WEBVTT 00:00:00.000 --> 00:00:03.980 You've probably typed something into Gemini, gotten an answer, and closed the tab. 00:00:04.500 --> 00:00:09.720 The same way you'd use any other chatbot. Here's the thing. That's maybe 10% of what it actually 00:00:09.720 --> 00:00:15.120 does. Use it right, and it can quietly take over almost half the busy work you're still doing by 00:00:15.120 --> 00:00:21.360 hand. I spent hours mapping every model, every mode, and every product Gemini is quietly wired 00:00:21.360 --> 00:00:27.480 into. And the number that stopped me was this. Gemini's app alone has over 650 million monthly 00:00:27.480 --> 00:00:33.640 users and that's before you count everyone using it inside search. Most of them are using maybe 20% 00:00:33.640 --> 00:00:38.040 of it with no idea the rest even exists. Look at this data from the Census Bureau. 00:00:39.000 --> 00:00:43.320 Only about one in five U.S. businesses actually use AI in their operations. 00:00:44.360 --> 00:00:48.600 So if you run a business and you're even thinking about this, you're ahead of most of your 00:00:48.600 --> 00:00:54.680 competition. What you might not know is that alongside covering AI news, we work with business 00:00:54.680 --> 00:00:59.320 owners to help them implement AI in their business. Our engineering team gets to know 00:00:59.320 --> 00:01:04.440 how your business runs, then builds the automation with you. You'll find the link in the description 00:01:04.440 --> 00:01:10.040 below. Click it, fill out a short form about your business, and we'll get in touch to set up a call. 00:01:10.040 --> 00:01:17.640 So in this video, I'm breaking down exactly what Gemini is as of mid-2026. Every current model, 00:01:17.640 --> 00:01:23.240 every mode, and everywhere Google has quietly built it in. By the end, you'll know exactly 00:01:23.240 --> 00:01:28.160 which Gemini tool to reach for depending on what you're actually trying to do, instead of just 00:01:28.160 --> 00:01:32.760 typing into whichever box is in front of you. First, let's clear up the biggest misconception. 00:01:33.380 --> 00:01:38.940 Gemini isn't one product at all. What Gemini actually is. Here's the mental model you need 00:01:38.940 --> 00:01:44.300 before any of this makes sense. Gemini isn't a single AI. It's Google's umbrella name for a 00:01:44.300 --> 00:01:49.400 whole platform, a family of models underneath, and a set of products on top that let you actually 00:01:49.400 --> 00:01:54.220 talk to them. Think of it in two layers. The bottom layer is the models themselves. Things 00:01:54.220 --> 00:02:01.180 like Gemini 3.6 Flash or Gemini 3.1 Pro. These are the engines, tuned for different jobs. Some 00:02:01.180 --> 00:02:06.640 built for speed, some for heavy reasoning, some for images or audio. You never see these names 00:02:06.640 --> 00:02:12.320 unless you go looking. The top layer is everything you actually click on. The Gemini app, AI mode 00:02:12.320 --> 00:02:18.400 inside Google search, Gemini inside Gmail and Docs, the voice assistant on your phone. All of those 00:02:18.400 --> 00:02:22.800 are just different doors into the same underlying models. That's the whole point of this video. 00:02:23.200 --> 00:02:28.300 Google isn't trying to build one great chatbot. It's trying to put the same AI brain behind every 00:02:28.300 --> 00:02:34.120 product you already use. So let's start with the brains, the actual models, because once you know 00:02:34.120 --> 00:02:38.880 what each one is built for, everything else clicks into place. The current model lineup. 00:02:39.260 --> 00:02:44.880 This is a demo checklist, so we're going model by model. What it is, what it's actually good for, 00:02:44.880 --> 00:02:52.020 and where you can get it. Gemini 3.7 Flash. Launched on August 13, 2026, this is Google's 00:02:52.020 --> 00:02:57.560 newest Flash model and its most capable workhorse yet. It's built primarily for coding and AI agents, 00:02:57.820 --> 00:03:03.220 with major improvements in software engineering, web development, and complex multi-step workflows. 