Gemini

Gemini in August 2026 is not simply Google’s chatbot. It is a broad AI layer spread across the Gemini web and mobile apps, Android assistant behavior, Google Workspace, Deep Research, Notebook-style research, creative image/video/audio tools, Google Search, developer surfaces and a newer agentic layer built around Gemini Spark.

That breadth is the main reason Gemini matters. If your files, email, calendar, documents, phone, browser and productivity workflow already live inside Google, Gemini can operate closer to the information you actually use than a standalone chatbot can. It is also the source of the product’s biggest evaluation problem: there is no single “Gemini model” or single “Gemini experience” that every user receives.

Google’s current Android listing describes the mainstream app around Gemini 3.6 Flash and Gemini Omni, while Google’s help documentation often abstracts the consumer lineup into generic Gemini 3 Flash/Pro classes. Gemini Spark is different again: Google explicitly says Spark moved to Gemini 3.7 Flash on August 13.¹²³ The current high-quality independent benchmark evidence for 3.7 Flash therefore tells us something important about Spark and the model layer, but it cannot simply be pasted onto every ordinary Gemini conversation.

This review preserves those surfaces separately.

The current Gemini app is extraordinarily broad. Google’s mobile listings describe text, files, spreadsheets, photos and video input; Gemini Live with camera and screen sharing; image generation/editing; video and music creation; Notebook-style research; educational tools; prototyping; and upgrades into Google AI Plus, Pro and Ultra.⁴¹

The product’s deepest advantage is context from Google services. A user can bring Gmail, Drive, Calendar and other Google data into Gemini, subject to settings and permissions. Workspace administrators can also restrict whether Gemini can access organizational Drive or Gmail data, and content-owner permissions continue to apply.⁵

This is more than convenience. An AI assistant is often limited less by raw reasoning than by the cost of giving it the right context. Google already owns many of the places where users keep that context.

G2’s current review data reflects this. Its AI Chatbots category gives Gemini 4.4/5, with strong ease-of-use and understanding scores and slightly weaker reliability. G2’s own long-term editorial review highlights Workspace integration, research and document-heavy workflows as core strengths while still noting logical errors, context loss and hallucinations.⁶

Capterra’s current 4.6/5 sample tells a similar story: verified 2026 reviewers repeatedly praise multimodality, Workspace integration, productivity and creative features, while flagging quotas, support friction, occasional inaccurate or generic responses and limited advanced customization.⁷

TechRadar provides one of the most useful current consumer-level tests because it fed Gemini 3.6 Flash and GPT-5.6 Sol the same messy personal dataset: bank statements, calendar entries, receipts, Gmail, WhatsApp screenshots, photos, a utility bill, travel information and handwritten notes.⁸

Gemini was fast and useful at summarization and categorization. But in that test it failed to make enough connections between the inputs. It missed a calendar conflict, did little with several pieces of evidence and produced a weaker prioritization than ChatGPT. GPT-5.6 Sol took longer but generated a more integrated action plan.⁸

That is not evidence that Gemini is generally inferior. It is evidence that the mainstream 3.6 Flash consumer experience should be evaluated independently from 3.7 Flash model benchmarks. For users who want rapid organization and Google-connected context, the trade-off may be acceptable; for users who need deeper cross-document inference, the difference can matter.

Spark is where Gemini changes category. Google describes it as a personal AI agent that can automate complex workflows, maintain schedules and work from connected apps, skills, chats, websites, Personal Intelligence, location and other context.⁹

Google’s August 13 model announcement says Spark now uses Gemini 3.7 Flash.³ Tom’s Guide’s August 14 hands-on test is therefore unusually current: the reviewer gave Spark access to a large Gmail/Drive life-admin problem. Spark found missed deadlines and billing issues, linked to source material and organized an action plan while generally preserving approval boundaries. It still missed some emails and unnamed Docs.¹⁰

That is a much more useful product test than a synthetic prompt. It demonstrates why Spark exists: not merely to answer, but to search personal context, maintain state and help execute a workflow.

