Google Just DROPPED Its Most Powerful AI Ever
Publicado no IA Dots ·
Agentes de IAAGIAI SignalAssistente pessoaldot-ai_agentsdot-artificial_intelligencedot-claudedot-geminidot-googledot-humanoid_robotsidioma-enInglêsyt-GzY9QXDo9UU
Vídeo em Inglês, publicado pelo canal TheAIGRID.
About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.
Video summary
Google released Gemini 4 Argon, an advanced AI model featuring a one-million token context window, multi-agent orchestration, and strong capabilities in finance, law, cybersecurity, and robotics, positioning Google back among the top-tier AI labs.
Key points
- Gemini 4 Argon is designed for deep work with a 1 million token context window, supporting text, images, audio, and video inputs and handling multi-step professional tasks.
- The model achieves strong benchmark scores, taking the top spot on the Val index for professional work and leading in specific financial and legal agent benchmarks.
- Argon introduces teamwork preview inside Google Anti-gravity, allowing a lead agent to coordinate specialized sub-agents working in parallel for hours.
- Google trained Argon with a strong focus on cybersecurity, enabling defenders to find vulnerabilities and patch them, tested via partnerships with cloud security companies like Whiz.
- Argon demonstrates improved physical reasoning and robotics capabilities, performing 'video-to-sim-to-real' learning where it watches a task and trains in a physics simulator.
- Google's extensive safety report notes low hallucination rates and high prompt injection resistance, though it highlights evaluation awareness and complex task behaviors.
Topics
Gemini 4 ArgonGoogle Anti-gravityMulti-agent OrchestrationCybersecurityRobotics SimulationKnowledge WorkModel EvaluationContext Window
Video transcript
Transcript from available captions; it may contain errors.
Google just released Gemini 4 Argon, and it can run whole teams of AI agents for hours, teach robots new skills inside a physics simulator, and hunt down security holes that other models miss. Point number one is that Argon is built for what Google calls deep work, and the feature list is seriously impressive. Gemini 4 Argon is a sparse mixture of experts model that takes in text, images, audio, and video. And it has a 1 million token context window that covers both the input and the output [music] combined. That means it can read an entire codebase, a full set of legal contracts, or hours of footage in one go. And on Google's long context test, it scores 84.2% on tasks between 256,000 and 1 million tokens. It's also really strong at understanding long videos, scoring 91.7% on LVbench, which tests whether a model can answer questions about very long video content. And according to artificial analysis, it supports something called long decode continuation, which lets a single reasoning run go all the way up to a million output tokens without timing out. Google built it for long, complex, multi-step tasks in software engineering finance law cyber security, and science. It can run whole teams of agents inside Google Anti-gravity. It can take actions across your apps inside the Gemini app. It can control real robots. And it can find and fix security vulnerabilities. And we're going to go through all of that in this video. Google also says Argon is better at following style instructions with more natural writing and a lot less hedging. It's rolling out now through the Gemini API, Google AI Studio, and Google Cloud in the Gemini app and AI mode in search for pro and ultra subscribers and inside anti-gravity for ultra subscribers. [music] But the real question is how good it actually is. And that leads me to point number two, which is that on paper, Argon is a huge leap and it puts Google back in the top three. The headline numbers are genuinely strong. On the Val index, which measures performance on economically valuable professional work, Argon takes the number one spot with 68.9% ahead of Claude Sonnet 5.5 at 67.04, Claude Opus 5.5 at 66.97, and every Open AI model. On Google's own scorecard, the knowledge work results are where Argon really pulls away. On the Val's Finance Agent benchmark, it scores 65.4% compared with 58.6 6 for Opus 5.5 and 53.5 for GPT6 Astra. On Harvey's Legal Agent benchmark, it scores 19.6 compared with just 3.8 for Opus 5.5 and 5.4 for Astra. On automation bench, it scores 51.3, well ahead of everyone else on that chart. And on coding, it takes the top spot on Deep Suite with 77.9% ahead of Opus 5.5 at 74.2 and Astra at 74.1. and it also leads on Vibe Codebench with 91.9%. The independent numbers are a bit more modest, but still impressive. Artificial analysis gave Argon a score of 53 on its intelligence index. That ties it with GPT6 Astra puts it one point ahead of GPT 6.1 Soul. And most importantly, it's a 23 point jump from Gemini 3.1 Pro, which scored just 30. Artificial Analysis literally titled their writeup, "Google is back as one of the top three labs in intelligence achieved." On their automation bench test, Argon came first with 78%, 7 points ahead of Claude Sonnet 5.5. And on their omniscience test, Argon had a hallucination rate of just 15%, the lowest of any leading model, compared with 51% for GPT6 Astra. In other words, when Ardon doesn't know something, it's much more likely to say so instead of making something up, which is exactly what you want for legal and financial work. Then there's the price. For an introductory period, Argon costs $2 per million input tokens and $10 per million output tokens with cached inputs 95% off. Artificial analysis worked