What Google is in AI
Google occupies a structurally unusual position in the AI landscape: it is simultaneously a frontier model developer (Gemini), an open-weights distributor (Gemma), a compute and cloud infrastructure provider (TPUs, Google Cloud), a research lab (Google DeepMind), and a strategic investor and partner to rivals including Anthropic. No other organization in the current AI ecosystem plays all five roles at scale.
The Gemini family: closed frontier models
Google's flagship closed-model line is Gemini, with Gemini 3.1 Pro as the current top tier. The Gemini 3.5 Flash release at Google I/O 2026 extended the efficient Flash sub-line with multimodal capabilities, Gemini 3.5 Live Translate (covering 70+ languages in real time), and — critically — computer-use functionality, enabling the model to interact with desktop and web interfaces autonomously. This brings Gemini into direct competition with Anthropic's Claude computer use and OpenAI's equivalent capability for agentic workflows.
Google's Aletheia agent, built on Gemini 3 Deep Think, demonstrated the frontier of reasoning capability: applied to 200 unsolved Erdős mathematical problems, it produced 13 correct solutions, 4 of which were genuinely novel contributions not found in existing literature. Gemini 3 Deep Think benchmarks include 48.4% on HLE, 84.6% on ARC-AGI-2, and 93.8% on GPQA Diamond.
NotebookLM was upgraded to Gemini 3.5, and Veo 2 — Google's second-generation video model — rolled out to Gemini Advanced subscribers and Whisk Animate, enabling text-to-video and image-to-animation generation.
The Gemma family: open-weights strategy
Google's open-weights strategy runs in parallel through the Gemma line. Gemma 3 (March 2025) introduced multimodal, multilingual, and long-context capabilities to the open-weights tier. Gemma 4 (April 2026) positioned itself as frontier-level multimodal intelligence deployable on-device. Gemma 4 12B (June 2026) went further with an encoder-free unified architecture — eliminating the separate vision encoder common in most multimodal models — at the 12B parameter scale.
The Gemma ecosystem also includes ShieldGemma (a safety classifier), Gemma Scope (an interpretability toolset), and a 27B Gemma-based foundation model for single-cell biological analysis that contributed to the discovery of a new potential cancer therapy pathway.
DiffusionGemma, an experimental 26B MoE model, explores a fundamentally different generation paradigm: diffusion-based text generation producing 256-token blocks simultaneously, achieving over 1,000 tokens/second on H100 hardware — at the cost of lower output quality versus standard Gemma 4.
Google as infrastructure and partner
Google's role as infrastructure provider is as consequential as its model work. Anthropic signed a multi-gigawatt TPU compute agreement with Google and Broadcom for capacity coming online from 2027 — Anthropic's largest compute commitment to date at the time of signing. Google is also a major Anthropic investor, with deepening investment reported alongside Amazon in funding-for-compute deals.
Apple's AFM 3 on-device model family is distilled from Google Gemini, and Apple's new AI architecture is built around Gemini models — a significant consumer-AI partnership that embeds Google's models into iPhones and Macs at the OS level.
Research: biology, mathematics, and security
Google DeepMind's research portfolio spans well beyond language. AlphaGenome (April 2026) is an open-weights model that interprets the ~98% of human and mouse genomes that regulate gene expression rather than coding for proteins, taking up to 1 million DNA base pairs as input and outputting ~6,000 human and ~1,000 mouse gene properties. Across 50 evaluations, it matched or exceeded prior models in 47 cases.
CodeMender, announced in October 2025, is a DeepMind agent targeting automated identification and remediation of critical software security vulnerabilities — a domain where Google also participates as a founding member of Anthropic's Project Glasswing cybersecurity consortium.
Google's Paper Assistant Tool (PAT) automates scientific peer review using agentic AI, achieving a 34% improvement over zero-shot recall on mathematical errors in the SPOT benchmark, and was piloted at STOC and ICML.
Ecosystem and standards roles
Google is a co-founder of the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation established in December 2025 alongside Anthropic, OpenAI, Block, AWS, Microsoft, Cloudflare, and Bloomberg. The AAIF governs the Model Context Protocol (MCP), which has reached 10,000+ active public servers and 97M+ monthly SDK downloads, with Gemini as one of the integrated platforms.
On the safety and regulatory front, Google voluntarily agreed to submit models to NIST's TRAINS (Testing Risks of AI for National Security) task force for pre-deployment evaluation covering cybersecurity, biosecurity, and chemical weapons risks. Google is also co-developing an industry jailbreak severity framework with Anthropic, Amazon, Microsoft, and other Glasswing partners — a direct response to the export-control episode that temporarily suspended Anthropic's Claude Fable 5 and Mythos 5 globally in June 2026.
Competitive position and tensions
Independent benchmarks show Gemini 3.1 Pro Preview competing closely with GPT-5.4 Pro on aggregate intelligence indices, with a price and multimodal advantage noted. The SearchGEO adversarial evaluation found Gemini-3-Flash had a 31.4% attack success rate on web-content manipulation — the highest among models tested — a vulnerability pattern distinct from Claude's over-rejection tendency and GPT's over-trust. A Munich court ruled Google liable for false claims in AI Overviews, adding a legal dimension to the accuracy challenges facing search-integrated AI.
Gray-market API proxy networks in China have been found to substitute cheaper models for Gemini access, with proxy "Gemini-2.5" achieving only 37% on MedQA versus 83.82% via Google's official API — a signal of both demand for Google's models and the difficulty of enforcing access controls.
Where it's heading
The bundle points toward Google deepening on three vectors: (1) embedding Gemini into consumer hardware via the Apple partnership and its own on-device Gemma line; (2) scaling compute infrastructure for both internal training and external partners like Anthropic; and (3) extending agentic capability — computer use, background agents (Spark), and scientific research tools — across the Gemini platform. The regulatory environment, including TRAINS and the emerging pattern of government pre-deployment review, will increasingly shape how and when Google's frontier models reach the public.
