Google's AI across the stack: Gemini family releases, DeepMind research, Google Cloud AI and Vertex, and the Search and Workspace integrations happening under AI Overviews.
Topic brief
Brief updated Aug 3, 2026
Google, through its parent Alphabet, is one of the largest AI infrastructure and model providers, spanning the Gemini model family, Google Cloud, DeepMind research, and consumer surfaces such as Search AI Overviews and Gemini in Chrome. Its scale gives it a distinct footprint: Gemini models run inside Google Cloud, Android, Chrome, Workspace and a growing set of third-party products from automakers to retailers and television formats, while DeepMind under Demis Hassabis continues to publish frontier research, extend Gemini into embodied robotics, and shape public debate over AI safety standards.
Google Cloud has become the company's clearest growth engine, and Google is simultaneously racing to close a self-acknowledged coding-capability gap, investing heavily in AI infrastructure, and shipping agent products such as CodeMender and AlphaEvolve to enterprise customers through the Gemini Enterprise Agent Platform. Its model lineup is unusually wide, running from Flash and Flash-Lite tiers aimed at cost-sensitive high-volume inference up to the Pro models that compete at the frontier, and it is one of the few vendors that designs its own accelerators for the workloads it serves.
For practitioners, Google is relevant both as a model and cloud vendor and as a proxy for how the largest AI companies handle safety governance proposals, copyright exposure over training data, and legal liability for AI-generated output in products used by billions of people. It is also the company whose search behavior sets the terms of the traffic bargain for publishers. Its coding-agent surface, its inference pricing, and its willingness to expose agents to third-party enterprise workloads are the details most likely to affect day-to-day engineering decisions.
The through-line across late July and early August 2026 is that Google is opening new Gemini surfaces faster than it can stabilize them, so an announced capability and a dependable one are no longer the same thing. On July 30 Google DeepMind launched Gemini Robotics 2, a vision-language-action model it says extends control from upper-body manipulation to whole-body humanoid movement, alongside Gemini Robotics ER 2, a high-level orchestrator available through the Gemini API and Google AI Studio with a private preview on the Gemini Enterprise Agent Platform; Google reported 57.4% accuracy on progress classification and 91.3% on moment finding, results the retrieved independent coverage did not reproduce. Within two days two other launches went the other way. Google withdrew the Google Earth image-generation feature on July 31, one day after its global web release, after researcher Henk van Ess documented that prompts placing refugees, a nuclear plant and a bomb crater into recognizable locations were not refused; Google said the images carried SynthID watermarks and never entered the shared Earth imagery, and gave no date for the feature's return. On the same day Google canceled the standalone AI Studio mobile app it had shown at I/O on May 19, after roughly 800,000 pre-orders, saying the app-building experience will move into Gemini with no rollout date or feature-parity plan.
The financial and governance frame around that build-out hardened at the same time. Alphabet's July 22 second-quarter results showed $119.8 billion of revenue, up 24%, with Google Cloud up 82% to $24.8 billion, but $44.9 billion of property and equipment spending against $39.1 billion of operating cash flow produced negative quarterly free cash flow of $5.9 billion under the company's own definition, and management raised 2026 capital-expenditure guidance to $195 billion to $205 billion from $180 billion to $190 billion. Reuters reported the Nasdaq fell 2.15% on July 23 with Alphabet down 7%. On the same call Sundar Pichai said coding and agentic coding are areas Google needs to improve, while Gemini 3.5 Pro remained in partner testing and Gemini 4 was in pretraining with no disclosed date. For buyers, that leaves the July 21 Flash tier as what is actually purchasable now, at $1.50 and $7.50 per million input and output tokens for Gemini 3.6 Flash and $0.30 and $2.50 for Gemini 3.5 Flash-Lite. Meanwhile the traffic bargain behind Search is being renegotiated in public: Google began testing a Search Console opt-out with a subset of UK site owners on June 3 under a UK Competition and Markets Authority requirement, and NewzDash reported on July 28 that Top Stories carousels appeared inside AI Overviews in 15.5% of its tracked US cases and 17.46% of UK cases where Top Stories were shown.
