Gemini 3.6 Flash: Benchmarks, Pricing, and the 1M Context Breakthrough

[00:00:00]Google just dropped Gemini 3.6 Flash today and uh honestly if you are a developer looking purely at the coding leaderboards right now you have probably already dismissed it. Oh >> almost certainly >> because I mean it only scored 58.7% on SW Bench Pro. It is officially lagging behind Claude Sonnet 5 and it's trailing GPT 5.6 Luna. So on paper it really looks like Google just lost the pure coding war. But, and this is a massive bet, but if you talk to enterprise teams right now, like the engineers who are actually running those heavy chaotic agent workflows, they're absolutely thrilled today, right?

[00:00:36]>> Because Google did not build this model to win a pure coding leaderboard. They built it to score an 83% on OS World verified for actual literal computer use.

[00:00:45]>> This is wild.

[00:00:46]>> It is. And they finally made a 1 million token context window mathematically reliable. Plus, they engineered this uh 17% reduction in output token usage.

[00:00:54]They essentially sacrificed the ability to write code from scratch so they could build an AI that perfectly operates your entire computer.

[00:01:00]>> Exactly. And that is exactly what we are breaking down for you in today's deep dive. We are going straight through the data on this release.

[00:01:08]>> No fluff, >> right? No fluff. Just the benchmarks, the actual real world API pricing and whether you should migrate your backend workflows to 3.6 Flash today. So, let's start by looking at the architectural foundation of what Google is calling their new mid-tier workhorse.

[00:01:24]>> I love that term.

[00:01:25]>> Yeah, it fits perfectly. So, the key claim here is that Gemini 3.6 Flash is not a groundup rebuild, right? It's natively multimodal, meaning it handles text, image, video, audio, PDFs, and it's built squarely on the 3.5 flash architecture, but with this massive addition of built-in clientside computer use. And natively multimodal is really the operative phrase there because in older models, you know, if you fed an AI, a PDF with a chart, you often had this bolted on optical character recognition system running interference.

[00:01:55]>> Oh yeah, those OCR systems were a nightmare.

[00:01:56]>> Total nightmare. It would translate the visual chart into this messy text string and then feed that text to the AI, but 3.6 Flash processes the pixels in a text in the exact same latent space.

[00:02:08]>> Wow.

[00:02:09]>> Yeah. It handles a 1 million token input window and it caps out at a 64k output ceiling. And for the developers mapping out their runway, the input pricing sits at a $1.50 per 1 million tokens.

[00:02:23]>> Okay?

[00:02:23]>> And the output is 750 per 1 million. But honestly, the most important architectural shift for businesses paying for API tokens is that 17% reduction in output tokens compared to 3.5 flash. Okay, I actually have to push back on the 17% reduction right out of the gate because I'm looking at this from a strict engineering perspective, right? When I hear like uses 17% fewer output tokens, my immediate worry is about structural integrity.

[00:02:50]>> Ah, I see where you're going.

[00:02:51]>> Yeah. Like if the model is just being aggressively tur, don't you risk it failing to format complex JSON schemas properly? If you force a model to just cut 17% of its output, isn't it going to start skipping crucial reasoning steps?

[00:03:04]It's a totally valid concern because historically, if you told an AI to be brief, it just became stupid. But that is not what is happening under the hood here. The 17% reduction is not about omitting structural JSON tokens or skipping logical steps. It is about aggressively tuning out conversational filler during tool use.

[00:03:23]>> Wait, what do you mean by conversational filler?

[00:03:25]>> So, think about how this model is deployed in an enterprise agent loop. If you have an AI agent iterating through like a hundred different backend database queries, you do not want it giving you a polite preamble every single time it executes a SQL command.

[00:03:39]>> Oh, like when it says certainly I will now query the database.

[00:03:43]>> Exactly. Older models would say certainly I will now query the database for the user's information and then finally output the JSON.

[00:03:50]>> 3.6 flashes training forces leaner logic paths. It doesn't need to, you know, think out loud in plain English before generating the exact API payload. Okay, that makes so much sense. So it's less here's the JSON payload you requested and more just raw perfectly formatted brackets >> precisely leaner output means the agent executes the tool call instantly >> which actually perfectly explains why it performs so strangely on traditional benchmarks. Let's look at the data in the 3.6 flash versus cloud set 5 and GPT 5.6 Luna battle because this model essentially has a split personality.

