Latest / AI Fire Daily / #74 Robin: Where is Gemini Pro? Testing Google's 3.6 Flash vs 3.5 Lite Architectures
Transcript
- 0:00Everyone was holding their breath for a massive,
- 0:02world -changing Gemini Pro release. Yeah. Beat,
- 0:06I mean, we always expect the paradigm shift.
- 0:08Oh, absolutely, the big flashy drop. Right. But
- 0:10instead, Google slipped a highly tactical, almost
- 0:16boring update under the door. And ironically,
- 0:20It might be the most important thing they've
- 0:22released this year. Yeah, it really is the ultimate
- 0:24pivot from hype to utility. I mean, we've spent
- 0:26the last few years obsessed with AI that can,
- 0:28you know, write a symphony. Or pass the bar exam.
- 0:30Exactly. But for actual software engineering,
- 0:33we just needed something that works reliably
- 0:35at scale. Well, welcome to the deep dive. Our
- 0:38mission today is unpacking Google's July 21st,
- 0:412026 release of the Gemini Flash series refresh.
- 0:45We're going to figure out exactly what this means
- 0:47for your daily work. So we have three specific
- 0:50models on the table today. There's Gemini 3 .6
- 0:52Flash. There's the Blazing Fast Gemini 3 .5 Flashlight.
- 0:56And finally, there's the heavily restricted Gemini
- 0:583 .5 Flash Cyber. The restricted one is fascinating.
- 1:01It really is. And the goal here is to help you
- 1:04figure out exactly which model fits your architecture
- 1:07and why this feels like a highly calculated strategic
- 1:10move before Gemini 4 eventually drops. So let's
- 1:13start with the flagship of this release. This
- 1:15is Gemini 3 .6 Flash. Yeah. Like we said, it's
- 1:18not the pro model everyone wanted. Yeah. But
- 1:21honestly, I think it's the most practical upgrade
- 1:23we've seen in a while. How so? Well, because
- 1:25it directly attacks the unit economics of AI.
- 1:28That is the perfect way to look at it. Gemini
- 1:303 .6 Flash is the workhorse. It sits right in
- 1:34the middle of Google's lineup. It balances coding
- 1:36quality, document understanding, and multimodal
- 1:39work. But the entire architecture is optimized
- 1:42to drive down the cost of complex agent tasks.
- 1:45And that focus on cost is vital for anyone actually
- 1:47building software right now. Yeah. Two secs silence.
- 1:50People forget how expensive AI agents get in
- 1:53production. Totally. An agent doesn't just answer
- 1:55your prompt once. It loops. Right. It iterates.
- 1:58Yeah. It checks files. It makes a plan. It writes
- 2:00code. And then it checks its own code. And if
- 2:03a single debugging task takes 20 background steps?
- 2:05Which happens all the time. Exactly. And you're
- 2:07paying top dollar for every single step. Your
- 2:10profit margins just vanish overnight. Yeah. You
- 2:12literally cannot build a sustainable sauce business
- 2:16if your backend AI is bleeding cash on every
- 2:19loop. Which brings us to the most interesting
- 2:22mechanism in the source material. Gemini 3 .6
- 2:26Flash uses 17 % fewer output tokens overall.
- 2:31compared to 3 .5 flash. Right. Now, I want to
- 2:34unpack how that actually works. Output tokens
- 2:37are the expensive ones, right? They are. Generating
- 2:40a token requires significantly more compute than
- 2:42just reading an input token. Makes sense. So
- 2:45by reducing the output token count by 17%, Google
- 2:48didn't just lower the price on a pricing sheet,
- 2:51they made the model structurally more concise.
- 2:54Like, it gets to the point faster. Exactly. And
- 2:56it gets even more extreme when you look at coding.
- 2:58On DeepSWE, the model uses up to 55 % fewer tokens.
- 3:02That's a huge drop. Let's define that real quick
- 3:04for everyone. Yeah. DeepSWE is a benchmark for
- 3:07comp - complex software engineering tasks. Right.
