The Bard launch (March 2023) and the Bard–to–Gemini rename (February 2024)
Google's reactive answer to ChatGPT was announced on February 6, 2023 and shipped to early-access US and UK users on March 21, 2023 as Bard, an experimental conversational AI service backed by LaMDA. The launch demo's well-publicized error — the model misattributed the first exoplanet image to the James Webb Space Telescope — helped frame the early-2023 narrative that Google was caught flat-footed by OpenAI's November 2022 ChatGPT release. EU access was blocked initially over GDPR concerns and came online in July 2023.
At Google I/O on May 10, 2023, Bard's underlying model swapped from LaMDA to PaLM 2. Specialty variants — Med-PaLM 2 for medical question answering, Sec-PaLM for security — followed. Duet AI for Workspace shipped at $30/user/month on August 29, 2023 as the productivity-suite analog. The PaLM API was deprecated entirely once Gemini 1.0 shipped in December 2023.
The Bard–to–Gemini rename happened on February 8, 2024, the same day Gemini 1.0 Ultra reached general availability via the new Gemini Advanced subscription on the $19.99/month Google One AI Premium plan. bard.google.com began redirecting to gemini.google.com; Duet AI for Workspace was rebranded Gemini for Workspace; a new Android Gemini app shipped that day. The post is at blog.google. Gemini Advanced was itself rebranded Google AI Pro at I/O 2025 on May 20, 2025 alongside the higher-priced Google AI Ultra tier ($249.99/month).
The rebrand pattern continued on July 16, 2026, when Google renamed NotebookLM to Gemini Notebook — folding the research tool that launched as Project Tailwind at I/O 2023 into the Gemini brand. Google's post says it “remains a standalone product focused on being your premier research tool” while doing more across the Google ecosystem, and cites more than 30 million people and over 600,000 organizations using it. Shipping alongside the rename: a secure cloud computer per notebook that lets Gemini Notebook write and execute code natively against your sources (available at launch to Google AI Ultra users and Workspace customers with AI Ultra / AI Expanded Access, rolling out to all Pro users on the web), full cross-app syncing between the Gemini app and the standalone notebook experience, and a stated plan to bring notebooks into AI Mode in Search. The announcement, by Google Labs / Gemini app VP Josh Woodward, is at blog.google.
The Brain × DeepMind merger (April 2023)
On April 20, 2023, Sundar Pichai announced the merger of Google Brain and DeepMind into a single unit, Google DeepMind, with Demis Hassabis (DeepMind cofounder, then CEO) named CEO of the combined entity. The announcement is at blog.google/technology/ai/april-ai-update; DeepMind's side is at deepmind.google/blog/announcing-google-deepmind.
Jeff Dean, previously the Google Brain lead, was named Chief Scientist of Google + Alphabet, reporting directly to Pichai. The combined unit's accomplishments cited in the announcement: AlphaGo, the original Transformer paper, word2vec, WaveNet, AlphaFold, seq2seq, deep reinforcement learning, TensorFlow, and JAX. The merger is the corporate-side throughline that explains why Gemini was framed in December 2023 as “natively multimodal trained from the ground up” — Google's first model assembled by the unified post-merger team.
Demis Hassabis and DeepMind's John Jumper were jointly awarded the 2024 Nobel Prize in Chemistry with David Baker for AlphaFold's protein-structure-prediction work. The DeepMind announcement is at deepmind.google. Hassabis led Google DeepMind as CEO until August 5, 2026, when he became Chair of Google DeepMind and Chief Scientist of Alphabet and Koray Kavukcuoglu took over day-to-day leadership of the unit as SVP; that reshuffle, and Jeff Dean's departure in the same announcement, are covered in the people section below.
The February 2024 Gemini image-generation incident
Within days of the Bard–to–Gemini rename, users posted historically inaccurate images Gemini's image generator produced — racially diverse Nazi soldiers, multi-ethnic US Founding Fathers, Black female popes — that drew sharp criticism. Google paused people-image generation in Gemini on February 22, 2024. The explainer post by then-SVP for Search Prabhakar Raghavan said directly: “we got it wrong.”
On February 28, 2024 Pichai's internal memo — reported in NPR — called the outputs “completely unacceptable” and committed to structural changes, updated guidelines, and red-teaming. People-image generation was restored on August 28, 2024 with Imagen 3, gradually rolled out starting with Gemini Advanced / Business / Enterprise; TechCrunch coverage.
The episode is recorded here factually because it bears on the Gemini-1.0 / Imagen-2 rollout window and on Google's subsequent moderation framing for image generation. This page does not litigate the underlying decisions or their handling.
The DOJ search-monopoly case — US v. Google, D.D.C.