The nearest checkable signals are Google's own unfinished rollouts. Google says the Search Console AI opt-out will move beyond the subset of UK site owners after testing and the UK Competition and Markets Authority gave it nine months to implement all required changes, so watch whether the control reaches other markets and whether Google ever publishes AI-specific click data; Digiday reported it currently supplies AI-feature impressions but not clicks. Google gave no date for restoring the Google Earth image generator or for describing the additional guardrails it plans, and no rollout date or feature-parity plan for moving AI Studio's app-building workflow into Gemini. On models, Google said on July 21 that Gemini 3.5 Pro was testing with partners and would be broadly available "as soon as it's ready," and it has disclosed no date, pricing or evaluations for Gemini 4. In the courts, watch discovery in the Delaware Starbuck case, Google's formal response and any class-certification ruling in case 26-cv-5870, and whether Google appeals the non-final Munich injunction in case 26 O 869/26. On silicon and financing, the evidence does not establish Google commitments: Fubon's projection of 12 million to 15 million TPU v9 chips in 2028 is an analyst estimate Google has not confirmed, Frozen v2 is a reported experimental project with a 2028 target and no public specification, and the roughly $15 billion Nexus financing remains reported advanced talks, with Octus saying the bridge facility was expected to be replaced by bonds or loans in the fall. Hassabis told Axios he wants his proposed standards body operating before year-end; no government has adopted the proposal.
| model | availability | listed api price | google positioning and caveats |
|---|---|---|---|
| Gemini 3.6 Flash | Released July 21 through the Gemini API and Google AI Studio; also rolling out gradually across GitHub Copilot | $1.50 per million input tokens, $7.50 per million output tokens | Google positions it as its workhorse model for coding, knowledge work and multimodal tasks and says it uses 17% fewer output tokens than 3.5 Flash on the Artificial Analysis Index; these are vendor-reported benchmark results, not production guarantees |
| Gemini 3.5 Flash-Lite | Released July 21 through the Gemini API and Google AI Studio | $0.30 per million input tokens, $2.50 per million output tokens | Targeted at high-throughput, low-latency workloads such as agentic search and document processing; Google cites a measured speed of 350 output tokens per second |
| Gemini 3.5 Flash Cyber | Limited-access pilot through the CodeMender agent, offered only to governments and trusted partners | Not disclosed in the reviewed material | Fine-tuned to find, validate and patch software vulnerabilities, with Google saying multiple Flash Cyber agents can work together on a single report; because vulnerability research is dual use it is not a general release |
| Gemini 3.5 Pro | Testing with partners as of Google's July 21 statement, to be broadly available "as soon as it's ready" with no date given | Not disclosed in the reviewed material | Bloomberg reported on July 16 that the model was months behind schedule while Google worked on coding performance; Google has published no API availability, pricing, context limits or model card |
| Gemini 4 | Pretraining underway; no release date announced | Not disclosed in the reviewed material | Google says Gemini 4 has entered its most ambitious pretraining run and Pichai said Google will need it to compete at the level where the frontier is expected to be at release; no parameter count, context window or reproducible coding evaluations are available |
Google released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite on July 21 through the Gemini API and Google AI Studio, priced at $1.50 and $7.50 per million input and output tokens for 3.6 Flash and $0.30 and $2.50 for Flash-Lite. Gemini 3.5 Flash Cyber is not a general release; Google says it will be offered only to governments and trusted partners through a limited CodeMender pilot. Gemini 3.5 Pro was still testing with partners as of Google's July 21 statement, and Google says Gemini 4 has entered pretraining with no disclosed release date, pricing or API details. Google's efficiency figures, including 17% fewer output tokens than 3.5 Flash, are vendor-reported benchmark results rather than guarantees for a specific workload.
Alphabet reported $119.8 billion of revenue for the quarter ended June 30, 2026, up 24%, with operating income up 30% to $40.8 billion and Google Cloud up 82% to $24.8 billion on $8.8 billion of cloud operating income. The $112.1 billion net income figure is not a clean operating measure: Alphabet said a $99.0 billion net gain on equity securities added $77.1 billion to net income and $6.26 to diluted earnings per share. The company spent $44.9 billion on property and equipment against $39.1 billion of operating cash flow, producing negative quarterly free cash flow of $5.9 billion under its own definition, and raised 2026 capital-expenditure guidance to $195 billion to $205 billion. Reuters reported the Nasdaq fell 2.15% on July 23, with Alphabet down 7%, though Brent crude settling above $100 a barrel also contributed to the broader selloff.