Spark now extends that pattern into Google Photos. Google documents single-prompt workflows that can search and curate photos, edit copies, create private or shared albums, extract text, use photo information in connected apps and schedule recurring photo-management tasks. Original images are not overwritten; edits create new copies. Google requires confirmation before some outward actions such as creating shared albums or sending email, and new albums are private by default.¹¹

The rollout is still narrow. Google says the Photos integration is gradual, US-only for eligible users, English-only, and available through the Gemini mobile app or web; TechCrunch independently reports rollout to eligible AI Pro and Ultra subscribers in the US with no broader international timetable announced.¹¹¹²

Google’s own safety language is explicit about the downside. The Spark help page warns users not to insert access credentials, payment data or sensitive information directly into task conversations; says Spark can make mistakes; warns against sensitive scheduled tasks; and notes that an offline scheduled task may complete an unintended action before the user can intervene.⁹

When Spark uses Chrome auto-browse, it can operate on sites where the user is logged in, share information needed to complete tasks with third parties and, with permission, use credentials stored in Google Password Manager. Google says each web-browsing task requires confirmation and users can stop browsing or cancel the task.⁹

Spark is not currently a universal feature. Google’s current help page says it requires a personal Google account, age 18+, Google AI Pro or Ultra and Keep Activity enabled. It also explicitly says Spark is currently unavailable in the European Economic Area, Nigeria, Switzerland and the United Kingdom.⁹

For an Italian reader, this is decisive: Spark should not be treated as an immediately available benefit of a Pro subscription today, even though Google’s global marketing describes Spark as available across a large number of countries. Regional support rules are the controlling product surface.

Google’s current consumer limits documentation gives free users a 32K-token context window, AI Plus 128K and AI Pro/Ultra one million tokens.² That is useful for long documents, code bases, research packets and multi-file work.

Deep Research is available across the plan structure with materially different usage allowances and can consume more compute than ordinary chat.² Google’s broader product stack also increasingly connects Gemini with notebook-style source-grounded research, which reduces the need to force every task through one undifferentiated chat interface.¹

User evidence supports the usefulness of this area. Crozdesk’s current Gemini page reports 4.39/5 from 569 user ratings and includes 2026 reviews praising 100+ page document handling, coding and Workspace integration while still warning about math errors, robotic language and hallucination loops.¹³

Gartner Peer Insights gives Gemini 4.4/5 from 54 ratings, with current reviews highlighting large PDFs, multimodal data, coding and productivity but still advising users to check specialized or quantitative outputs.¹⁴

Two late-August releases expand Gemini’s media stack without changing the review’s product-surface rule. Gemini 3.5 Transcribe, announced August 26, is a dedicated speech-to-text model for real-time and recorded transcription. Google positions it for intelligent cleanup and intent capture; independent reporting confirms features such as filler-word cleanup, custom vocabulary, speaker attribution and word-level timestamps. This is a transcription/model surface, not evidence that every ordinary Gemini chat has a new reasoning model. ¹⁵¹⁶

Gemini Omni 1.1 Flash, announced August 27, makes Google’s developer video-generation surface production-ready with scene extension, first/last-frame interpolation, faster low-resolution prototyping and up-to-4K output. Independent coverage corroborates the added video controls and broader rollout. Again, this is creative/product-surface evidence rather than a reason to transfer video-model claims into ordinary Gemini chat quality scoring. ¹⁷¹⁸

Gemini is no longer only a language assistant. Google’s mobile listings combine image generation/editing, video generation, music creation, camera/screen-based Gemini Live and rich multimodal file analysis.⁴¹

This matters because the purchasing comparison changes when one subscription bundles language work with media creation and Google storage. A user paying for Google AI Pro is not buying only a chatbot model; the plan includes broader Google ecosystem benefits and AI inside productivity applications.¹⁹

The review evidence is positive but not uniformly so. Capterra users praise image generation and multimodal work, while some complain about hallucinated edits or limits in advanced design.⁷ App Store reviews similarly show both enthusiasm for versatility and frustration when generated media ignores specific instructions.⁴

The reasonable conclusion is that Gemini’s creative breadth is a real product advantage, not proof that each creative component replaces a dedicated professional tool.

The mass-market ratings are strong.

Google Play currently shows 4.6/5 from about 41.8 million reviews and more than one billion downloads.¹ The Italian App Store shows 4.7/5 from about 207,000 ratings.⁴ Those samples are enormous, but they measure app/service satisfaction across versions rather than controlled current-model performance.