The case, US v. Google, D.D.C. case number 1:20-cv-03010 before Judge Amit P. Mehta, was filed October 20, 2020 and went to a non-jury trial in 2023–2024. The CourtListener docket is the canonical entry point.
The August 5, 2024 liability ruling held that Google “is a monopolist, and it has acted as one,” with Sherman Act §2 violations in the general-search-services and general-search-text-advertising markets. The September 2, 2025 remedies opinion rejected a Chrome divestiture but barred Google from exclusive distribution contracts for Search, Chrome, Google Assistant, and the Gemini app; ordered the search index and user-interaction data to be shared with rivals; and installed a six-year technical-oversight committee. The opinion explicitly states the judgment “covers Google's GenAI tool (Gemini), any successors, and any other Google products that fit the GenAI definition,” on the theory that Google cannot use its search monopoly to box out GenAI competitors. The DOJ press release is at justice.gov.
Final remedies were entered December 5, 2025; the case was formally terminated at the district level December 9, 2025 and is now in appellate proceedings on three tracks. Google filed its notice of appeal January 16, 2026, challenging the data-sharing requirements and the technical-committee oversight; DOJ and a majority of state AGs filed a cross-appeal by February 3, 2026, seeking the Chrome / Android divestitures and Apple-default-deal break-up that Judge Mehta declined to order. In a May 2026 order, Judge Mehta required DOJ to give 45 days' advance notice before any Qualified Competitor may access the data-sharing or syndication remedies — allowing Google to seek a stay at that juncture — but did not grant the blanket stay Google had sought. Google filed its 111-page opening appellate merits brief at the D.C. Circuit on May 22, 2026, contesting both the August 2024 liability ruling and the September 2025 remedies order and arguing the district court conflated harm to competitors with harm to the competitive process; the filing also objected that the data-sharing remedies open Google's index and user data to GenAI rivals such as OpenAI that “did not even exist” during the conduct at issue, per Courthouse News. DOJ and the co-plaintiff states filed their combined 144-page response and cross-appeal brief on July 28, 2026, asking the D.C. Circuit to affirm both the liability finding and the data-sharing remedies while vacating Judge Mehta's refusal to ban Google's search-distribution payments; the brief argues that “emerging [generative AI] products pose ‘a potential threat to Google's dominance,’ and a competitive window may close if Google can continue using its monopoly profits — fruits of its monopolizing conduct — to squelch the threat,” per Courthouse News. A wave of amicus briefs landed on August 4, 2026, several of them pressing the GenAI question directly: Public Knowledge asked the D.C. Circuit to affirm liability and send the case back for stronger relief — including a ban on all payments for default or preferential placement, and Chrome divestiture held in reserve if the current order fails — citing the growth of AI-assisted search and Google's continuing Apple partnership as evidence the same playbook will run again (press release), and the Open Markets Institute filed on the theory that Search is the foundation of Google's next generation of AI products, its counsel Tara Pincock framing the stakes as whether remedies arrive “before today's search monopoly becomes tomorrow's AI monopoly” (announcement). Google filed its reply brief on September 29, 2026, closing the main briefing on the consolidated appeal and cross-appeal; per MLex it again asks the D.C. Circuit to overturn the ruling, arguing the remedies would penalize Google and distort the market and that Apple and Mozilla would have chosen Google as the default without the agreements. On October 5, 2026 DOJ and Google jointly asked the court to give the cross-appeals priority and to hear argument as soon as December 9, 2026, per Law360; no order setting argument had surfaced as of October 19, 2026. Remedies implementation continues in the district court in parallel — on August 3, 2026 Judge Mehta granted the plaintiffs an unopposed extension to October 27, 2026 to submit the template Web Search Index and User-Side Data Sharing license agreement, the instrument through which a Qualified Competitor (including a GenAI rival) would actually obtain the shared index and user-interaction data, and the parties have been filing joint status reports on a roughly weekly cadence since the July 14, 2026 status conference (docket entries 1537, 1539, and 1542). Judge Mehta's August 17, 2026 minute order set the next of them for August 28, 2026; Google filed it on schedule (docket 1544), and on August 31, 2026 Mehta entered a further minute order calling for another joint status report by September 11, 2026, this one updating the court on the Technical Committee's proposal regarding indemnification. The plaintiffs filed it on schedule (docket 1548): it records that Google's questions at an August 25 meet-and-confer sent the plaintiffs back to the TC and its counsel, that a September 4, 2026 meet-and-confer cleared an interim issue so the TC could gather information bearing on how to resolve its indemnification concerns, and that the parties jointly propose the next report on that topic be due October 2, 2026. Mehta entered another minute order the same day on a separate instrument — an additional joint status report due September 16, 2026 on the status of the Source Code Agreement, with a status conference set for September 18, 2026. That sequence has since run its course: the Source Code Agreement report went in under seal on September 16 (docket 1549) with a public redacted version and sixteen exhibits the next day (docket 1551), Google filed a supplemental status report of its own on September 17 (docket 1550), and the United States filed its second status report on Google's compliance with the Final Judgment the same day (docket 1552). The September 18, 2026 status conference was held as scheduled, and Mehta entered a minute order from the bench setting two further deadlines: an additional joint status report on the Source Code Agreement by September 25, 2026, and a