- 3:09And to give you an idea of what that benchmark
- 3:11entails, it's not just asking for a simple Python
- 3:13script. No. It's asking an AI to navigate a massive
- 3:17legacy code base, find a bug, and write the patch.
- 3:21Wow. Yeah. So 65 % fewer tokens on a task like
- 3:24that. It completely changes the math of running
- 3:27a coding agent. The pricing structure really
- 3:29reflects this push for high volume utility, too.
- 3:32We're looking at $1 .50 per $1 million input
- 3:36tokens and $7 .50 per $1 million output tokens.
- 3:42It's a massive shift. They basically want you
- 3:44to build autonomous systems that run continuously.
- 3:47I have to admit, I still wrestle with prompt
- 3:49drift myself. Oh, don't we all? You set up this
- 3:51beautiful complex prompt, and by the fourth or
- 3:54fifth iteration in the loop, the AI just loses
- 3:56the plot completely. It forgets what it's doing.
- 3:58Right. You have to scrap it and start over. Saving
- 4:01tokens on those inevitable frustrating retries
- 4:03is an absolute lifesaver. And beyond saving money,
- 4:06fewer tokens to generate physically means the
- 4:08model finishes its thought faster. Oh, right.
- 4:11So you get lower latency on top of the cost savings.
- 4:13There's also a major security component baked
- 4:15into this 3 .6 upgrade. The source notes it features
- 4:18new frontier safety safeguards. Yeah, this is
- 4:21critical. These specifically target CBRN. That
- 4:25stands for chemical, biological, radiological,
- 4:27and nuclear risks. Right. It also specifically
- 4:30targets cyber -offense misuse. They are actively
- 4:35trying to stop bad actors from using the model
- 4:38to automate network exploitation. The tricky
- 4:41part here is the balance to sex silence. We both
- 4:45know security safeguards usually make models
- 4:47incredibly annoying to use for legitimate work.
- 4:49Oh, definitely. You get those false positive
- 4:52refusals where the AI apologizes for not being
- 4:54able to help you write a basic bash script. Yeah,
- 4:57the dreaded I cannot fulfill this request. But
- 5:00Google's documentation claims this model resists
- 5:02actual jailbreaks without those constant annoying
- 5:05refusals. That's refreshing. It is. It's trained
- 5:08to understand the context of your legitimate
- 5:10developer environment. Much better. I want to
- 5:13circle back to that massive token reduction,
- 5:15though, because it raises a critical question.
- 5:17OK, shoot. If it uses 65 % fewer tokens for coding,
- 5:21is it actually thinking less or is it just communicating
- 5:23more efficiently? It's communicating more efficiently.
- 5:26The model cuts out the conversational fluff.
- 5:28OK. It stops writing. Certainly, I can help you
- 5:31with that code, and just outputs the raw diffs.
- 5:33Nice. And we know it isn't thinking less because
- 5:36its core reasoning scores actually went up. It
- 5:38hit 63 .9 % on MLEBench. Which is a solid jump
- 5:42in machine learning engineering tasks. Exactly.
- 5:45Got it. It edits itself better, saving us money
- 5:48on every single output. It's all signal, no noise,
- 5:52which is exactly what developers have been begging
- 5:54for. OK, so 3 .6 Flash is a fantastic, precise
- 5:58workhorse. But what if I'm running a customer
- 6:00support pipeline that gets 10 ,000 pings a minute?
- 6:03That's a different story. Yeah, even at $7 .50
- 6:06million output tokens, I would still go bankrupt.
- 6:09And that is exactly why Google quietly dropped
- 6:11a second, purely volume -driven model alongside
- 6:14it, Gemini 3 .5 Flashlight. This is the budget
- 6:18speedster. This model exists for one singular
- 6:21reason. Raw throughput. The numbers here are
- 6:24just wild. It hits a blistering 350 output tokens
- 6:27per second. To put 350 tokens a second into perspective,
- 6:30you are basically watching a short novel generate
- 6:32in under a minute. That is fast. It's so fast
- 6:35that, to the end user, the response feels instantaneous.