00:03:03.940 --> 00:03:09.600 Google has already made it generally available through the Gemini API, positioning 3.7 Flash 00:03:09.600 --> 00:03:14.100 as the new go-to model when you want strong intelligence without giving up the speed and 00:03:14.100 --> 00:03:19.880 efficiency the Flash lineup is known for. Gemini 3.6 Flash. This is Google's current flagship, 00:03:20.420 --> 00:03:26.620 announced in a company blog post on July 21, 2026. It's built as a workhorse, strong at coding, 00:03:27.020 --> 00:03:30.920 knowledge work, and multimodal tasks. And according to Google's own numbers, 00:03:31.360 --> 00:03:35.960 it does the job using about 17% fewer tokens on average than its predecessor. 00:03:36.360 --> 00:03:41.140 Fewer tokens means faster answers and a lower bill if you're paying for it through the API. 00:03:41.140 --> 00:03:49.080 You can reach it through the Gemini API, through AI Studio, or simply by using the Gemini app and Search's AI mode. 00:03:49.440 --> 00:03:50.820 No extra setup required. 00:03:51.180 --> 00:03:57.800 If you only remember one model name from this video, make it this one, because it's what most of Gemini is quietly running on right now. 00:03:58.120 --> 00:03:59.800 Gemini 3.5 Flash. 00:04:00.120 --> 00:04:08.820 This one launched back in May 2026, and it's the model that was actually powering AI mode in Search before 3.6 Flash took over. 00:04:08.820 --> 00:04:12.940 Google described it as frontier-level intelligence at exceptional speed, 00:04:13.300 --> 00:04:18.100 and by its own benchmarks, it pushed output throughput to roughly four times faster than 00:04:18.100 --> 00:04:22.740 other top models at the time. It's still very much in active use across the Gemini app, 00:04:23.220 --> 00:04:29.220 Google's anti-gravity platform, and enterprise tools. It's slightly behind 3.6 now, but it's 00:04:29.220 --> 00:04:34.840 the model quietly sitting behind a huge share of what shipped this year, Gemini 3.5 flashlight. 00:04:34.840 --> 00:04:37.780 Same July announcement, different job entirely. 00:04:38.240 --> 00:04:44.020 This is the stripped-down, high-throughput sibling Google cites roughly 350 tokens per second, 00:04:44.380 --> 00:04:46.320 which is built for volume, not depth. 00:04:46.620 --> 00:04:48.540 You wouldn't use this for a hard reasoning problem. 00:04:48.860 --> 00:04:53.980 You'd use it for background tasks and agent pipelines that need to move fast and cheap at scale. 00:04:54.340 --> 00:04:55.740 Gemini 3.1 Pro 00:04:55.740 --> 00:04:59.800 Released in February 2026, this is the reasoning specialist. 00:04:59.960 --> 00:05:04.700 Google's benchmarks claim roughly double the logic performance of the earlier Gemini 3 Pro. 00:05:04.840 --> 00:05:13.600 It's mostly gated behind preview access through the API, Antigravity, Vertex AI, and Pro or Ultra subscriptions in the consumer app. 00:05:13.600 --> 00:05:18.480 If 3.6 Flash is built for speed, 3.1 Pro is built for depth. 00:05:18.480 --> 00:05:22.200 The model you'd want on a genuinely hard problem, not a quick one. 00:05:22.200 --> 00:05:27.880 And if you're actually building on top of these through the API, price is where the real-world decision gets made. 00:05:27.880 --> 00:05:34.440 According to Google's own published rates, 3.6 Flash runs about $1.50 per million input tokens 00:05:34.440 --> 00:05:37.040 and 750 per million output tokens. 00:05:37.040 --> 00:05:40.640 For context, that's noticeably cheaper on the output side 00:05:40.640 --> 00:05:44.840 than GPT 5.6 LUNA's roughly $6 per million. 00:05:44.840 --> 00:05:49.340 That gap is exactly why so many developers default to flash tier models 00:05:49.340 --> 00:05:51.040 for anything running at volume. 