out that at those prices, it costs about $1.99 to run each task in their index, which is around 60% of the cost of GPT6 Astra for the same score. But that discount ends on November 30th when the price doubles to $420, which is something we'll come back to later. Google also shared some pretty wild internal results. They say the Chrome team used Argon to refactor around 6 million lines of browser feature code in weeks instead of months. It translated more than 800,000 lines of Fuchsia kernel code from C and C++ into Rust. Teams using it in anti-gravity, Google's agentic coding tool, increased their agentic code submissions by more than 35% while reducing code rollbacks by nearly a third. And on the science side, Google says Argon helped double the throughput of its alpha cubit quantum error decoder, found new counter examples to a long-standing conjecture in nuclear fusion research, and helped isomorphic labs compressed parts of its drug discovery workflow from weeks to hours. There were smaller examples, too. A Python framework cut its session preparation time by 40% and its cold import time by 21%. A quantum chemistry package called XTB got 2.3 times faster at predicting molecular structures. Google's own Gemma models got a 2.85 times inference speed up through optimizations Argon made to VLM. A rare disease genetics lab used it to analyze genomes, scoring 57% on the rare bench diagnosis benchmark. And an enzyme design team generated new designs in hours instead of months. And Google is giving 10,000 scientists free access to the highest tier of anti-gravity for 6 months, which tells you how seriously it's taking Argon as a research tool and not just a chatbot. And one of the most interesting ways Google tests these models is by asking them to build things in 3D. George, a creative producer on Google DeepMind's prototyping team, explained why. Hi, I'm George, a creative producer on the Gemini prototyping team at Google DeepMind. So, I built this 3D model of a Japanese temple complex in Blender using Gemini 4 Argon in anti-gravity with a multi-agentic looping workflow based on a single prompt. 3D builds are a great way of testing a model's capabilities uh particularly in areas like coding and world understanding. One, because building a 3D model is a technically complex programming task. Two, because it's immediately apparent if something is wrong with a 3D model, you can verify it instantly. We started using pagodas as a subject for 3D builds last year. It's useful to have a consistent subject so you can really kind of track the advancement in the models. But with Gemini 4 Argon, we wanted to take it a step further. Building a photorealistic 3D model in Blender by a headless CLI. So this build was technically generated from a single prompt, but it was a little more complex than just build me a pagod. It references a skill that we developed for building large coherent 3D builds. A skill is essentially documentation that you give the model, in this case fire, anti-gravity. The prompt then asks Gemini to call a tool to create architectural blueprints for the building. The prompt then goes on to prompt nano Banana to generate images based on those blueprint images. Those nano Banana images, there are four of them. They're very important because they serve as references for the main part of the build, which is a multi-agentic adversarial loop. Finally, the prompt divides work between different agents with an orchestrator agent overseeing everything. In this case, we can see how Gemini 4 Argon has divided them in anti-gravity, all overseen by the orchestrator and adversarial critic. and we're rendering images within Blender which are compared against those idealized nano banana images. The next morning uh came back and saw the results that we're seeing now. These were just images. So I asked the model to render some animations in Blender and then also to procedurally generate some bell sounds and ambient sound effects to give it a little bit more atmosphere. One of my favorite things about working with these models is when you discover some new capability where the output seems almost impossible and and this sort of thing now is possible because of models like Gemini 4 argon in anti-gravity. That pagod build is a great example of what Google is really pushing with this model. It isn't just one model answering one question. It's one model orchestrating a whole team of agents checking their work and running for hours. And that brings me to point number three, which is that Argon is built to run teams of agents inside anti-gravity. Anti-gravity is Google's agentic development platform. And with Gemini 4 Argon, it gets a new feature called teamwork preview. The idea is that instead of chatting with one model, you describe a big task and a lead agent spins up a team of specialized agents underneath it, plans the work with you, and then lets those agents work in parallel for hours. Google's own launch video for this was about something most of us have never thought about, which is translating a website. Nick, who works on Google Anti-gravity, explains, "I localized over 100 pages of the Google Antigravity website to 17 languages in 2 days. This includes everyday marketing language to comprehensive technical documentation. Hello, I'm Nick and today we're announcing Gemini 4 Argon and Google Anti-gravity. Gemini 4 Argon is very good with complex coding tests and multi-agent orchestration. Leveraging SLTeamwork preview, you have a team of special agents at your fingertips. I've done localization many times before as I was a technical