Three separate tracks are live in the evidence. The Delaware Superior Court denied Google's motion to dismiss commentator Robert Starbuck's defamation suit over allegedly false AI-generated statements on July 24, which allows the case into discovery but is expressly not a finding that the allegations are true or that Google is liable. Hachette Book Group, Cengage Learning, Elsevier, Scott Turow and S.C.R.I.B.E. filed a proposed class action on July 10 in case 26-cv-5870 alleging Google used books obtained for limited purposes to develop Gemini. And Munich Regional Court I issued a temporary injunction on May 28 in case 26 O 869/26 that, according to The Decoder, treated an AI Overview as Google's own content rather than a neutral search result; Google said that decision is not final and is under review.
The evidence supports treating it as a controlled measurement problem rather than a universal rule. Google began testing a Search Console control on June 3 with a subset of UK site owners that excludes a site's links and content from AI Overviews, AI Mode and generative features in Discover without leaving traditional Search, and says the choice will not be used as a ranking signal outside those features but will cost traffic and impressions from them. It is distinct from Google-Extended, which governs training and grounding uses of crawled content. NewzDash reported on July 28 that Top Stories carousels appeared inside AI Overviews in 15.5% of tracked US cases and 17.46% of UK cases where Top Stories were shown, and observed an either-or placement, so opting out can also forfeit that position. Digiday reported that Google provides AI-feature impression information but not AI-specific click data, which is why the traffic cost remains hard to quantify.
Google DeepMind released two models with different jobs. Gemini Robotics 2 is a vision-language-action model that DeepMind says extends control from upper-body manipulation to whole-body humanoid movement, can run locally on-device, and can adapt to new robot bodies within a few hours. Gemini Robotics ER 2 is the high-level reasoning and orchestration layer, publicly available through the Gemini API and Google AI Studio with a private preview on the Gemini Enterprise Agent Platform; Google reported 57.4% accuracy assigning video frames to five progress bands and 91.3% accuracy on moment finding with a mean absolute timing distance of 0.96 seconds. Those are company-reported evaluations that the retrieved independent coverage did not reproduce, and DeepMind acknowledged that robot movement speed still needs improvement.
Several distinct issues surfaced in July. Pillar Security disclosed seven sandbox escapes and boundary bypasses on July 20 across Cursor, OpenAI Codex, Gemini CLI and Google Antigravity; Pillar's master report says Google classified two Antigravity findings as valid but downgraded their severity because exploitation would require social engineering or a repository containing indirect prompt injection, and does not list those two as patched. Varonis Threat Labs disclosed Rogue Agent in Dialogflow CX on July 7, and Axios reported Google said the issue is fully mitigated with no known customer compromise and no customer action required. A lock-screen flaw reported on July 17 let someone with physical access to an affected Android 16 phone use Gemini to send SMS without entering the PIN; Google told The Register it had implemented a fix and expected full deployment during the week of July 17, and the retrieved reports do not independently verify that rollout is complete. On the defensive side, Google says AI-assisted workflows helped its Chrome team fix 1,072 security bugs across Chrome 149 and Chrome 150.
Latest coverage
Showing newest first with full-topic pagination
Google canceled its unreleased standalone AI Studio mobile app on July 31 after about 800,000 pre-orders across Android and iOS. The company said it will bring the app-building experience into Gemini on mobile and desktop instead, while the browser-based AI Studio service will remain available; it did not give a rollout date for the Gemini features.
On August 2, 2026, 9to5Google reported that Google is testing a redesigned voice-search screen in the Android Google app beta, with Search, AI Mode, Search Live, and Song Search in one bottom toolbar. The test remembers whether a user last chose Search or AI Mode, but it is not widely available and Google has not announced a broader rollout.
Google began testing a Search Console control on June 3 that lets a subset of UK website owners leave AI Overviews, AI Mode, and AI features in Discover without leaving traditional Search. The tradeoff grew more concrete on July 28, when NewzDash reported that Top Stories appeared inside AI Overviews in 15.5% of its tracked US cases and 17.46% of UK cases where Top Stories were shown.
Google rolled back its new Google Earth image-generation feature on July 31, one day after launch, after screenshots of synthetic scenes appeared to violate its policies. The company said generated images were watermarked and never entered the shared Earth imagery, but it is withdrawing the tool while it develops stronger guardrails.