- 6:38And the pricing drops completely through the
- 6:41floor. It's dealer 30 cents per $1 million input
- 6:44tokens. Output tokens are just $2 .50. Two secs
- 6:49silence. Whoa! Imagine scaling to a billion queries
- 6:54without bankrupting your startup. That unlocks
- 6:56entirely new sauce business models that were
- 6:58financially impossible yesterday. Teams building
- 7:01internal search pipelines, scraping tools, or
- 7:04real -time gaming NPCs can finally run thousands
- 7:08of background tasks every hour without the latency
- 7:10and cost stacking up aggressively. But cheap
- 7:13doesn't automatically mean dumb. No, not at all.
- 7:15The data shows it actually beats the older Gemini
- 7:173 Flash in some surprising areas. Yeah, it scored
- 7:2054 .2 % on SWBench Pro, and it hit an impressive
- 7:2474 .0 % on Always World Verified. Always World
- 7:27Verified is a fascinating metric. That's a computer
- 7:29use benchmark. Right. It measures how well the
- 7:32AI can physically navigate a virtual desktop
- 7:34environment. And Flashlight actually comes with
- 7:36a built -in computer use tool. Yeah, it's designed
- 7:38out of the box to interact with websites and
- 7:41apps. It can click buttons, read DOM elements,
- 7:43and navigate browser environments. That's a staggering
- 7:46amount of capability for 30 cents a million input
- 7:49tokens. It is. But the most unique feature mentioned
- 7:51in the sources is its adjustable thinking levels.
- 7:54You can actually dial this cognitive effort up
- 7:56or down. It gives builders incredible granular
- 8:00control over their API bills and their latency.
- 8:03Wait, but if I lower the thinking level, aren't
- 8:05I just intentionally making the AI dumb beat.
- 8:10Why would anyone manually lower the thinking
- 8:12level if the model is already this cheap? Because
- 8:15in high volume, simple background tasks like
- 8:18extracting a date from a receipt, you don't need
- 8:21deep philosophical reasoning. True. When you
- 8:23lower the thinking level, you are literally telling
- 8:25the model to skip its internal loop of self -correction.
- 8:28I see. It bypasses the chain of thought scratch
- 8:31pad. It just grabs the data and runs. Makes sense.
- 8:34Lower thinking means faster reflexes for simple,
- 8:37repetitive background tasks. You don't need a
- 8:39chess grandmaster to sort your mail. You just
- 8:41need someone with fast hands. Reducing that internal
- 8:45processing drops the latency to almost nothing,
- 8:48making the system ultra responsive. Sponsor.
- 8:51So having explored the heavy lifter in 3 .6.
- 8:55And the speedster in flashlight, the real question
- 8:58is application. How do these actually fit together
- 9:01in a real world software architecture? The secret
- 9:05isn't picking one model over the other. That's
- 9:07a trap a lot of developers fall into. The real
- 9:10secret is using them together in a master worker
- 9:12setup. We're talking about agent workflows here.
- 9:15Exactly. Let's define that for clarity. Agent
- 9:17workflows are AI systems that run multiple steps
- 9:20automatically to solve problems. And building
- 9:22a resilient workflow is all about assigning the
- 9:24right brain to the right task. You want to match
- 9:27the cognitive overhead to the specific job. I
- 9:31like to think of this setup like stacking Lego
- 9:32blocks of data, beat. But to push that metaphor
- 9:35deeper, 3 .6 Flash is the master builder holding
- 9:38the instruction manual, while the flashlight
- 9:41models are the hands just snapping the repetitive
- 9:43parts together. That's a great way to visualize
- 9:45it. So 3 .6 Flash acts as the master agent. Right.
