Topic deskAI and data science news

Google AI, Gemini and DeepMind News

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.

Stories

1,128

Latest source update

August 21, 2026

Coverage

Live

Topic brief

What to know about Google AI

Brief updated Aug 6, 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 eyewear and television formats, while Google DeepMind 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, now including capacity leased to rival labs.

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.

What changed recently

The most consequential recent change is organizational. Google said on August 5 that Demis Hassabis will give up day-to-day leadership of Google DeepMind to become its chair and Alphabet's chief scientist, with Koray Kavukcuoglu taking over as a senior vice president reporting to Sundar Pichai and overseeing Gemini models, frontier research, the Gemini app and developer teams. In the same reshuffle, Jeff Dean and Sanjay Ghemawat are leaving to form Discovery Loop, a public-benefit research company with Google as a founding investor and cloud partner. That lands while the frontier tier is already behind: Bloomberg reported on July 16 that Gemini 3.5 Pro was months late while Google worked on coding performance, Google said on July 21 it remained in partner testing with no broad-availability date, and Pichai said on the July 22 earnings call that coding and agentic coding are areas Google needs to improve, with Gemini 4 in what Google calls its most ambitious pretraining run. What Google did ship in the interim was breadth rather than a new frontier model: Gemini 3.6 Flash on July 21 at $1.50 per million input tokens and $7.50 per million output tokens, which Google says uses 17% fewer output tokens than 3.5 Flash, alongside Gemini 3.5 Flash-Lite and a restricted Gemini 3.5 Flash Cyber pilot delivered through CodeMender; a two-part robotics stack on July 30 in which Gemini Robotics 2 extends vision-language-action control to whole-body humanoid movement and Gemini Robotics ER 2 handles high-level planning, video-based progress tracking and multi-robot coordination through the Gemini API and AI Studio; and CodeMender itself, in preview since July 21 as a managed code-security agent that reruns a verified proof-of-concept exploit against its own candidate patch before a developer applies it. If you are planning a migration, keep pricing and benchmark assumptions provisional until release documentation is public.

The bill and the liability surface both grew. Alphabet reported $119.8 billion in second-quarter revenue, up 24%, with Google Cloud up 82% to $24.8 billion and cloud operating income of $8.8 billion, but $44.9 billion of capital spending pushed quarterly free cash flow to negative $5.9 billion, most of the net-income jump came from a $99.0 billion gain on equity securities rather than the operating business, and management raised 2026 capital-spending guidance to $195 billion to $205 billion; the Nasdaq fell 2.15% on July 23 and Alphabet dropped 7%. Some of that capacity is being monetized through competitors: Bloomberg reported on August 4 that Blackstone held early investor discussions on a second debt package of at least $36 billion to finance Anthropic's use of Google TPUs, after an earlier roughly $35 billion package covering five data centers, and CNBC reported Google would backstop Anthropic with its investment-grade credit in the roughly $15 billion Nexus campus financing in Hubbard, Texas. Meanwhile generated output kept producing exposure: the Delaware Superior Court denied Google's motion to dismiss Robert Starbuck's defamation suit over allegedly false statements from Google AI products on July 24, allowing discovery; Hachette Book Group, Cengage Learning, Elsevier, author 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 train Gemini; Google withdrew its Google Earth image generator on July 31, one day after launch, after screenshots of synthetic scenes appeared to violate its policies; and Pillar Security disclosed on August 3 two now-patched privilege-boundary weaknesses in Google's ADK for Python repository, one of which let prompt injection cross from a public triage agent into a maintainer-only workflow. For teams building on generated answers and agent handoffs, the practical read is that provenance, complaint handling and trust boundaries between agents now need product owners, not just review checklists.