00:05:51.040 --> 00:05:54.640 Now, a handful of specialty models worth knowing by name, 00:05:54.640 --> 00:05:56.640 even if we don't dwell on each one. 00:05:56.640 --> 00:06:01.240 Nano Banana 2 is Gemini's current image generation and editing model, 00:06:01.240 --> 00:06:04.040 replacing the older Imogen line entirely, 00:06:04.040 --> 00:06:08.840 And that's not a small detail, because Amagen is actually shutting down on August 17, 00:06:09.160 --> 00:06:14.840 2026. If you've had workflows built on Amagen, that clock is already running. There's also a 00:06:14.840 --> 00:06:20.040 Nano Banana 2 Lite variant built purely for speed, trading a small amount of quality for 00:06:20.040 --> 00:06:25.720 much faster, cheaper output at high volume. VO 3.1 is Google's video generation model, 00:06:25.720 --> 00:06:30.360 still in beta, built to turn a text prompt into a short clip with matching audio. 00:06:30.360 --> 00:06:36.360 Gemini Audio 3.5 Live Translate handles real-time speech-to-speech translation 00:06:36.360 --> 00:06:40.760 across more than 70 languages already built into Google Meet and Android. 00:06:40.760 --> 00:06:43.560 And if you're curious about the more niche end of the lineup, 00:06:43.560 --> 00:06:47.960 there's a security-focused variant called 3.5 Flash Cyber Built 00:06:47.960 --> 00:06:50.840 to coordinate with vulnerability scanning tools, 00:06:50.840 --> 00:06:54.120 and Lyria 3.5, Google's music model, 00:06:54.120 --> 00:06:57.960 which can now generate tracks up to three minutes long from a text prompt. 00:06:57.960 --> 00:07:00.280 Here's the honest limitation worth naming. 00:07:00.360 --> 00:07:05.080 Google ships a lot of these models fast, and the naming gets confusing on purpose or not. 00:07:05.400 --> 00:07:10.680 3.5, 3.6, 3.1 Pro, Flashlight, FlashCyber. 00:07:11.080 --> 00:07:16.520 If you're not building on top of the API professionally, you genuinely don't need to memorize this list. 00:07:16.520 --> 00:07:21.540 You just need to know the shape of it. Fast and cheap, deep reasoning, and multimodal. 00:07:21.840 --> 00:07:24.880 That's really three categories wearing a lot of different name tags. 00:07:25.440 --> 00:07:28.800 The modes you actually interact with models are the engine. 00:07:28.800 --> 00:07:33.560 Modes are the steering wheel. Here's where things get useful for anyone who isn't a developer. 00:07:34.000 --> 00:07:39.860 AI mode inside Google search turns your search bar into a conversation. As of IO 2026, 00:07:40.420 --> 00:07:46.020 it's globally powered by Gemini 3.5 Flash, and instead of 10 blue links, you get a written 00:07:46.020 --> 00:07:50.900 answer with follow-up questions and sometimes an interactive widget built on the fly. Anyone 00:07:50.900 --> 00:07:56.000 with search can use it. No subscription required. Ask something like, what's a quick dinner with 00:07:56.000 --> 00:08:00.000 what's in my fridge and it answers in full sentences not a list of recipes 00:08:00.860 --> 00:08:03.800 DeepThink is the extra effort version of the Gemini app. 00:08:04.180 --> 00:08:08.640 It spends more compute per answer to reason through harder problems step by step. 00:08:08.960 --> 00:08:11.260 Google gates this one behind Google AI Ultra, 00:08:11.260 --> 00:08:16.620 and it's built for genuinely difficult science or engineering questions, not everyday chat. 00:08:16.980 --> 00:08:21.360 Now, deep research is where this stops being a chatbot and starts being an assistant. 00:08:21.800 --> 00:08:25.920 You give it a topic, and instead of one reply, it plans a research strategy, 00:08:26.480 --> 00:08:32.360 opens webpages, reads them, and, if you allow it, pulls from your own Gmail and Drive too. 00:08:32.680 --> 00:08:34.120 What comes back isn't a paragraph. 00:08:34.660 --> 00:08:37.740 It's a full multi-page report inside Gemini's Canvas. 00:08:37.740 --> 00:08:44.760 This is the part of Gemini that actually earns the word agent, and we're coming back to why that matters in a few minutes. 