenAI, Anthropic, and the broader frontier-model landscape will be the principal forward variable across the rest of the calendar year.