[00:04:22]>> It really does. Yeah, it leases pure knowledge and pure generation tests, but it absolutely sweeps the board on reasoning, context, and UI tasks. I mean, let me run the pure coding numbers for you. SWEBench Pro, which measures resolving real GitHub issues from scratch, puts 3.6 Flash at 58.7% versus Claude Sonnet 5's 63.2%.

[00:04:44]And then on MLE bench, it hits 63.9% versus Claude 66.9%.

[00:04:49]Even the knowledge work metric GDP valet v2 ELO which is a mouthful that one also favors Sonnet 5 and Luna >> right but you have to remember what those benchmarks actually test they test a model's ability to act as a principal software engineer you know staring at a blank page and writing a complex application from thin air >> which is really hard >> it is but look at the flip side of the data when you test reasoning with tools on the charts given benchmark 3.6 flash hits 89.4%. And on long context recall, the GDM MRCR V2 benchmark at the 128K range, it reaches an incredible 91.8%.

[00:05:23]>> Okay, so it's basically the ultimate middle manager.

[00:05:26]>> Yes.

[00:05:26]>> And I actually mean that as a high compliment. Like it might not be able to write the complex Python script from zero, but it can perfectly read the 100page architectural manual about the script.

[00:05:36]>> Exactly.

[00:05:37]>> It organizes the dense financial charts and it directs the specialized tools to get the project done. That is a perfect mechanical analogy. If you are doing long context recall and tool orchestration, this model leads every single column. It's not trying to be your lead developer. It's trying to be the flawless project manager who knows exactly where every single piece of data is burdened >> and knows exactly which button to click to move that data around. And honestly, that middle manager skill set is exactly what you need for what Google is really going after here, which is dominating robotic process automation or RPA.

[00:06:07]>> Huge market, >> massive. Google clearly sacrificed the pure coding crown to win the computer use race. The hard data here is the 83.0% on OS World Verify, >> which is a huge deal.

[00:06:19]>> It really is. That barely edges out Sonet 5's 81.2%, but it completely leaves GPT 5.6 Luna in the dust at 72.7%.

[00:06:29]>> And Alice World Verified is such a fascinating benchmark because it measures operating a real computer interface. Uh, >> we're talking about the AI actually taking control of a mouse and keyboard, >> literally clicking things, >> literally clicking applications, navigating drop- down menus, filling out forms, and completing multi-step tasks across an operating system. And the reason this 83% score matters so much to developers is how Google actually shipped it, >> right? Because it's native.

[00:06:55]>> Yes, they shipped this capability as a built-in clientside tool natively within the Gemini API and Gemini Enterprise. It is not some experimental third-party add-on you have to hack together yourself using Python rappers.

[00:07:07]>> Let's paint a picture of why this is so valuable for you. Coding benchmarks just measure if you can write a clean sorting algorithm, but a SS world measures if you can automate the absolute worst, most mind-numbing parts of white collar work.

[00:07:19]>> Oh, absolutely.

[00:07:20]>> Think about a chaotic real world scenario. You have a legacy HR dashboard built in 2008 by a vendor that went out of business a decade ago.

[00:07:29]>> We've all seen them.

[00:07:29]>> We've all suffered through them. It has no API. You cannot send a clean JSON request to it. So, historically, if you wanted to automate data entry on that system, you had to write incredibly brittle web scraping scripts that would instantly break if the company changed the color of a single button.

[00:07:46]>> Exactly. Because older automation relied on reading the underlying HTML code. But 3.6 Flash operates visually. It looks at the screen coordinates.

[00:07:55]>> Wow. Okay.

[00:07:55]>> Yeah. It understands that the weirdly shaped blue rectangle in the corner is a submit button. Even if the underlying HTML is a total mess, it can move data between two ancient completely siloed systems by literally having the AI click copy in one window, switch tabs, and click paste in another.

[00:08:13]>> That is incredible.