00:08:45.180 --> 00:08:47.400 Gemini Live is the voice and camera mode. 00:08:47.800 --> 00:08:54.940 Say, hey Google, let's chat, and you're talking to it hands-free, with the option to point your camera at something and ask what it's looking at. 00:08:55.240 --> 00:08:55.600 Live. 00:08:55.920 --> 00:08:57.600 And Canvas is the workspace mode. 00:08:58.000 --> 00:09:04.220 Type, create a quiz app about planets, and it writes the code, the interface, and the content in one pass. 00:09:04.700 --> 00:09:05.780 Right there for you to edit. 00:09:05.780 --> 00:09:09.120 One more worth a mention, briefly, because it's still early. 00:09:09.500 --> 00:09:13.600 Gemini Spark, a personal agent announced at IO 2026, 00:09:14.180 --> 00:09:17.500 meant to run continuously in the background handling things like scheduling. 00:09:17.980 --> 00:09:20.140 Right now it's limited to early ultra testers. 00:09:20.580 --> 00:09:22.740 So treat this one as coming, not here. 00:09:23.100 --> 00:09:24.700 Quick gut check before we move on. 00:09:25.040 --> 00:09:28.060 If all of that sounds like a lot of separate tools, that's fair. 00:09:28.300 --> 00:09:29.260 But notice the pattern. 00:09:29.680 --> 00:09:35.540 Every single one of these modes is just Gemini 3.5 or 3.6 Flash wearing a different job title. 00:09:35.780 --> 00:09:38.020 You're not learning six different AIs. 00:09:38.020 --> 00:09:41.300 You're learning six different ways to ask the same brain for help. 00:09:41.300 --> 00:09:42.980 Multimodal and practice. 00:09:42.980 --> 00:09:46.100 Let's talk about what multimodal actually means day to day, 00:09:46.100 --> 00:09:48.340 because it's more than a buzzword on a slide. 00:09:48.340 --> 00:09:50.660 Gemini reads and writes text and code. 00:09:50.660 --> 00:09:52.580 Obviously, that's the baseline. 00:09:52.580 --> 00:09:56.020 But drop a photo into a chat and ask it to caption or edit it, 00:09:56.020 --> 00:09:57.700 and Nano Banana handles that. 00:09:57.700 --> 00:09:59.700 Ask it to speak and answer out loud, 00:09:59.700 --> 00:10:03.060 and Gemini's audio models generate that voice on the spot 00:10:03.060 --> 00:10:05.700 with actual control over tone and pacing. 00:10:05.780 --> 00:10:08.780 Ask for a short video, and VO builds one from scratch. 00:10:09.160 --> 00:10:10.840 Ask it to edit an existing clip. 00:10:11.280 --> 00:10:15.740 Swap the sky, change the style, and that's a separate tool called Gemini Omni, 00:10:16.040 --> 00:10:18.280 doing frame-by-frame editing by voice command. 00:10:18.700 --> 00:10:24.220 Inside Google Docs and Sheets, the same underlying models can draft a document from your meeting notes 00:10:24.220 --> 00:10:29.280 or build a spreadsheet out of a pile of invoices, complete with formulas and charts, 00:10:29.500 --> 00:10:31.760 not just raw numbers dumped into cells. 00:10:31.760 --> 00:10:38.520 In Slides, hand it a list of bullet points and it can lay out an actual deck, not just text on blank slides. 00:10:38.940 --> 00:10:43.900 And through Gemini Live's camera mode, you can point your phone at a menu in a language you don't speak 00:10:43.900 --> 00:10:46.940 and get a live translation overlaid on what you're looking at, 00:10:47.140 --> 00:10:50.260 or ask it to identify an object it's looking at through the lens. 00:10:50.620 --> 00:10:51.600 No typing involved. 00:10:51.920 --> 00:10:53.200 Here's the part worth remembering. 00:10:53.640 --> 00:10:54.720 You never pick the model. 00:10:55.080 --> 00:10:56.180 You just say what you want. 00:10:56.480 --> 00:10:57.620 Make this an infographic. 00:10:58.140 --> 00:10:58.940 Translate this. 00:10:59.360 --> 00:11:00.240 Write this in Python. 00:11:00.240 --> 00:11:04.940 and Gemini quietly roots the request to whichever model actually does that job. 00:11:05.280 --> 00:11:07.300 That's the design philosophy in one sentence. 00:11:07.880 --> 00:11:11.160 One platform, and it decides the plumbing so you don't have to. 00:11:11.500 --> 00:11:12.920 Where Gemini actually lives. 00:11:13.280 --> 00:11:15.040 This is the part that's easy to underestimate. 