plaintiffs-only status report on enforcement of and Google's compliance with the Final Judgment by November 10, 2026. The transcript of the conference issued September 19, 2026 (docket 1553). The parties filed the Source Code Agreement report on schedule on September 25, 2026 (docket 1555), reporting that they had reached final agreement on the template Source Code Agreement — the instrument under which Technical Committee members and staff get access to Google's source code, algorithms, and other sensitive information — and attaching it as Exhibit A; Google had also proposed a Supplemental Protective Order, which the plaintiffs were still reviewing (report, template agreement). The indemnification report went in on October 2, 2026 (docket 1557), and on October 5 Mehta called for another by October 23, 2026. October was otherwise about syndication — the remedy that lets a Qualified Competitor resell Google's search results and search text ads while it builds its own index. After an unopposed extension granted October 6 (Google had delivered roughly 240 objections to the draft on September 29 and wanted time to update them), the plaintiffs and the Technical Committee filed their Joint Recommendation for a template Search and Search Text Ads Syndication License Agreement on October 9, 2026 (docket 1559 under seal; redacted at 1560), with Google's roughly 220 remaining objections attached. The template is written “to apply equally to any Qualified Competitor, including providers of General Search Engines, Search Text Ads, or GenAI Products,” caps a competitor's use of Google syndication at 40% of its annual US queries in the first year, and defers the tapering schedule for later years to a supplemental recommendation; a Google-redacted copy of the template itself followed on October 13 (docket 1561). The joint status report on the outstanding objections was due October 16, 2026, and the United States filed a sealed document with seventeen exhibits that day (docket 1562). Still ahead: the template Web Search Index and User-Side Data Sharing license agreement, due October 27, and the plaintiffs' compliance report, due November 10. Separately, on August 25, 2026 he had entered a five-page order (docket 1543) resolving the issues the Technical Committee had raised at that July 14 conference: it shields TC staff from being compelled to testify about work done under the Final Judgment, and it defines the post-service employment bar by naming the entities that count as “Competitor” for a departing committee member — a list that runs Anthropic, Brave, DeepSeek, DuckDuckGo, Ecosia, Meta, Microsoft, OpenAI, Perplexity, Yahoo, and SpaceXAI (formerly xAI), alongside any certified or applicant Qualified Competitor. The bar is written against work supporting a general search engine, search text ads, or “GenAI Products” — the same Final-Judgment category that covers the Gemini app. The ad-tech case, US v. Google in the Eastern District of Virginia (case 1:23-cv-00108, Judge Brinkema), held Google liable on April 17, 2025 for monopolizing the publisher ad-server and open-web display ad-exchange markets; remedies-phase closing arguments were November 2025, and Judge Brinkema issued the remedies ruling on September 2, 2026 (docket entry 1857). She rejected the plaintiffs' structural remedies — the divestiture of AdX, the open-sourcing of DFP's final auction logic, and the contingent divestiture of DFP Remainder — and accepted most of the parties' proposed behavioral remedies, as modified by the court, ordering the parties to meet and confer and file one jointly proposed Final Judgment within 30 days, with any still-disputed provision submitted in both sides' versions and labeled as to proponent. The accompanying memorandum opinion (docket entry 1858) was filed under seal for 14 days because so much sealed material was referenced at trial — and on September 16, 2026, no redactions having been sought, Judge Brinkema ordered it unsealed (docket 1860, following a joint notice at docket 1859). The 106-page opinion is now public. It adopts Google's proposed six-year term for the Final Judgment over the plaintiffs' fifteen — matching the six-year term in the D.D.C. search case — with the court retaining authority to extend it if the judgment is not fully satisfied; sets the effective date 60 days after entry except for appointing the Monitor and forming the Technical Committee, which begin immediately; expects implementation within roughly 15 months; and applies the judgment globally rather than only in the United States. On the question this page cares about, the opinion treats generative AI as a live but not-yet-arrived disruption: “The integration of AI within the ad tech industry is in its early stages,” and while the D.D.C. search court found that AI — “particularly generative AI” — may “prove to be game changers” and “wrest[] market share from Google” in general search, “[h]ere, by contrast, AI has not yet had such a disruptive impact on the ad tech industry.” The court nonetheless cites “the rapid emergence of AI and supply-path optimization” as likely to change the industry dramatically, and uses that expectation to justify the shorter term. The PDF is on CourtListener's RECAP archive. The parties filed their jointly proposed Final Judgment on October 2, 2026, the last day of the 30-day window (docket 1861; the 41-page proposed judgment is attached). It carries the six-year term and the Monitor and Technical Committee machinery from the opinion, and where the two sides could not agree it shows both versions labeled by proponent — Google wants the effective date made “subject to the outcome of any motion to stay this Final Judgment pending appeal” and the Monitor's term tied to the judgment's, and the AdX–Prebid and AdX–publisher-ad-server interoperability deadlines read “6” months in the plaintiffs' version and “12” in Google's. The September 2 order says the court will notify the parties if it wants further briefing or argument on the disputed provisions. On an October 19, 2026 check of both CourtListener dockets, that October 2 notice was still the ad-tech case's Date of Last Known Filing, and the search case's newest entry was the October 16 sealed filing above. Both cases bear on Gemini distribution — verify the docket status at write time.