What to watch

The clearest model question is whether Gemini 3.5 Pro leaves partner testing with public release and API documentation, and whether Gemini 4 emerges from pretraining with coding results that can be independently verified against Pichai's acknowledgment that coding and agentic coding need improvement; the leadership change under Koray Kavukcuoglu and the departure of Jeff Dean and Sanjay Ghemawat to Discovery Loop make the next release cadence a real signal rather than a formality. A hard date to mark is September 4, when Google begins removing Google Assistant from Android phones, tablets and paired devices in a staged rollout that affected users cannot reverse; teams supporting voice workflows across mixed device fleets and non-Gemini regions should test before then. On the publisher side, watch whether Google extends the Search Console AI opt-out beyond the subset of UK website owners that began testing on June 3, and whether it ever provides AI-specific click data, which is what publishers would need to compare opting in against opting out now that NewzDash finds Top Stories appearing inside AI Overviews in 15.5% of tracked US and 17.46% of UK cases. Legally, the open items are discovery in Robert Starbuck's Delaware defamation case and Google's response in the publishers' proposed class action, case 26-cv-5870. On products, watch whether the Google Earth image generator returns with the stronger guardrails Google says it is developing, whether CodeMender exits Pre-GA with published accuracy and regression metrics, and whether the Gemini 3.5 Flash Cyber pilot widens beyond governments and trusted partners. On infrastructure, the questions are whether Alphabet holds to its $195 billion to $205 billion 2026 capital-spending guidance after a quarter of negative $5.9 billion free cash flow, whether the second Anthropic TPU debt package moves past early discussions, and whether Google confirms anything about the reported Frozen v2 inference chip or Fubon's projection of 12 million to 15 million TPU v9 chips in 2028.

Timeline

2026-08-05Blackstone Weighs Second $36B Anthropic TPU Debt Package
2026-08-05Google Reshapes DeepMind Leadership as Hassabis Moves to Alphabet Science Role
2026-08-05Google Retires Assistant From Android Devices September 4
2026-08-03Pillar Discloses Patched Agent-to-Agent Flaws in Google ADK Repository
2026-07-30Google Launches Gemini Robotics ER 2 for Multi-Robot Tasks
2026-07-30Google DeepMind Launches Gemini Robotics 2

Key players

Google DeepMindThe research unit behind Gemini, now led operationally by Koray Kavukcuoglu after the August 5 reshuffle. It launched Gemini Robotics 2 and Gemini Robotics ER 2 on July 30, released Gemini 3.6 Flash and 3.5 Flash-Lite on July 21, and says Gemini 4 has entered its most ambitious pretraining run.
Demis HassabisMoving from day-to-day leadership of Google DeepMind to chair of the unit and chief scientist of Alphabet. In a July 14 signed essay he proposed a federally overseen U.S. frontier AI standards body loosely modeled on FINRA, with labs voluntarily submitting qualifying models up to 30 days before launch and mandatory approval only after the evaluation system proved reliable; he told Axios he wants it operating before year-end.
Koray KavukcuogluTakes over Google DeepMind as a senior vice president reporting to Sundar Pichai, with responsibility spanning Gemini models, frontier research, the Gemini app and developer teams. His remit covers exactly the areas Google has publicly said need to improve, including coding and agentic coding.
Sundar PichaiAlphabet's CEO, who said on the July 22 second-quarter earnings call that coding and agentic coding are areas Google needs to improve and that a larger base model will be needed to compete at the next frontier. On the same call Alphabet raised 2026 capital-spending guidance to $195 billion to $205 billion.
Jeff Dean and Sanjay GhemawatBoth are leaving Google to form Discovery Loop, a public-benefit research company for which Google is a founding investor and cloud partner. Their departure was announced alongside the August 5 DeepMind leadership change.
AnthropicGoogle's largest publicly visible external TPU customer. Blackstone arranged a roughly $35 billion package for Anthropic to lease Google chips across five data centers and, per Bloomberg's August 4 report, discussed a second package of at least $36 billion; CNBC reported Google would backstop Anthropic with its investment-grade credit in the roughly $15 billion Nexus financing in Hubbard, Texas.
Pillar SecurityDisclosed two now-patched privilege-boundary weaknesses in Google's ADK for Python repository on August 3, showing prompt injection crossing from a public triage agent into a maintainer-only workflow and, separately, an Antigravity-based path that could expose a bot token and Google Cloud credentials. The same firm disclosed seven sandbox escapes across four coding agents on July 20, including Gemini CLI and Antigravity, two of which Google downgraded as difficult to exploit.
Book publishers and authorsHachette Book Group, Cengage Learning, Elsevier, author Scott Turow and S.C.R.I.B.E. filed a proposed class action against Google on July 10 in case 26-cv-5870, alleging it copied books supplied for limited Google Books uses, scraped other works, removed copyright-management information and used the material to develop Gemini. The plaintiffs seek damages, an injunction and destruction of allegedly unauthorized copies; the allegations remain unproven.