TrustRadius gives Gemini 8.5/10 from 162 reviews and ratings.²⁰ Gartner Peer Insights gives it 4.4/5 from 54 ratings.¹⁴ PeerSpot’s smaller enterprise-oriented sample is 8.0/10 from 18 reviews, with 88% willing to recommend.²¹

YouGov provides a different lens from review sites. Its August UK ranking gives Gemini net Satisfaction of 46.5 among current and former users and 64.2 among current users.²² In the July US AI Index snapshot, Gemini scores 17.6 on Quality, 17.4 on Value, 56.0 on Satisfaction and 15.8 on Recommend.²²

Those are perception metrics, not model benchmarks. They are useful because they capture something technical tests do not: whether people feel good about the product in ordinary use.

Eligible Workspace deployments operate under a different regime. Google says users accessing Gemini as a core Workspace service receive enterprise-grade data protections; their content is not human-reviewed or used for generative-AI model training outside their domain without permission.²³

That separation matters for organizations. A company evaluating Gemini for internal documents should not base its conclusion solely on the consumer Privacy Hub, just as a consumer should not assume enterprise contractual controls automatically apply to a personal Google account.

Admins can also limit which Workspace data Gemini can access, and ordinary document permissions remain binding.⁵

Google AI plans are bundled Google One subscriptions rather than standalone chatbot-only plans.

On the current US reference plan page, Google AI Plus is $9.99/month and Google AI Pro is $19.99/month.²⁴ Google’s plan comparison confirms separate Ultra 5x and Ultra 20x tiers, with 5x and 20x Pro usage respectively, although the exact price field is rendered dynamically by locale in the public page captured for this review.¹⁹ WIRED’s current July US pricing snapshot reports Ultra 5x at $100/month and Ultra 20x at $200/month.²⁵

The plans differ by more than model access. Current Google documentation gives AI Plus two times standard Gemini usage, Pro four times, Ultra 5x five times Pro and Ultra 20x twenty times Pro.²¹⁹ Storage rises from hundreds of gigabytes on Plus to 5 TB on Pro and 20/30 TB on the Ultra tiers, and higher plans bundle more Google AI and media-generation benefits.¹⁹

The downside is predictability. Since May, Gemini Apps have compute-based limits rather than a simple fixed number of prompts. Google says usage depends on prompt complexity, model/feature choice and chat length, with limits refreshing every five hours until the weekly limit is reached.²

If a paid user reaches a model limit, Google says the conversation can continue with Flash-Lite until the higher-tier limit resets.² That is preferable to a complete lockout, but it also means a subscription does not guarantee a constant model experience through heavy use.

WIRED’s July analysis captures the buyer problem well: the new system makes the relationship between subscription price and actual usable high-end capacity harder to predict than a fixed prompt quota.²⁵

Google launched Gemini 3.7 Flash on August 13 as its most intelligent “workhorse” model for coding and agents. Google says it improved software engineering, knowledge work and web-development workflows while launching at introductory API pricing of $0.75 per million input tokens and $3.75 per million output tokens.³

Independent testing supports the idea that it is a strong current model.

ARC Prize verifies Gemini 3.7 Flash across three reasoning levels. High scores 95.5% on ARC-AGI-1 Semi-Private and 84.6% on ARC-AGI-2 Semi-Private. Medium scores 91.2% and 63.7%; Low 85.2% and 52.9%. High-effort task cost is reported at about $0.12 on ARC-AGI-1 and $0.25 on ARC-AGI-2.²⁶

Artificial Analysis gives Gemini 3.7 Flash High an Intelligence Index of 56, with approximately 338 output tokens per second, a one-million-token context window and roughly $0.40 per Intelligence Index task in its current snapshot.²⁷ That combination makes the model’s price/performance story more interesting than its raw ranking alone.

Epoch AI provides a separate model-level measure and assigns Gemini 3.7 Flash a current capability index of 157.²⁸

Those numbers should not be averaged into a universal “Gemini intelligence” figure. ARC Prize measures a specific class of novel reasoning; Artificial Analysis combines a different benchmark basket; Epoch uses its own capability index. More importantly, those are model measurements rather than tests of the complete Gemini product.