The Anthropic investments
Google has been one of Anthropic's two major cloud and capital partners alongside Amazon, which remains Anthropic's primary cloud provider. The October 27, 2023 $2 billion commitment ($500M upfront, $1.5B over time) was reported in CNBC, alongside Anthropic's $3B Google Cloud commitment over four years and adoption of TPU v5e for Claude inference.
On April 24, 2026, Google announced an additional commitment of up to $40 billion in Anthropic — $10B initial, with the remaining $30B contingent on performance milestones — at Anthropic's then-latest valuation of $380 billion. Coverage in TechCrunch, CNBC, and Bloomberg, which first reported the deal.
On the compute side, Anthropic announced on April 6, 2026 a new agreement with Google and chip partner Broadcom for multiple gigawatts of next-generation TPU capacity expected to come online starting in 2027 — 5 gigawatts over five years per CNBC — which Anthropic called “our most significant compute commitment to date” in its own announcement. That post also notes that Amazon remains Anthropic's primary cloud provider and training partner; Google is a capital partner and one of three platforms Claude runs on. A widely-circulated figure putting Anthropic's Google Cloud spend at $200 billion over five years originates with The Information (May 5, 2026) and was relayed by Reuters, which noted it could not verify the report and that neither company commented; neither Google nor Anthropic has confirmed it, so it is recorded here as reported, not confirmed. The scale of the bet surfaced in Alphabet's Q2 2026 results, reported July 22, 2026: roughly $99 billion in quarterly gains on equity investments — primarily the Anthropic and SpaceX stakes, after Anthropic's May 2026 raise at a $965 billion valuation — drove a record $112.1 billion net-income quarter, per Fortune; Bloomberg reported July 23 that Alphabet's SEC filing valued its private-company investments — primarily the Anthropic stake — at roughly $124.3 billion as of June 30, 2026. The Anthropic relationship is the corporate-strategy counterweight to the in-house Gemini line.
Project Astra, Gemini Live, and Project Mariner
Project Astra — the “universal AI agent” prototype with the smart-glasses demo — was first unveiled at Google I/O on May 14, 2024 as a research preview. The DeepMind page is at deepmind.google/models/project-astra. Astra's multimodal real-time understanding capabilities folded into Gemini Live (consumer) and into Search's AI Mode (May 2025).
Gemini Live, the conversational voice surface (10 voices, interruptible, runs in background), launched alongside the Pixel 9 series on August 13, 2024, initially as a Gemini Advanced subscriber feature. Camera and screen-sharing capabilities followed in 2025 as Astra's surface migrated into the Gemini app. On August 26, 2026 Google gave Live its own agentic upgrade: it now hands multi-step, long-running jobs off to Gemini Spark (which works across Docs, Sheets, Drive, and the web and keeps running across days or weeks), reads out a spoken Daily Brief assembled from Gmail and Calendar, and searches, summarizes, stars, archives, or deletes mail hands-free. Google says 63% of Gemini users talk to it out loud; the post is at blog.google. On September 15, 2026 the surface got a new engine: Gemini Live in the app now runs on Gemini 3.8 Live Extended Thinking, which reasons in the background while it keeps speaking, and the same release put Gemini 3.8 Live behind Search Live and Extended Thinking behind Workspace's Docs, Gmail, and Keep Live surfaces — see the Gemini 3.8 Live row above. On October 1, 2026 Live gained Guided Vision on compatible Android devices: with the camera shared, it describes the surroundings, answers follow-ups, and speaks reframing cues (“pan slowly to the right”) — built with the blind and low-vision community and trained with Aira's visual interpreters, per the post by Isha Sheth and Scott Adams at blog.google.