- 9:49The master agent's job is understanding the big
- 9:52picture goal. It takes the initial user prompt,
- 9:55breaks that massive task into smaller, manageable
- 9:57microjobs, and writes the structured JSON instructions
- 10:01for the other models to follow. And then? Then,
- 10:04crucially, it reviews the final decisions. Meanwhile,
- 10:08Gemini 3 .5 Flashlight. acts as the worker agent.
- 10:11It's the muscle doing the heavy lifting in parallel.
- 10:14Yeah, the worker agents take those JSON instructions
- 10:16and execute them. They generate the drafts, they
- 10:19run the quick web searches, they do the high
- 10:20-volume data extraction. Stacking the Lego blocks,
- 10:24you use the cheap, fast blocks for the foundation,
- 10:26and you save the complex, expensive master builder
- 10:29for the detailed quality control on top. You
- 10:32ensure you only pay for high reasoning when it's
- 10:35absolutely necessary. Let's trace a real -world
- 10:37example from the sources to see how the architecture
- 10:39architecture physically operates. Say you are
- 10:41building an internal research tool to scan corporate
- 10:44filings. Okay, so you give the system a massive
- 10:47300 -page financial report. Right. The master
- 10:49agent, 3 .6 Flash, reads your prompt. It realizes
- 10:53it needs to extract Q3 revenue numbers, check
- 10:57historical trends against last year, and summarize
- 11:00any new legal risks. But it doesn't do all that
- 11:02reading itself. It farms it out. Exactly. It
- 11:05sends three separate, highly specific prompts
- 11:08to three flashlight worker agents simultaneously.
- 11:11And because they are moving at 350 tokens per
- 11:14second, those workers hit the massive document,
- 11:17pull the data, and return it almost instantly.
- 11:19And they cost absolute pennies to run. Yeah.
- 11:21The workers hand their raw findings back up the
- 11:24chain to the master agent. The master agent then
- 11:26synthesizes the final report using its higher
- 11:28reasoning to ensure the narrative actually makes
- 11:30sense and answers your original prompt. But there's
- 11:32an There's an obvious risk here, right? What's
- 11:34that? If the cheaper worker agents hallucinate
- 11:36on a task, can the master agent reliably catch
- 11:39those mistakes? The data suggests it can, provided
- 11:42your system architecture is solid. Gemini 3 .6
- 11:46Flash has the advanced reasoning bandwidth to
- 11:48act as an evaluator. It can cross -reference
- 11:50the worker's output against the original text
- 11:53to filter out the weak or incorrect data. So
- 11:56the smarter model acts as a strict... quality
- 11:59control filter for the workers. It is a hierarchical
- 12:02management structure built entirely out of AI
- 12:05models communicating with each other. Which is
- 12:07incredibly efficient, to sex silence. So we can
- 12:11use those first two models to build our Lego
- 12:13castles. But there is a third model Google released
- 12:16in this update. Oh, yes. And this one is strictly
- 12:19locked behind closed doors. This is the restricted
- 12:21enigma of the July release. Gemini 3 .5 Flash
- 12:26Cyber. You cannot find this in the Consumer Gemini
- 12:28app. You cannot just call it up on an API for
- 12:30your personal coding project. No. It runs exclusively
- 12:33inside Codemender, which is Google's internal
- 12:36AI code security agent. A source has explained
- 12:38this model is fine -tuned specifically to discover
- 12:41and repair security vulnerabilities. It's a highly
- 12:44specialized digital detective. Multiple Flash
- 12:47Cyber agents can work together in that same master
- 12:50worker architecture we just talked about. They
- 12:52inspect massive code bases, tracing data flows
- 12:54in ways human engineers simply can't keep in
- 12:57their working memory, hunting for structural
- 12:59weak points. But the access is severely locked
- 13:02down. It is restricted only to governments and
- 13:05trusted partners in a closed pilot program. And
- 13:08we really have to talk about the dual -use nature
- 13:10of this technology. Dual use. Beat. That's a
- 13:14terrifyingly clinical term for what we're actually
- 13:16discussing. It is. The fundamental reality of
- 13:18cybersecurity is that if an AI is incredibly
- 13:21good at finding and fixing vulnerabilities in
- 13:24code, it is mathematically just as good at exploiting
- 13:26them in the wrong hands. Exactly. A tool that
- 13:29tells you exactly how to lock a complex vault
- 13:31door can also show someone exactly how to pick
- 13:34it. If a malicious actor got raw API access to
- 13:38Flash Cyber, they wouldn't use it to secure their
- 13:40own code. They would point it at open source
- 13:43repositories and automate the discovery of zero
- 13:45-day vulnerabilities at massive scale. They could
- 13:48find the cracks in the global software supply
- 13:50chain before the maintainers even know they exist.