00:11:15.680 --> 00:11:17.560 Gemini isn't confined to one app. 00:11:18.000 --> 00:11:20.240 It's spread across nearly everything Google ships. 00:11:20.700 --> 00:11:22.980 In search, it's AI mode, already covered. 00:11:23.460 --> 00:11:26.860 In Gmail, it's behind Smart Compose and Auto Reply suggestions. 00:11:27.320 --> 00:11:28.900 In Docs, Sheets and Slides. 00:11:28.900 --> 00:11:35.360 Ultra and Pro subscribers get Gemini drafting text, building formulas, and designing slide layouts, 00:11:35.880 --> 00:11:38.240 pulling context from your own files when you let it. 00:11:38.500 --> 00:11:41.520 In Drive, it can find and summarize documents for you. 00:11:41.920 --> 00:11:44.320 In Google Meet, it's doing live caption translation. 00:11:44.820 --> 00:11:49.160 On Android, especially Pixel devices, it's baked straight into the Voice Assistant, 00:11:49.540 --> 00:11:53.660 and there's a Chrome extension that lets the browser send page content straight to Gemini 00:11:53.660 --> 00:11:56.420 so you can ask questions about whatever tab you're on. 00:11:56.420 --> 00:12:01.680 And for developers, all of it is exposed through Google AI Studio and the Gemini API, 00:12:02.220 --> 00:12:07.040 plus a newer platform called Anti-Gravity for building multi-agent workflows on top of it. 00:12:07.280 --> 00:12:09.080 The strategic point here isn't subtle. 00:12:09.460 --> 00:12:12.320 Google isn't trying to win the best standalone chatbot argument. 00:12:12.700 --> 00:12:15.940 It's trying to make sure you're never more than one product away from Gemini, 00:12:16.300 --> 00:12:19.380 no matter what you're doing on a Google device or in a Google app. 00:12:19.780 --> 00:12:22.000 Agents, the part that actually matters. 00:12:22.380 --> 00:12:24.220 Now here's the shift I promised earlier. 00:12:24.220 --> 00:12:27.240 the one that actually changes what this platform is for. 00:12:27.640 --> 00:12:31.160 Everything so far has been ask a question, get an answer. 00:12:31.560 --> 00:12:34.480 Agents are Google trying to move Gemini past that entirely. 00:12:34.900 --> 00:12:40.620 Deep research is the clearest example already live, plan, browse, synthesize, write, 00:12:40.960 --> 00:12:42.580 without you babysitting every step. 00:12:42.980 --> 00:12:46.920 Spark is the early, still limited attempt at a persistent personal agent 00:12:46.920 --> 00:12:48.820 running continuously in the background. 00:12:49.280 --> 00:12:50.440 And on the developer side, 00:12:50.440 --> 00:12:55.920 anti-gravity lets companies build coordinated teams of sub-agents. Google's own blog post gave 00:12:55.920 --> 00:13:01.320 an example of businesses running parallel agents to analyze data at scale, rather than one model 00:13:01.320 --> 00:13:06.280 doing everything sequentially. Picture the difference in practice. The old way. You ask 00:13:06.280 --> 00:13:11.800 Gemini, what should I know before a trip to Japan? And it gives you a paragraph. The agent way. You 00:13:11.800 --> 00:13:17.940 say, plan my trip to Japan. And it checks flights, compares hotel options, and comes back with an 00:13:17.940 --> 00:13:23.040 actual itinerary, pausing to confirm with you before it books anything. That's the same underlying 00:13:23.040 --> 00:13:27.660 model, just given permission to take more than one step before handing control back to you. 00:13:28.000 --> 00:13:31.620 None of this is science fiction anymore, and none of it is fully finished either. 