Project Mariner — the Gemini 2.0–powered research-prototype Chrome extension that controls the browser (cursor, clicks, forms) — was announced December 11, 2024 alongside the Gemini 2.0 Flash Experimental release. Single-tab and foreground-only at launch; 83.5% on the WebVoyager benchmark. Mariner's computer-use capability was folded into Gemini 2.5 Flash by I/O 2025 and into the Antigravity agent platform that launched alongside Gemini 3 Pro. Antigravity has since grown a multi-agent framework in which autonomous teams of agents critique and iterate on one problem over hours or days; on August 31, 2026 Google reported running it on Gemini 3.7 Flash to close seven open problems in math and theoretical computer science — among them Knuth's Cycles Conjecture, with a 40-plus-page proof machine-verified in Lean — alongside 71% on TCSBench, a cycle-accurate out-of-order RISC-V simulator that boots xv6, and accepted upstream optimizations to Eigen and ParlayHash. All figures vendor-stated; the post is at blog.google. The same agentic-loop pattern reached video on September 1, 2026, when Google shipped agentic video understanding for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite: instead of ingesting a video at a fixed frame rate (1 FPS by default), the model decides what to watch, at what speed, and through which modality — frames, audio, or transcript — calling an internal tool to load only the segments it needs. Vendor-stated gains across standard video benchmarks are up to 88% fewer tokens, 66% lower cost, and 7% better accuracy, concentrated on long-form video; it is enabled by setting the API configuration to “agentic” in Google AI Studio or the Gemini Enterprise Agent Platform. The announcement, by Rohan Doshi (Senior Product Manager) and Mario Lučić (Research Director), both of Google DeepMind, is at blog.google. The managed agent itself was re-cut on September 17, 2026: antigravity-preview-09-2026 replaced and deprecated antigravity-preview-05-2026, which shut down October 5, 2026. Remote-sandbox callers reading only output_text or model_output need nothing but the new agent string, but anyone running tools locally or parsing function_call steps faces a rewritten built-in toolset — PascalCase parameters instead of snake_case, line-range replacements (replace_file_content) instead of full-file rewrites, and new find_by_name / grep_search tools where agents previously shelled out; shell execution and Google Search are unchanged. The change is recorded under the September 17, 2026 heading of the Gemini API changelog. Since October 13, 2026 the agent can also run on Gemini 4 Argon (set model in agent_config), and the same day Google put Antigravity's agentic coding interface directly into the Gemini macOS app and added a chat-to-tasks toggle in the main chat window, per the Gemini app post; Google says Argon in Antigravity raised its own engineers' agentic code submissions by more than 35% while cutting rollbacks by nearly a third (vendor-stated).
At Google I/O 2026 on May 19, 2026, Google launched Gemini Spark — a 24/7 personal AI agent that runs on Gemini 3.5 Flash via the Antigravity platform. Spark operates in the background and is designed to take actions on the user's behalf (scheduling, web tasks, custom sub-agents, authorized payments) while surfacing confirmations before major actions; the announcement is at blog.google. The initial rollout was to trusted testers and Google AI Ultra subscribers in the US at the end of May 2026; Spark began rolling out to Google AI Pro subscribers in the US on July 23, 2026, with other countries to follow, per 9to5Google. On the July 22, 2026 Alphabet Q2 earnings call, Sundar Pichai described Spark as “now available in the U.S. and internationally” (transcript). On July 30, 2026 Google widened both the surface and the geography: Spark gained a Chrome auto-browse integration that, with permission, uses the visitor's logged-in accounts and saved passwords to run web errands end to end — handing sensitive steps such as payments back to the user and screening for prompt injection — and AI Pro access opened in 160+ additional countries; the Chrome capability is US-only at launch, per blog.google. On September 9, 2026 Google added a Google Photos connection next to the Chrome one — Spark can now edit photos and curate albums on request — still scoped to Google AI Pro and Ultra subscribers in the US, per the fall plan-update post at blog.google. On September 30, 2026 the reusable-instruction mechanism Spark already used, skills, rolled out into Gemini chat globally (invoked with a leading slash, stackable, and now able to carry reference files), with Google saying skills will replace Gems in the coming weeks, per blog.google. Gemini Spark is the consumer-facing instantiation of the agentic direction that Astra and Mariner prototyped in 2024–2025.