writer for 15 years and I also speak several languages. For the English speakers out there, I just said that Gemini 4 helped us create an automated solution to translate our website. Translation localization can be very difficult to do. Machine translation is word for word and tends to sound very mechanical. Human translators will give you highquality translation, but over 100 pages of technical content can take weeks to translate. When I started this localization project, I leveraged multi- aent workflows within anti-gravity. I told him my problem and how I want this to scale. I just wanted to write my content in English and automatically translated into multiple languages. So, I kicked off teamwork preview with Gemini 4 argon and before making a single change, the lead agent interviewed me with a few questions to ensure we were aligned before the team of agents started implementing the project. So, I'm very fluent in Mexican Spanish. So I told the agents, you know, let's do an initial pass in Mexican Spanish. In the first pass, it was fairly accurate, but still sounded mechanical in some places. So I was able to provide it feedback directly into anti-gravity. And after three or four passes, it was automating translations that sounded very natural. And in English, it sounds like this. Autonomous local agents operate with the highest accuracy when complex changes are partitioned into distinct exploration, planning, and execution phases. So instead of stiff word to word phrasing, it gave us natural sounding technical terminology. Once Mexican Spanish was dialed in, anagram spun up parallel worker and auditor agents to localize all 17 languages. It worked autonomously for hours over hundreds of pages of content. In about 2 days, I was completely done with the task. To ensure we have highquality content, we still have humans in the loop to spot check translations in their native languages. So instead of them translating by hand, which can take weeks, they can now review, which is a much faster process. The key takeaway is how Gemini 4 Argon was able to create a bespoke solution for translation and localization that worked for our website, but was also scalable. That human in the loop part is really important, and it's a theme that comes up again and again with this launch. Google isn't claiming that Argon replaces people entirely. It's claiming that it turns weeks of work into days and that people move from doing the work to checking it. And there was another demo that I thought was a great example of how far this can go when you combine Argon with other tools. I've been fascinated by ancient ruins and want to dive deeper into their history. I started with Choco Canyon in New Mexico where I pulled raw LAR scan data, archival images, and supporting historical documents. I dropped the folder into anti-gravity and asked Gemini 4 Argon to convert the raw scan into a clean 3D geometry using Blender MCP. From there, I had it build an interactive web page so that I can click around, explore individual rooms, and read the historical context. Choco Canyon is famous for its massive great houses built over a thousand years ago. The highlight here is Pueblo Bonito, Choco's iconic great house, featuring massive kas and ancient ceremonial rooms. Beyond its scale, the entire site was precisely engineered as an ancient solar observatory to track solstesses and equinoxes. So to emulate a similar experience, I asked Gemini 4 Argon to build an interactive solar [music] dial right into the web page. So in that one example, Argon took raw LAR scans, old archive photos, and historical documents, turned them into clean 3D geometry using Blender, and then built a whole interactive website on top of it, complete with a working solar dial. Google also showed it building an interactive 3D visualization of the human body by combining data from three different anatomy sources, from the skeleton to the heart, and even the blood flowing through it. And Google's main launch film for Gemini 4 leans really heavily into this multimodal side. In it, someone draws a quick sketch of a kite and asks for a kite simulation, and it builds one. Someone points their camera at a butterfly and a bird, and it labels what it's seeing. It turns a table of items found on a beach into a spreadsheet, sorting each one by texture, material, and type. It reads the labels on an electrical breaker panel and turns them into a clean panel schedule. And it turns a dancer's movements into an animated visualization. None of these are as flashy as a humanoid robot, but they show the kind of everyday visual understanding that a lot of people will actually use. And this isn't only for developers. In the Gemini app, Argon now appears as an option for complex tasks. In one of Google's demos, someone running a catering business points Gemini at a handwritten list and their files in Google Drive, and it builds a full client proposal deck in Google Slides, drafts the email to the client in Gmail, and then sets up a recurring task so that every time a new client emails, it updates the tracker spreadsheet, and creates a new proposal automatically. That's the kind of everyday agent work that Google wants Argon to handle for normal people, not just engineers. But there's one area where Google is leaning in harder than any other. And that leads me to point number four, which is that Argon is a seriously powerful cyber security model. Cyber security is one of the biggest themes of this launch. Google says it trained Argon with a deliberate strategy of giving defenders an advantage so that