- 13:54Two secs silence. But I have to push back on
- 13:57Google's logic here. Does locking this model
- 14:00away actually keep us safer? Or does it just
- 14:03slow down independent security defenders? That
- 14:06is the core debate in open source AI right now.
- 14:09Keeping it restricted definitely frustrates independent
- 14:12defenders. But Google's calculation is that releasing
- 14:15it openly would arm malicious actors overnight.
- 14:19A rogue hacker could use it to automate massive
- 14:21zero -day cyber attacks before the defense community
- 14:24had time to patch the holes. Right. Keeping it
- 14:26restricted stops hackers from automating massive
- 14:28zero -day cyber attacks. Google is taking the
- 14:31cautious route. By deploying it only inside controlled
- 14:34environments like code mender, the outputs and
- 14:37the targets can be strictly monitored. It's a
- 14:40fascinating, slightly chilling glimpse into the
- 14:42future of cyber warfare. The battles of the next
- 14:45decade won't be fought by humans typing on keyboards
- 14:48in a dark room. No, they won't. They'll be fought
- 14:50by swarms of specialized security models attacking
- 14:53and defending code bases at machine speed. Which
- 14:56is exactly why the restriction makes sense for
- 14:58now. The potential blast radius of a rogue security
- 15:02AI is just too large for public release. Let's
- 15:05slowly pull these threads together. Beat. We've
- 15:08looked at all three parts of this July 2026 release.
- 15:12And a very clear picture emerges about Google's
- 15:15strategy. This update is not about flashy generalized
- 15:19intelligence. No. It's about pure practical utility
- 15:22and unit economics. Google is handing developers
- 15:25highly efficient specialized tools. You get a
- 15:28brilliant master builder and 3 .6 flash. You
- 15:31get a blazing fast cheap worker and 3 .5 flashlight.
- 15:35You can finally use them to run automated systems
- 15:37affordably. And they are clearly doing this to
- 15:39buy time. They're giving builders the precise
- 15:41tools they need to optimize their current architectures,
- 15:44while Google finishes cooking the eventual massive
- 15:46release of Gemini 3 .5 Pro and Gemini 4. It's
- 15:49a bridge. But it's a structurally vital, very
- 15:52useful bridge for the industry. My recommendation
- 15:54to anyone listening today is to deeply audit
- 15:57your current AI workflows. Look at where you're
- 16:00spending money on API calls. You are almost certainly
- 16:03overpaying for reasoning on simple tasks. You
- 16:06should absolutely try building a master worker
- 16:08setup with these two new models. The efficiency
- 16:11gains are just too massive to ignore. Two sec
- 16:13silence. It makes total sense for the bottom
- 16:16line, but it does leave me thinking about the
- 16:19underlying structure we are building. How so?
- 16:21If our top -tier AI models are officially transitioning
- 16:23into middle managers whose main job is to delegate
- 16:26tasks and oversee swarms of cheaper, faster AI
- 16:29workers, what happens when those worker models
- 16:31become perfectly capable of managing themselves?
- 16:34Where exactly will you and I sit in a fully automated
- 16:37AI hierarchy? Out Tiro music.