The agentic pattern reached physical robots on July 30, 2026 with Gemini Robotics ER 2, Google DeepMind's embodied-reasoning model — framed in the announcement as “a high-level brain for robots” that talks to humans, plans multi-step tasks, natively calls tools such as Google Search, and hands motor execution off to a lower-level vision-language-action (VLA) model. It is built on Gemini 3.5 Flash and ships as two preview endpoints, gemini-robotics-er-2-preview and the Live-API streaming variant gemini-robotics-er-2-streaming-preview, publicly available to developers through the Gemini API and Google AI Studio and in private preview on the Gemini Enterprise Agent Platform. The headline gains over Gemini Robotics ER 1.6 (April 14, 2026) are continuous-video progress tracking — 57.4% on progress classification, 91.3% accuracy and 0.96s mean absolute distance on moment-finding, all vendor-stated — and multi-robot collaboration, demonstrated with Boston Dynamics' Spot and with Apptronik's Apollo 2 working alongside a Franka F3 Duo. gemini-robotics-er-1.6-preview shut down on August 31, 2026 per the deprecations page, leaving the two ER 2 endpoints as the only served embodied-reasoning ids — though the summary models page has not caught up: at its October 13, 2026 stamp it still lists ER 1.6 as a served specialized-task model, six weeks after the shutdown date the deprecations page gives — so this is a stale roster entry rather than a rebuild lag. The announcement is at blog.google, the model card on the DeepMind site, and the developer guide at ai.google.dev. ER 2 was one of three robotics models announced that day: Gemini Robotics 2, the vision-language-action model that converts vision and language input into motor control and is the first in the line to drive a full humanoid “from feet to fingertips” as well as bi-arm robots; Gemini Robotics On-Device 2, the efficiency-optimized VLA that runs locally on robot hardware and adapts to an entirely new robot embodiment in a few hours of data; and ER 2 itself, the reasoning layer that sits above them. Only ER 2 is publicly available — both VLA models are limited to early-access partners — per the DeepMind post at deepmind.google. Robotics models are deliberately not enumerated as rows in the version table above — same convention as the embedding and other specialty lines.
Imagen and Veo
The image and video lines under the Gemini umbrella ship on a parallel cadence to the chat models. Imagen 1 (May 2022) was a Google Brain research model; Imagen 2 (December 2023) powered Bard's image generation and was the model behind the February 2024 incident; Imagen 3 (announced May 14, 2024 at I/O; ImageFX rollout August 15, 2024; people-image return August 28, 2024) introduced the photorealistic-rendering jump. Imagen 4 shipped at I/O 2025 (May 20, 2025) in Standard, Fast, and Ultra variants with major text-rendering improvements and up to 2K resolution; Workspace Docs / Slides / Vids integration followed. The Imagen 4 line is now deprecated — imagen-4.0-generate-001, -fast-generate-001, and -ultra-generate-001 shut down August 17, 2026, announced in the June 15, 2026 Gemini API changelog entry; the deprecations page now names gemini-nano-banana-2.1 as the recommended replacement for all three.
Nano Banana Pro (November 20, 2025; preview id gemini-3-pro-image-preview, stable GA id gemini-3-pro-image released May 28, 2026) followed Gemini 2.5 Flash Image as a professional design engine with a reasoning core — studio-quality 4K visuals, complex layouts, and precise text rendering, framed as the high-end successor to the original Nano Banana model. Nano Banana 2 (February 26, 2026; preview id gemini-3.1-flash-image-preview, stable GA id gemini-3.1-flash-image released May 28, 2026) shipped as the speed-and-volume sibling on the 3.1 base — high-efficiency production-scale visual creation, optimized for the rapid-iteration use cases the original Nano Banana shape was named for. Both preview ids are deprecated and shut down June 25, 2026 per the Gemini API deprecations page; the stable ids are documented at ai.google.dev/gemini-api/docs/image-generation. Nano Banana 2 Lite (gemini-3.1-flash-lite-image, stable GA released June 30, 2026) followed as the efficiency specialist of the family — ultra-low latency (text-to-image in ~4 seconds), $0.034 per 1K-resolution image, and 1K output only — and is now Google’s recommended replacement for the original Nano Banana (gemini-2.5-flash-image) on high-volume, cost-sensitive pipelines; the announcement is at blog.google. Nano Banana 2.1 (gemini-nano-banana-2.1, stable GA October 6, 2026) is the first image id to drop the Gemini-version prefix from its string. Google describes it as an update to Nano Banana 2 at Flash-level speed and cost, with better visual quality, prompt adherence, multi-turn character consistency, and text rendering; fixed tiling artifacts on the 1:4, 4:1, 1:8, and 8:1 panoramic aspect ratios at 2K and 4K; fusion of up to 14 reference images; Google Web and Image Search grounding; and configurable thinking (minimal / medium / high). Priced at $30 per million image output tokens — about $0.0336 per 1K image, $0.0504 per 2K, $0.113 per 4K — per the pricing page. It landed quietly: the entry under the October 6, 2026 heading of the Gemini API changelog, the model page, and a generally-available listing on the Vertex AI roster, with no launch post on blog.google. The same changelog entry deprecates gemini-3.1-flash-image (Nano Banana 2) with no shutdown date and points it at 2.1; the deprecations table now also names 2.1, rather than Nano Banana 2, as the replacement for the retired Imagen 4 ids and the old image previews. The Nano Banana shape now runs from the 2.5 Flash Image table row above through Pro, 2, 2 Lite, and 2.1 in this prose section — the original codename, the professional tier, the balanced workhorse, the speed/cost tier, and the workhorse's refresh.