Frequently asked questions

Is Gemini 3.5 Pro available yet, and what should I plan around?

Not as of the evidence here. Bloomberg reported on July 16 that Gemini 3.5 Pro was months behind schedule while Google worked on coding performance, and Google said on July 21 that the model was testing with partners and would be broadly available as soon as it is ready, without a date. Sundar Pichai separately said on the July 22 earnings call that coding and agentic coding are areas Google needs to improve. Keep migration, pricing and benchmark plans provisional until release and API documentation are public.

What do the new Flash models cost, and when do they make sense?

Gemini 3.6 Flash launched July 21 at $1.50 per million input tokens and $7.50 per million output tokens, with Google reporting 17% fewer output tokens than 3.5 Flash. Gemini 3.5 Flash-Lite shipped alongside it, and a restricted Gemini 3.5 Flash Cyber pilot is limited to governments and trusted partners through CodeMender. GitHub began rolling 3.6 Flash across Copilot the same day, though Business and Enterprise administrators must enable a preview policy first. Because the token-efficiency claim is vendor-reported, benchmark task success, total tokens, tool calls, retries and review overhead on your own workload rather than comparing list prices.

Does the DeepMind leadership change affect Google's model roadmap?

It changes who owns it. Demis Hassabis becomes chair of Google DeepMind and Alphabet's chief scientist, giving up day-to-day operations, and Koray Kavukcuoglu takes the unit as a senior vice president reporting to Sundar Pichai with responsibility for Gemini models, frontier research, the Gemini app and developer teams. Jeff Dean and Sanjay Ghemawat are leaving to found Discovery Loop with Google as founding investor and cloud partner. Google announced no accompanying change to model availability or pricing, so the checkable consequence is release cadence rather than anything you need to act on today.

How serious were the ADK agent-to-agent flaws, and what do they imply for my own multi-agent setup?

Pillar Security disclosed two now-patched weaknesses in Google's ADK for Python repository on August 3. In the first, prompt injection in a public triage agent could cross a natural-language handoff into a maintainer-only workflow and fabricate convincing review activity, though a maintainer still had to merge the pull request. A later Antigravity-based path could expose a bot token and Google Cloud credentials and enable runner code execution without maintainer interaction; Google fixed that in July. The generalizable lesson is that a handoff between agents is a privilege boundary, and content from an untrusted agent should not be able to select or parameterize a higher-privilege workflow.

How does Google's robotics stack divide responsibilities between the two models?

Gemini Robotics 2, launched July 30, is the vision-language-action model that translates vision and language into motor commands, extended from upper-body actions to whole-body humanoid movement including walking, crouching and two-handed manipulation. Gemini Robotics ER 2, released the same day through the Gemini API and Google AI Studio, is the embodied-reasoning layer that plans multi-step tasks, monitors video for progress and coordinates between robots. Google reports 57.4% progress-classification accuracy and 91.3% moment-finding accuracy for ER 2, figures the retrieved independent coverage did not reproduce. Physical deployments still require robot-specific safety validation, and latency and cross-body transfer remain open evaluation questions.

What is the Assistant-to-Gemini migration timeline on Android, and does it affect my users?

Google will begin removing Google Assistant from Android phones, tablets and paired devices on September 4, according to an email reported by The Verge and Android Authority. The rollout may take several weeks and affected users will not be able to switch back. Gemini becomes the remaining assistant option only on eligible devices in Gemini-supported regions, so mixed device fleets and unsupported regions are where breakage is most likely. If you ship voice workflows on Android, test across model years, device types and regional configurations before that date.

Latest coverage

Latest Google AI news

Showing newest first with full-topic pagination

See the full news feed