A lock-screen flaw reported in July let someone with physical access to an affected Android 16 phone use Gemini to send SMS or reconnect WhatsApp without entering the device PIN. Google told The Register that it had implemented a fix and scheduled full deployment for the week of July 17, but the retrieved reports do not independently verify that rollout is complete.
Android Authority reported on July 31 that code in Google Health version 5.04.1 reveals upcoming preferences that could let users disable AI Coach insights. The strings describe a "No Coach insights" option that would retain basic fitness metrics while removing extra commentary, although the code does not confirm public availability.
Fubon Research projects that Google could deploy 12 million to 15 million ninth-generation TPUs in 2028, according to reports by Tom's Hardware and TechSpot. The analyst estimate would put Google's annual custom-chip volume near Fubon's forecast of 12.4 million Nvidia data center GPUs, but Google has not confirmed the target or disclosed TPU v9 specifications.
Nexus Data Centers is seeking roughly $15 billion in financing to build a large Anthropic-linked campus in Hubbard, Texas. CNBC reported that Morgan Stanley is leading advanced bank talks and that Google would backstop Anthropic with its investment-grade credit, while Octus described the package as bridge financing expected to be replaced by bonds or loans.
Similarweb's 2026 Generative AI Landscape report found that 461 million of ChatGPT's 494 million users also used Google between March and May 2026, a 95% overlap. The same research says generative-AI platforms averaged 9.5 billion monthly web visits from June 2025 through May 2026, showing rapid growth without evidence that users have broadly abandoned traditional search.
Google says AI-assisted workflows helped its Chrome team fix 1,072 security bugs across Chrome 149 and Chrome 150, more than the total fixed across the preceding 23 release milestones. Google's Q2 security update describes deterministic and AI tooling for reproduction, severity analysis, routing, and validation, while independent reports document the sharp increase in patch volume.
Google began a gradual rollout of Gemini visual tools in Google Docs on July 28, enabling eligible users to create and edit images, diagrams, and infographics within a document. The web-only features can use document context and apply one natural-language request to multiple visuals. Access is limited to specified Workspace, education, and Google AI subscription tiers.
Google DeepMind launched Gemini Robotics 2 on July 30, adding whole-body humanoid control to its robotics AI stack. According to DeepMind, the vision-language-action model can translate vision and language into motor commands for motions from walking and crouching to two-handed object manipulation, while Gemini Robotics ER 2 handles high-level reasoning and multi-step task planning.
Google launched Gemini Robotics ER 2 on July 30, making the embodied-reasoning model available through the Gemini API and Google AI Studio. Google reports 57.4% accuracy on progress classification and 91.3% on moment finding, but independent benchmark results were not available in the retrieved coverage. The model coordinates lower-level robot controls, monitors video for task progress and supports collaboration between robots.
Google updated Gemini for macOS on July 29 with intelligent voice dictation triggered by a long press of the Fn key and an optional reasoning mode that can use onscreen context. The English-language feature is rolling out globally, and Google says it can remove filler words, handle mid-sentence corrections, and insert formatted text at the active cursor.
Google released its first ATLAS study on July 23, analyzing about 15 million de-identified interactions across the Gemini app, AI Mode and Gemini API. It found AI activity across 68% of occupations representing 90% of U.S. employment, but only about 21% of tasks in a typical job. Manual and technical workers who used the tools were twice as likely to use multimodal AI.
Google Research published a mathematical account on July 15 explaining how diffusion models can generate novel outputs rather than reproduce training examples. The post attributes this capability to score-function smoothing during denoising, which leads to interpolation along a hidden data manifold, while Creative Bloq situated the claim in an ongoing debate over whether AI image generators invent new images or copy training material.
U.S. stocks fell on July 23 after Alphabet and Tesla earnings revived concern about heavy AI spending, while surging oil prices added inflation pressure. Reuters reported the Nasdaq fell 2.15% and Alphabet sank 7%; Alphabet’s filed results showed $119.8 billion in quarterly revenue and $24.8 billion from Google Cloud, while management raised 2026 capital-spending guidance to $195 billion–$205 billion on its earnings call.