Veo 1 debuted as a Vertex research preview at I/O 2024. Veo 2 (December 16, 2024) brought 4K resolution and clips up to two minutes. Veo 3 at I/O 2025 was the first Veo with native synchronized audio — dialogue, sound effects, and ambient sound generated in the same pass — available initially through the AI Ultra tier and via the Google Flow filmmaking surface. Veo 3.1 Lite Preview (veo-3.1-lite-generate-preview, March 31, 2026) is the cost-efficient, developer-first addition to the Veo 3.1 family for rapid iteration and high-volume applications, per the Gemini API changelog. The whole Veo 3.1 preview trio on the Gemini API — veo-3.1-generate-preview, veo-3.1-fast-generate-preview, and veo-3.1-lite-generate-preview — now carries an October 22, 2026 shutdown date on the deprecations page, with gemini-omni-1.1-flash as the recommended replacement — which, if it holds, leaves the Gemini API with no Veo id at all and Omni as the only video line there. The earlier veo-2.0-generate-001, veo-3.0-generate-001, and veo-3.0-fast-generate-001 ids shut down June 30, 2026 (June 15, 2026 changelog entry); the Veo 3.1 preview ids and the GA models on the Gemini Enterprise Agent Platform are the recommended replacements.
At Google I/O 2026 on May 19, 2026, Google introduced Gemini Omni Flash, a multimodal “any-input, any-output” video generation and editing model that accepts text, image, video, and audio inputs and generates high-quality video output. The I/O announcement is at blog.google. Omni Flash is designed for creative workflows — scene transformation, action reimagining, multi-turn conversational video refinement, and avatar-based video creation — and is built on Gemini’s world knowledge for physically plausible scene generation. On June 30, 2026, Google brought Omni Flash to developers for the first time as gemini-omni-flash-preview (public preview) via the Gemini API, Google AI Studio, and the Gemini Enterprise Agent Platform, alongside the Gemini app and Google Flow; it is priced at $0.10 per second of video output (matching Veo 3.1 Fast) and generates 10-second clips at launch, with SynthID watermarking on all output. The joint developer-availability post is at blog.google; the developer docs are at ai.google.dev/gemini-api/docs/omni. Image and video models are deliberately not enumerated as separate rows in the version table above (Gemini 2.5 Flash Image is the one row exception); if either line grows substantially, a dedicated /ai/gemini/image-models/ sibling page is in scope.
Omni Flash reached general availability on August 27, 2026 as gemini-omni-1.1-flash, announced by Google DeepMind product managers Anish Nangia and Alisa Fortin at blog.google, with the GA entry under the August 27, 2026 heading of the Gemini API changelog. The 1.1 release is about control rather than a new base model: scene extension in 10-second increments to a cumulative 40 seconds, reading up to 10 seconds of prior context where earlier models referenced only the final second; first-and-last-frame interpolation for generating the footage between two keyframes; a resolution parameter covering 360p, 720p (default), 1080p, and 4K, with the two top rungs produced by upscaling; and video references of up to three seconds in the multimodal input for character consistency. The 360p tier is the draft mode — Google states up to 60% faster generation and a third of the 720p cost. Billing moved to tokens: $1.50 per million input tokens and $17.50 per million video output tokens, which at 5,792 tokens per second of 720p video works out to roughly the same $0.10 per second the preview charged, per the pricing page. Google names Adobe (Firefly), Figma Weave, Runway, and GMI Cloud as production users. The preview id gemini-omni-flash-preview is deprecated; its shutdown, originally set for September 30, 2026, was pushed to October 22, 2026 on the deprecations page, the same day as the Veo 3.1 previews. Note the split availability: the Gemini API lists gemini-omni-1.1-flash as Stable GA while the Vertex AI roster still files both Omni ids under Preview at its October 15, 2026 stamp.