00:13:31.940 --> 00:13:37.140 That's the honest read. Deep research genuinely works today. Spark is still in early testing, 00:13:37.140 --> 00:13:42.080 but the direction is unmistakable. Google wants Gemini to eventually take a task, 00:13:42.500 --> 00:13:47.400 break it into steps, and execute most of them without you typing a follow-up for every single 00:13:47.400 --> 00:13:52.640 one. What actually makes Gemini different? So how does this stack up against everyone else building 00:13:52.640 --> 00:13:58.060 the same kind of thing? Let's be balanced here, because Google's advantages are real, but so are 00:13:58.060 --> 00:14:04.200 its weak spots. The clearest edge is data. Gemini can pull from live search results, maps, Gmail, 00:14:04.560 --> 00:14:10.580 and Drive in ways that a closed, sandbox chatbot simply can't match without plugins bolted on. 00:14:10.580 --> 00:14:16.380 The second edge is reach. Every Android phone is a potential Gemini client, and every workspace 00:14:16.380 --> 00:14:21.620 business account already has it available. No competitor has that kind of built-in distribution. 00:14:22.100 --> 00:14:28.120 And on raw benchmarks, Gemini 3 Pro topped the LM Arena leaderboard, which, regardless of how much 00:14:28.120 --> 00:14:32.680 weight you put on any single leaderboard, says Google's infrastructure and DeepMind's research 00:14:32.680 --> 00:14:38.520 are producing real, top-tier results, not just hype. But, and this matters for credibility, 00:14:39.060 --> 00:14:42.420 Google is genuinely more conservative about rollout than some competitors. 00:14:42.420 --> 00:14:47.420 DeepThink and Spark are still gated behind ultra subscriptions or limited testing, while 00:14:47.420 --> 00:14:50.580 some rivals ship new capabilities to everyone at once. 00:14:50.580 --> 00:14:57.660 And the tier structure itself — free, pro, ultra, API pricing — can be genuinely confusing 00:14:57.660 --> 00:15:00.640 next to a simpler flat subscription from a competitor. 00:15:00.640 --> 00:15:04.200 If you've ever opened the Gemini pricing page and closed it five minutes later still 00:15:04.200 --> 00:15:07.260 unsure which plan you need, that's not just you. 00:15:07.260 --> 00:15:08.620 Where this is actually headed. 00:15:08.620 --> 00:15:12.600 A few things are confirmed, and a few are still rumor, and it's worth keeping those separate. 00:15:12.960 --> 00:15:18.360 Confirmed. Gemini 3.5 Pro is currently in partner testing with a public release expected soon, 00:15:18.360 --> 00:15:23.860 and Google has already started training on Gemini 4, according to its own July 2026 announcement, 00:15:24.200 --> 00:15:28.560 though there's no public timeline for that yet. Workspace AI rollout continues expanding, 00:15:29.080 --> 00:15:34.300 and Gemini Live's regional language support keeps growing. Speculative, and worth labeling clearly 00:15:34.300 --> 00:15:40.000 as such. There's talk of a dedicated on-device AI chip for future Pixel phones, and some experimental 00:15:40.000 --> 00:15:45.100 deep-mind research around 3D avatars and world simulation that hasn't shipped as a product. 00:15:45.480 --> 00:15:50.500 Treat both of those as possible, not coming. Nothing official has confirmed either one. 00:15:50.820 --> 00:15:56.100 The verdict. So where does that leave things? Gemini in 2026 isn't a chatbot you occasionally 00:15:56.100 --> 00:16:02.560 open. It's an AI layer Google has threaded through search, Gmail, your documents, and increasingly, 00:16:02.560 --> 00:16:07.380 your phone itself. The models handle the thinking, the modes handle how you ask, 00:16:07.720 --> 00:16:11.640 and agents like Deep Research are the clearest sign of where all of it is actually heading. 00:16:12.020 --> 00:16:16.080 If there's one thing worth trying this week, it's Deep Research on something you'd normally 00:16:16.080 --> 00:16:20.100 spend an evening looking into yourself, and actually watching it work instead of just 00:16:20.100 --> 00:16:23.940 reading the final report. Drop a comment with which piece of this surprised you most, 00:16:24.300 --> 00:16:28.760 the model lineup, the agent side, or just how much of this you were already using without 00:16:28.760 --> 00:16:33.340 realizing it. I'll be back soon with a deeper breakdown on how deep research actually performs 00:16:33.340 --> 00:16:37.120 against a real research task. Thanks for watching, and I'll see you in the next one.