[00:08:14]>> That is where the massive commercial value lies for Fortune 500 companies.

[00:08:18]It's that unglamorous heavy lifting work that usually requires an army of interns. But to navigate those massive legacy systems or to parse through mountains of company data to figure out what to copy and paste, you need a flawless memory. Which brings us to the actual context window data, >> right? The million tokens.

[00:08:37]>> Yeah. Because we have all seen the marketing mirages from AI companies over the last two years where they proudly announce, you know, a 1 million token context window. But if you actually put a million tokens of text in, the model completely forgets everything in the middle of the prompt. the infamous needle in a hay stack credibility problem.

[00:08:54]>> Yes, >> older models with massive context windows were essentially like a student speedreading a massive textbook. By the time they reached chapter 10, they had completely forgotten the core definitions they read in chapter 1, >> right?

[00:09:06]>> They could hold the text in memory theoretically, but they couldn't reliably recall specific data points buried deep within it.

[00:09:12]>> But the recall data for 3.6 flash is wild. Point-wise recall on GDM MRCRV2 roughly doubled to 54.0% at the full 1 million token range. And just for context, the older Gemini 3.1 Pro was sitting at a dismal 26.3% on that exact same test, >> which is practically guessing.

[00:09:33]>> Exactly. And if you zoom in to the 128K range, accuracy jumped from 77.3% to 91.8%.

[00:09:41]>> And to put that in practical terms, 91.8% An 8% recall fundamentally changes how you build applications. You can now realistically feed this model an entire enterprise code base or a massive stack of complex M&A legal contracts or a 100page SEC financial filing >> and actually trust it.

[00:09:58]>> Yes, actually trust it to find the needle. It rescues long context from his credibility problem.

[00:10:04]>> This model isn't speedreading anymore.

[00:10:05]It is operating like a perfectly indexed filing cabinet.

[00:10:08]>> That's a great way to put it. Before this release, asking an AI agent to find a specific liability clause buried on page 342 of a 500page document was a total gamble. You might get the right answer, or you might get a highly confident hallucination. Now, at nearly 92% accuracy in that 128K range, it shifts from being a neat party trick to a dependable feature for deep corporate due diligence.

[00:10:31]>> And it is not just indexing plain text.

[00:10:34]This is a massive generational leap in extracting insights from complex mixed media visuals. The vision upgrades are nuts.

[00:10:40]>> They really are. Let's look at the multimodal upgrades. The numbers back up how powerful this is. On Sheriff, which explicitly tests the model's ability to reason through complex charts while using external tools, it scored 89.4%.

[00:10:54]>> Yeah.

[00:10:54]>> And then on LV Bench, which tests long video understanding, it hit 83.2%.

[00:11:00]Just for perspective, Claude Sonnet 5 scored 68.5% on that exact same video benchmark.

[00:11:07]>> Wow. That is not a marginal win. That is a full tier of separation.

[00:11:11]>> It really is because think about the actual mechanics of what LVB forces a model to do. It is not just watching a video and giving a generic summary. To score an 83.2%, the model has to hold a visual memory of a specific fleeting frame from minute two of a video >> and seamlessly synthesize that visual data with a subtle audio cue that happens at minute 45.

[00:11:30]>> That is crazy.

[00:11:31]>> It requires absolute temporal and spatial awareness across thousands of frames.

[00:11:35]>> And we are already seeing this translate to the real world. I mean, if you look at the community tests flooding Reddit and X right now, developers are showing this model generating incredibly complex automotive line art.

[00:11:45]>> I saw those.

[00:11:46]>> Yeah. And parsing impossible engineering diagrams from just a single prompt. It is not just reading the text of a PDF.

[00:11:53]It is mathematically understanding the dense complex scatter plots embedded within that PDF.

[00:11:59]>> Right? If your workflow involves mixing unstructured text, dense financial charts, and long- form video, this visual processing leap is honestly the single biggest reason to switch your routing to 3.6 Flash.

[00:12:11]>> Absolutely. When you combine the 91.8% long context recall with that level of visual acuity, you have an AI that can ingest a company's entire quarterly earnings report, like the text, the charts, the recorded video, the CEO's presentation, >> everything.