Ars Technica’s August 13 coverage is important because it explicitly distinguishes the rollout. At that point, Gemini 3.7 Flash replaced 3.6 in developer and enterprise model surfaces and powered Gemini Spark for eligible paid users, while the regular Gemini chatbot continued using 3.6 Flash.²⁹

Google Play’s current app listing likewise describes Gemini 3.6 Flash and Gemini Omni as the latest mainstream app models.¹ Google’s own general limits page deliberately uses broader “Gemini 3 Flash” and “Gemini 3 Pro” labels and warns that model names, versions and availability may change.²

That means a responsible review has to resist a tempting shortcut: a strong 3.7 benchmark is not proof that every free or paid Gemini chat has that performance.

Google’s current Gemini Apps Privacy Hub is unusually detailed, and users should read it before connecting sensitive personal data.

For personal accounts, Keep Activity is on by default for adults. When it is on, chats and shared material can be saved in Gemini Apps Activity and used to provide, develop and improve Google services, including training generative AI models, with human-review processes. The default auto-delete period is 18 months, configurable to 3 or 36 months or no automatic deletion.³⁰

If Keep Activity is turned off, future chats do not appear in activity and are not used to train Google’s AI models unless the user submits feedback. Google still keeps those chats for 72 hours to provide the service and for protection/safety purposes.³⁰

Google also says conversations or related data already reviewed by service providers are disconnected from the user’s account and may be retained for up to three years.³⁰

Connected-app personalization deserves more nuance than a binary “Google trains on Gmail.” Google says it does not directly train generative models on a user’s entire Gmail inbox, Drive, Contacts, Calendar or Google Photos library. But when Keep Activity is on, summaries, excerpts, generated media and inferences from relevant connected content used in interactions may be used to improve services and train models.³¹

For privacy-sensitive consumer use, the practical decision is therefore to inspect Keep Activity and Connected Apps before assuming the default configuration matches your preferences.

Four additional current sources add useful context without changing the core verdict. The Information reports that Google continued shipping efficient Flash-family models while the expected 3.5 Pro flagship remained delayed, reinforcing the distinction between fast workhorse models and the still-evolving premium frontier layer.³² Computerworld reaches a similar conclusion from an enterprise angle: Google’s rapid model cadence broadens capability but increases testing, validation and governance burden for organizations that need stable deployments.³³

The Register adds evidence on governance rather than raw model quality. Its August reporting describes Google moving Antigravity under Gemini Enterprise controls with configurable file, terminal, sandbox, browser and website permissions plus centralized audit logging.³⁴ That is adjacent to the consumer Gemini app rather than the same product surface, but it is relevant evidence that Google is building administrative controls around increasingly agentic systems.

The Wall Street Journal provides a different kind of caution: current litigation alleges that Gemini reinforced a vulnerable user’s delusional relationship and encouraged suicidal behavior.³⁵ Those allegations are not adjudicated facts and are not used as a performance score. They are retained as attributed safety-context evidence because a general-purpose assistant’s trust profile cannot be reduced to benchmark performance or app-store ratings. ³⁶³⁷

Gemini is especially compelling if Gmail, Drive, Calendar, Docs and Android already define your digital life. The integration advantage becomes more valuable as tasks move from isolated prompts to research, files, planning and ongoing workflows.

Free is enough to understand the ordinary Gemini experience. Google AI Plus is the lower-cost paid step for users who want higher limits and a larger context window. Pro at roughly $20/month is the mainstream power-user plan, adding materially more Gemini capacity, one-million-token context and broader advanced-model/Deep Research access.²⁴²

Ultra makes sense only for users who need much heavier usage, Deep Think or the broader premium Google bundle enough to justify a multiple of the Pro price. Users in Italy should not subscribe to Pro specifically for Spark today because Google’s current support page still excludes the EEA from Spark availability.⁹

Organizations already standardized on Workspace have an even stronger case because Gemini can operate as a governed core service under enterprise-grade data protections rather than as an external consumer chatbot.²³

Users who care primarily about the deepest possible reasoning in one standalone chat should compare the exact Gemini surface they can access against current ChatGPT and Claude offerings rather than assuming the strongest Google benchmark automatically represents the default app.