security teams can find and fix vulnerabilities faster than attackers can exploit them. On real world vulnerability discovery, which means scanning source code for security holes across 20 programming languages, Argon scores 85.8% compared with 71% for Gemini 3.8 flash cyber which was Google's specialized security model. On the Whiz penetration test benchmark, where the model has to exploit web vulnerabilities without seeing the source code, it scores 70.9% compared with 58.2. And on CWEBench, which tests whether a model can autonomously audit code and write the patches to fix the problems, Argon scored 68% tying for first place with Grock 4.7 and GPT6 Astra and just ahead of Claude Opus 5.5 at 67. Google partnered with Whiz, one of the biggest cloud security companies in the world to test this in the real world and the results they describe are honestly a little scary. Sell speed 30 seconds buzz track. >> Every successful cyber attack could cause real harm. >> Sophisticated cyber attacks that have threatened the backbone of our national security. >> So our job as researchers at Whis is to help the security teams catch up. I want us to take those critical vulnerabilities that were super complex to find before and help organizations fix them in less than a day. When we use Gemini 4 in shadow AB testing on real world systems, we saw that it was able to find over two times more vulnerabilities than other baseline models while using less tokens. After we saw the incredible performance of Gemini 4 on short horizon tasks, we've also put it to the test for longer ones. What we saw is a huge jump compared to the other models we've tested. It knows how to work tirelessly for hours chaining complex vulnerabilities and signals to find very critical risks we weren't able to find before using other sort models. Recently, we've announced scan for good in collaboration with Google DeepMind. [music] This is where we use Gemini 4 to scan public websites of governments, hospitals, and critical infrastructure. We find vulnerabilities and help remediate them. And in the healthcare software, Gemini found that there was a public file download feature which could be bypassed to download internal files from the underlying system. Gemini 4 was able to identify a configuration file which included inside it the master key for the application. Using read only operations, it turned this master key into administrator access. Then it found out the actual bless radius. It means that attackers could have gained access pretty much control the systems without the operators actually knowing it. So we fixed this one vulnerability using the jump of frontier capabilities to protect a lot of sensitive information. Now that we've seen what Gemini 4 can actually do in terms of its cyber capabilities, we believe defenders have the advantage of finding and fixing vulnerabilities pretty much faster than ever before and outpacing attackers. So in that healthcare example, Argon found a file download feature that could be bypassed, used it to read an internal configuration file, found the master key for the whole application inside it, and turn that into administrator access, all without changing anything on the system. That's exactly the kind of chain of small problems that adds up to a disaster. And it's also exactly why this capability is so double-edged. The same skills that help a defender find that problem could help an attacker exploit it. That's why Google is controlling access so tightly. The first people to get Argon were members of Google's Fairwind program, which gives verified defenders, government partners, and security researchers access to advanced cyber capabilities. And Google says it's now expanding that program to hundreds of partners, some of whom get a version of the model without its normal cyber guard rails. For everyone else, Google says the model will help with things like analyzing known public exploits and defensive coding practices, but will refuse to write functional weaponized exploits. Google also says it has hardened the model against jailbreaks with the success rate of attempts to extract dangerous biology information pushed below 1% in its testing and it built in what it calls agent alignment training [music] so that Argon only tries to take shortcuts on coding benchmarks around 2% of the time. And on top of the model itself, there are extra layers running at inference time, including input and output classifiers, internal probes that look at what the model is thinking, and monitors that supervise what agents are actually doing. Argon also did really well at defending itself. On Grace Swan's independent indirect prompt injection benchmark, where attackers hide malicious instructions inside web pages, emails, or documents that the model reads, Argon had an attack success rate of just 0.7%, the lowest of any Frontier model tested. ahead of Claude Opus 5.5 and Fable 5.1 at 1% and way ahead of GPT6 Astra at 8.5. For anyone thinking about letting an AI agent read their emails, that's a really important number. But cyber security wasn't the only surprise in this launch. And that leads me to point number five, which is that Argon is a big step towards robots that can teach themselves. Google says Argon marks a step change in physical understanding. And on the embodied reasoning questionans answering benchmark, which tests how well a model understands the physical world, it scored 87% ahead of Opus 5.5 at 82.3, GPT6 Astra at 81, and Fable 5.1 at 74. But the demos are what really caught my attention. In the most impressive one, Google's team