The Delaware Superior Court denied Google's motion to dismiss Robert Starbuck's defamation lawsuit over allegedly false statements generated by Google AI products. The July 24 ruling allows the case to proceed into discovery under Delaware's plaintiff-friendly pleading standard. It does not establish that Starbuck's allegations are true, that users relied on the outputs, or that Google is liable.
In June 2026, plaintiffs filed proposed class actions against Amazon Ring and Google Nest alleging that familiar-face systems collected bystanders' biometric face data without consent. Reuters identified Virginia resident Charles Sigwalt as a Ring plaintiff seeking at least $5 million for a proposed class. The complaints remain allegations that courts have not resolved.
Alphabet CEO Sundar Pichai said on the July 22 Q2 earnings call that coding and agentic coding are areas Google needs to improve. Google says Gemini 4 has entered its most ambitious pretraining run, and Pichai said the larger base model will be needed to compete at the next frontier. No release date, API specifications or independently verified coding results are available.
Alphabet raised its 2026 capital-expenditure outlook to $195 billion-$205 billion, while Tesla said full-year capex will exceed $25 billion as it expands AI compute, robotaxi, Optimus, and manufacturing capacity. In both companies' second-quarter disclosures, equipment spending exceeded operating cash flow. The comparison measures current investment intensity, not the companies' broader quarterly performance or the eventual returns from these programs.
Google Cloud put CodeMender into preview on July 21 as a managed code-security agent for Gemini Enterprise Agent Platform customers. Its documented workflow scans code, verifies findings with proof-of-concept exploits, and generates patches that are built, tested, and checked against the exploit before developers apply them; Google classifies the service as a Pre-GA offering.
Alphabet reported $119.8 billion in second-quarter revenue, up 24% from a year earlier, while Google Cloud revenue jumped 82% to $24.8 billion. Net income available to common shareholders reached $112.1 billion, but most of the increase came from a $99.0 billion gain on equity securities rather than the operating business.
GitHub began a gradual rollout of Google's Gemini 3.6 Flash to Copilot on July 21, 2026, covering paid individual and business plans across major editors, the CLI, the cloud agent and the Copilot app. Business and Enterprise administrators must enable a preview policy before members can select it, while GitHub bills usage at provider list pricing.
Samsung introduced its Gemini-powered Intelligent Eyewear at Galaxy Unpacked on July 22, showing camera-equipped Android XR glasses developed with Gentle Monster and Warby Parker. Samsung claims up to nine hours of battery life and plans a fall release; pricing, exact markets, weight, camera specifications and independent performance tests remain unavailable.
Bengaluru police allege that Kenneth Arun Kumar used Google Gemini over several months while planning the June 22 killings of three members of his partner's family. Investigators say they found searches about attacking people, cleaning evidence and disposing of bodies, and have sought records from the AI provider. The claims remain under investigation, and the retrieved reports do not include the underlying chatbot logs.
Pillar Security disclosed seven sandbox escapes and boundary bypasses affecting Cursor, OpenAI Codex, Gemini CLI and Google Antigravity on July 20. The common failure was indirect: agents wrote workspace files or reached local services that trusted components outside the sandbox later executed. Pillar and independent reporting say most issues were patched, while Google downgraded the two Antigravity reports as difficult to exploit.
Google released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite on July 21 and announced a limited-access Gemini 3.5 Flash Cyber pilot through CodeMender. Google says 3.6 Flash uses 17% fewer output tokens than 3.5 Flash and costs $1.50 per million input tokens and $7.50 per million output tokens, while Flash Cyber is restricted to governments and trusted partners.
Google is reportedly developing Frozen v2, an experimental server chip that would embed parts of Gemini's architecture in silicon. Internal projections cited by The Information put its efficiency at six to ten times more tokens per unit of power than Google's newest TPUs, but the project is not an announced product and may not reach deployment.
Automotive assistants are moving from fixed command grammars toward context-retaining, multi-turn interfaces, making account identity, grounding, and driver-distraction safeguards central deployment concerns. Honda announced July 15 that Google Gemini is replacing Google Assistant in thousands of eligible Honda vehicles with Google built-in. Honda's release lists select Civic, Accord, Prelude, CR-V, Passport, Pilot and Prologue vehicles; Car and Driver reports that nine Honda and three Acura models qualify, with eligibility limited by model year and trim. The update supports conversational requests, follow-up questions, trip planning, news summaries, and Gemini Live, according to Honda.