[00:12:26]>> Everything that cross references all of it flawlessly. Okay, let's pivot and talk about what this actually costs when you run it at scale.

[00:12:32]>> Uh, the bottom line >> because cheapest per token on a pricing page does not necessarily mean cheapest per task on your monthly invoice. So, we established the sticker price for Gemini 3.6 flash is a $1.50 per million input tokens and 750 per million output tokens. Now, if you're a developer looking strictly at the pricing page, GP 5.6 Luna looks cheaper on paper. It's a dollar in and $6 out, >> right? But you have to do the math on the actual architectural usage of the tokens.

[00:13:02]>> Walk us through it.

[00:13:02]>> Let's break down a very real cost example for you. [snorts] Imagine you have a back-end agent workflow that burns exactly 2 million output tokens a month on the older Gemini 3.5 flash model. At the old rate of $9 per million, that costs you $18.

[00:13:18]>> Okay.

[00:13:19]>> But because 3.6 Six Flash was re-engineered to be 17% more token efficient. That exact same workload now only requires 1.66 million tokens to accomplish the same tasks. At the new 750 rate, your cost actually drops to $1245.

[00:13:33]>> Okay, I have to press you on this again because this feels like clean room lab math.

[00:13:37]>> Fair enough, >> right? In a perfectly sterile environment, yes, fewer tokens equals less money, but enterprise workflows are chaotic. Agents get confused, they hit API rate limits, they fail, and they retry constantly. Yeah, they do.

[00:13:50]>> So, do these fewer tool calls and greater token efficiencies actually translate to tangible invoice savings in the real world or does the agent just fail faster and burn through your budget anyway?

[00:14:01]>> It is a great point, but you have to look at how cascading failure loops actually work in agent architectures.

[00:14:08]Fewer reasoning steps and fewer tool calls intrinsically mean a massively reduced surface area for failure.

[00:14:14]>> Okay, I see.

[00:14:15]>> Right. In a chaotic environment, an agent that requires 10 sequential steps to complete a task has 10 different opportunities to fail, throw an error, and trigger a retry loop that burns tokens exponentially. Because 3.6 Flash is leaner and requires fewer reasoning steps to reach the correct UI action, it dramatically reduces those cascading failure loops.

[00:14:34]>> So it doesn't just fail faster.

[00:14:36]>> Exactly. It avoids the failure cascade entirely. Yeah.

[00:14:39]>> By taking a shorter path to the objective. So yes, when you aggregate that across thousands of automated tasks a day, the invoice savings are very real and very noticeable.

[00:14:49]>> That makes total sense. Now, while 3.6 Flash is saving money, there is a massive elephant in the room regarding Google's overall model lineup.

[00:14:59]>> Oh, I know what you're going to say.

[00:15:00]>> Where is Gemini 3.5 Pro?

[00:15:02]>> Yeah, >> people are extremely confused why 3.6 Flash is out today, but we never got the Pro version of the previous generation.

[00:15:09]The current industry rumor is that 3.5 Pro failed its internal coding benchmarks so spectacularly that Google just had to shulk it, >> which Google has not officially confirmed. Of course, >> of course not.

[00:15:20]>> Their official corporate stance is that 3.5 Pro is currently testing privately with select enterprise partners while the massive Gemini 4 model is in deep pre-training. But honestly, whether the rumor about the coding benchmarks is true or not, the reality on the ground is that 3.6 Flash is effectively functioning as a very capable stop gap right now. Yeah, it fills the void in the market while Google recalibrates their frontier level pro tier.

[00:15:43]>> Exactly.

[00:15:44]>> Which brings us to the other massive point of friction developers are experiencing today. The knowledge cutoff controversy.

[00:15:50]>> Oh, this one is driving me crazy.

[00:15:52]>> It's everywhere. If you look at the model card, it clearly states that the static training data ends in March 2026.

[00:15:58]But users all over developer forums are complaining that 3.6 6 Flash is aggressively failing at answering basic questions about current events from the first quarter of 2026. Did Google just mess up the date on their own documentation?