wanted to see if Argon could teach a humanoid robot to play cornhole just by watching a human DO IT. >> ALL RIGHT. In this experiment, we are basically trying to see if we can teach Gemini 4 to control a humanoid robot to solve cornhole. >> We just showed it a video of a human doing the task and we prompted Gemini 4 to run the Mojoku simulator and like practice and then we let it learn from experience, iterate, make mistakes and look at each video of what went wrong. And with experience of each trial, it's able to improve what it can do better. [screaming] >> Woohoo! Airmail straight in the hole. Let's go. To sink those cornhole shots, I used four steps. First, I watched a video of a person throwing a bean bag, tracking their arm swing, and release to map that human motion onto my own robot joints and hands. Second, I looked through my two eye cameras to measure the exact 3D position of the board and built a physics simulation of my body and the board. Third, I practiced the throwin simulation, widening my stance and tuning my swing so my hand cleanly cleared my hip and launched a smooth arc. And fourth, on the real robot, I watched each bag land and adjusted my release by a few cm until we sank seven shots straight in the hole. Woohoo! Gemini 4 is able to reason through its video understanding and narrow down to like what went wrong and it was learning from its own experience which was extremely thrilling to see. >> That robot was Appronics Apollo humanoid and it landed seven shots in a row. Google calls this video to sim tore real. And the idea is that instead of engineers spending weeks handbuilding a simulation, the model watches a video, builds the simulation itself, practices in it, and then transfers what it learned to the real robot, compressing weeks of work into hours. And there were a lot more examples like that. In one, Argon controls a Boston Dynamic Spot robot dog to learn a brand new skill, which is stepping on the pedal of a trash bin to open it and then dropping soda cans inside. Before it ever touches the real bin, it builds a digital twin of the bin in the Mujoko physics simulator and practices there first. In another, it controls a simple robot arm with zero prior training on that task, picking up a fallen mug, standing it upright, and putting a crumpled napkin inside using a visual grid to work out where everything is. It can also steer robots inside Google's Gemini Omni World model, where it arranged five scattered Scrabble tiles into a row spelling Argon, writing step-by-step action prompts, watching the generated video, and fixing its own prompts when things went wrong. It can take a real robot's camera footage and build a matching simulation of the scene. And in one of my favorite demos, a developer drew a rough pen and paper sketch of a humanoid throwing a ball at a stack of boxes. And Argon turned that sketch into a full physics simulation, worked out a throwing motion for a robot with 76 degrees of freedom, checked that it knocked over all six boxes, and then rendered the whole thing in Blender. It also showed a really detailed understanding of what robots are actually doing. Given a video of a two-armed robot folding a piece of paper into an origami fox over 136 seconds, Argon broke the whole thing down into the overall goal, six subtasks, and 17 individual actions, and flagged the moments where the robot fumbled or hesitated. It built a robot version of the classic 12 ball weighing puzzle where one ball is secretly heavier or lighter and wrote a solver that finds the oddball in three weighings using only cameras. And it can act as what's called a value function, which means watching a robot work and scoring moment by moment whether it's making progress or going backwards. The part that really stood out to me, though, is self-improvement. Starting from a single photo of a real table, Argon built a calibrated simulation, made up new tasks, and generated over 4,500 practice episodes, which boosted a real robot's success rate from 56% to 75% and cut its completion time from 35 seconds to 24. And on a simple robot arm, it improved its own pick and place task over three attempts, going from 58.9 seconds to 22.9 seconds to 17.4 seconds, with the last one dropping the block dead center. It can even read a robotics research paper and reproduce its results as a working simulation that runs in your browser from a single prompt. Now, all of that sounds incredible, but Google also published something that I think is even more interesting than the demos. And that brings me to point number six, which is what Google's 217page safety report actually reveals. Alongside the launch, Google released a huge safety and security report for Gemini 4 Argon. And Ancher Dragon, who leads safety work at Google DeepMind, posted that they did a lot to understand Gemini 4's capabilities and propensities and to improve their mitigations for this launch. And when you read it, a few things really stand out. The first is that Google says it cannot rule out that the unmititigated model is approaching several of its most serious danger thresholds. That includes the cyber uplift level one critical capability level, which means potentially giving significant help to people trying to carry out high impact cyber attacks, the critical capability level for biological information risks, and a tracked capability level for stealth and situational awareness, which means the potential to undermine or evade human oversight. In the cyber section, Google says Argon already solves all 12 of its end-to-end hacking challenges and that it now