People who shaped Gemini
Sundar Pichai — CEO of Google since 2015, CEO of Alphabet since December 2019. The 2023 ChatGPT-response framing, the Brain × DeepMind merger, the December 2023 Gemini 1.0 launch, the February 2024 image-generation memo, and the strategy around the DOJ search remedies all run through Pichai's office.
Demis Hassabis — CEO of Google DeepMind from the April 2023 merger until August 5, 2026, and Chair of Google DeepMind and Chief Scientist of Alphabet since. DeepMind cofounder. 2024 Nobel laureate in Chemistry (jointly with then-DeepMind colleague John Jumper and David Baker) for AlphaFold. The technical-leadership face of the post-merger Gemini program.
Google restructured its AI leadership on August 5, 2026. In a joint note to Google DeepMind staff published at blog.google, Pichai wrote that Hassabis would hand over his day-to-day operational responsibilities to become Chair of Google DeepMind and Chief Scientist of Alphabet while continuing to lead Isomorphic Labs, and that Koray Kavukcuoglu — Google DeepMind's Chief Technology Officer and Google's Chief AI Architect, at DeepMind for 13 years and a lead on WaveNet and DQN — would step up as SVP of Google DeepMind, reporting directly to Pichai, overseeing Gemini model development, Frontier AI research, and the Gemini app and developer teams. The same note announced that Jeff Dean is leaving Google after 27 years, together with Google Senior Fellow Sanjay Ghemawat, to launch an independent public benefit corporation aimed at accelerating discoveries in machine learning, science, and engineering; Pichai said Google will continue with them “as a founding investor and Cloud partner.” Dean named the company Discovery Loop in a follow-up post on X, per CNBC.
Google's note named two of the four Discovery Loop founders. Dean's own announcement the same day, and the company's site, add Oriol Vinyals — VP of research at Google DeepMind, AlphaStar co-lead, and a Gemini co-technical lead — and Quoc V. Le, a Google Brain cofounder whose work centered on automating model design. Both posted their own farewells to Google DeepMind on August 5, 2026; Vinyals wrote that Dean, who had recruited him to Google 13 years earlier, was “challenging me to take the next leap.” That detail matters to this page because Dean and Vinyals were the two people publicly identified as Gemini's co-technical leads, and both left on the day Kavukcuoglu was given the next flagship. The company is a public benefit corporation aimed at automating the research loop — starting with machine-learning research itself — and its seed round was co-led by Radical Ventures and Khosla Ventures, with Lightspeed, Kleiner Perkins, Doerr Capital, and Alphabet participating; Radical's Jordan Jacobs took a board seat and Google supplies the first year of compute. Reporting: The New York Times, Steven Levy's founder interviews in Wired, and TNW, which first flagged the gap between Google's two-name memo and the startup's four-name masthead. No party has disclosed the round size or valuation. This page records the departures; it does not handicap what they mean for the lab.
Two senior departures landed in June 2026. Noam Shazeer — VP of engineering and a co-lead of the Gemini models, co-author of the 2017 “Attention Is All You Need” paper that introduced the Transformer, and cofounder of Character.AI — announced on June 17, 2026 that he was leaving for OpenAI, less than two years after Google brought him back in the reported $2.7B Character.AI licensing deal of August 2024; CNBC covered the move. Days later, John Jumper — the AlphaFold lead who shared the 2024 Nobel with Hassabis — announced he was leaving Google DeepMind for Anthropic, per Fortune (June 23, 2026). Neither has publicly stated a reason. Recorded here because both sat at the technical center of the Gemini and AlphaFold programs; this page does not handicap what the departures mean for the lab.
Jeff Dean — Chief Scientist, Google + Alphabet from the April 2023 merger until his August 2026 departure, reporting directly to Pichai. Previously co-led Google Brain. Sissie Hsiao led the Bard product through the rename. Eli Collins (VP, Google DeepMind product) co-launched Bard alongside Hsiao. Prabhakar Raghavan (then SVP, Search) authored the February 2024 image-generation explainer; he transitioned out of Search leadership in October 2024.
Current leadership beyond Pichai / Hassabis / Kavukcuoglu shifts more frequently and should be re-verified at each refresh run; the Musk v. Altman–style governance episode that has played out at OpenAI does not have a Google analog.
The competitive landscape
Gemini competes most directly with OpenAI's ChatGPT (the line that Bard was Google's reactive response to in 2023; see ChatGPT Versions), Anthropic's Claude (Google's primary capital and cloud partner alongside Amazon; see Claude Versions), xAI's Grok (see Grok Versions), and Meta's Llama. Gemini's distinguishing positioning has been the integration into Search, Workspace, Pixel devices, and Chrome alongside the standalone Gemini app, and the on-device Nano line that no other major frontier lab has matched at scale. This page does not attempt a benchmark roundup or a ranking.