[00:16:12]>> This raises such an important point about the modern developer experience and how we actually interface with these APIs.

[00:16:18]>> It is almost certainly user error.

[00:16:20]>> Really?

[00:16:20]>> Yes, there is a fundamental difference between a model's static neural weights and its dynamic retrieval capabilities.

[00:16:26]Gemini 3.6 Flash inherently supports Google search grounding. If the model is failing on recent events from early 2026, it is not because the cutoff data is a lie. It is because the developer forgot to explicitly toggle grounding on in their API payload.

[00:16:41]>> Okay, let's explain the mechanics of that because it trips up so many people.

[00:16:44]When you send an API request, if you just send a text prompt, you are only querying the frozen mathematical weights of the model, >> right? To get current information, you have to pass a specific tool parameter in your code that essentially gives the model permission to open a browser and search Google in real time before it answers you.

[00:17:02]>> Precisely. If you do not explicitly pass that Google search grounding parameter in your JSON request, the model is trapped in its static training. You have to turn on the live web results if you wanted to know what happened yesterday or even what happened in February of 2026.

[00:17:17]>> Wow. It's a simple switch, but forgetting it is the root cause of like 90% of these hallucination complaints right now.

[00:17:24]>> Just flip the switch in your code, people.

[00:17:26]>> Seriously.

[00:17:27]>> Now, Google also quietly dropped two highly specialized sibling models alongside 3.6 Flash today. Flashlight and Flash Cyber. And they serve wildly different purposes.

[00:17:38]>> Very different. Flashlight is absolutely fascinating for teams dealing with massive bulk operations. It is designed purely for speed and cost. I mean, it runs at an absurd 350 output tokens per second.

[00:17:50]>> That is so fast.

[00:17:51]>> And it costs just 30 cents per million input tokens and $2.50 per million output tokens. Yeah.

[00:17:57]>> And remarkably, it actually beats the older Gemini 3 Flash on coding tasks, scoring 54.2% on SWB Pro.

[00:18:05]>> So, if you are like a logistics company processing millions of shipping manifests a day, and you just need high throughput data extraction where latency is everything, flashlight is your workhorse.

[00:18:15]>> Definitely. But then there is flash cyber. This model is built specifically for code mender which is Google's internal vulnerability detection system.

[00:18:23]The data shows it performs competitively at the absolute frontier on the cyber gym benchmark. But here is the catch.

[00:18:29]>> The big catch.

[00:18:29]>> It is strictly restricted to governments and highly trusted enterprise partners.

[00:18:33]So let me get this straight. Google essentially built a frontier level hacker model, realized it was far too good at finding zero-day vulnerabilities and live software and immediately locked it in a government vault.

[00:18:47]>> Essentially, yes. The dual use risk of an AI that is exceptionally good at finding and exploiting zero-day vulnerabilities and software is massive.

[00:18:55]Releasing a highly capable offensive cyber security model to the public API would be incredibly irresponsible.

[00:19:02]>> That makes sense. Restricting it to trusted partners for defensive patching is a very practical security measure, though keeping it locked away certainly adds to the mystique and the hype of today's release.

[00:19:13]>> It definitely builds the lore. All right, let's bring all of this data together. What is the final practical verdict for the developers and enterprise leaders trying to map out their architecture right now?

[00:19:23]>> The overarching takeaway from all of this data is that you have to route your tasks by strength. Do not use Gemini 3.6 six flash for pure iterative software engineering. If you need a model to architect an application from scratch, stick to Claude Sonnet 5 or Grock 4.5.

[00:19:40]But if you're doing agentic UI automation, massive document retrieval over a million tokens or chart heavy multimodal research, Gemini 3.6 Flash is mathematically the best tool for the job.

[00:19:51]>> It is the ultimate digital middle manager and it is going to completely change how we handle RPA. Which leaves us with one final thought for you to mle over. If models are moving away from chatting with humans and text boxes and moving towards silently clicking buttons and managing workflows in the background, wait, what is the very first mind-numbing daily task you are handing over to it? Are we looking at a near future where we stop talking to AI entirely and we just manage the outputs of invisible digital ghosts running our software for us? Let us know what you think.