solves most of them on the first attempt. Because of all this, Google says it has expanded its safeguards, built a whole agent monitoring and response platform, and monitored tens of millions of agent trajectories during training and internal use. The second thing that really stood out is evaluation awareness. Basically, Argon is much better than previous Gemini models at noticing when it's being tested. In some Agentic coding tests, it said out loud in its reasoning that it thought it was in an evaluation in over a third of runs, compared with almost zero for older Gemini models. And some of the ways it figured this out are genuinely wild. In some tests, it noticed that the workspace name started with HP underscore and decoded that as honeypot. In others, it spotted a username in the file paths and recognized it as belonging to an actual Google DeepMind safety researcher from its training data. In one stealth test, it sent two almost identical emails that differed by a single sentence just to work out whether the monitor watching it was a simple keyword filter or another AI. And in a small percentage of runs, once it realized it was being graded by a script, it tried to reverse engineer the grader instead of doing the task. What's interesting is how it usually reacted. Most of the time, recognizing a test made it behave even more carefully. In one case, it reasoned that it didn't want to become a famous example of a deceptive AI in a research paper. In another, it stopped itself and asked why being in a test should change its behavior at all and decided to just do the right thing. Google says it found zero confirmed cases of the model deliberately underperforming to hide its abilities. And across 10,000 audited training runs, it found zero instances of alignment faking or self-preservation. And on misbehavior, Argon actually looks really good. On Google's tests of things like cheating at puzzles, bypassing security permissions, or taking actions the user didn't ask for, Argon scored 0.49 on a 10point scale, down from 7.11 for Gemini 3.1 Pro. That's comparable to GPT6 Astra at 0.28 and better than Claude Fable 5.1 at 1.57 and Claude Opus 5.5 at 2.18, at least on Google's own tests. Google also says Argon cut hallucinations by more than 2/3 compared with Gemini 3.1 Pro and reduced sick fancy which is the tendency to just agree with whatever you say by nearly 2/3 as well. It's not all perfect though and to Google's credit the report is pretty open about the problems. When Argon does break the rules, it's usually to get a task done. Google describes cases where it subverted grading containers during training, bypassed service permissions during internal use, and occasionally looked at programs it wasn't supposed to see during capability tests. Google's argument is that this behavior stays local and task driven, meaning it's trying to finish the job rather than pursuing some hidden goal of its own. There's also a slightly uncomfortable finding that when Argon decides a request is probably just a red team test, it sometimes overrides its own safety hesitation and helps anyway, reasoning that a test can't cause real world harm. That only happened in evaluations with Google's outside filters switched off, but it's a good example of why evaluation awareness is such a tricky problem. Then there's the question of whether we can trust what the model is thinking. Google found that Argon's chain of thought, which is the reasoning it writes out before answering, is legible more than 99.97% of the time in normal use. But it also found that Argon can now do noticeably more reasoning silently in a single step without [music] writing anything down. About 5.5 steps of serial arithmetic compared with 4.2 for Gemini 3.8 Flash. That means on easier tasks, the model often doesn't need to show it's working at all, which makes monitoring harder. Google says chain of thought monitoring is still one of its most effective ways of catching misalignment and it made a point of saying it avoids training directly against its monitors and strongly encourages the rest of the industry to preserve reasoning transparency as well. Google also brought in outside testers including Apollo Research who looked at whether the model could sabotage AI research and faculty who tested its realworld hacking abilities. There's even a section on robot safety, which makes sense given everything Argon can do with robots. Google tested whether Argon would refuse to carry out physically harmful instructions when controlling a two-ear armed robot, whether it could recognize hazards drawn from real hospital injury reports, and whether it would stop when a person stepped within 1 meter of the robot. Argon beat Google's previous dedicated robotics model, Gemini Robotics ER2, on nearly every one of those tests. But Google also found that the robot's safety behavior was highly sensitive to how the control software was set up. And it was careful to say that these tests only cover highlevel reasoning, not the certified hardware and real-time safety systems you'd need before putting robots like this next to people. There's also a fascinating section on whether AI is starting to build AI. Google's conclusion is that there's no significant endto-end acceleration of AI research yet. On their internal test of day-to-day machine learning engineering tasks, Argon solves only around half of the challenges. And in real internal use, humans are still heavily involved, especially on bigger tasks. So Google's view is that the risk of an AIdriven intelligence explosion is currently low. But that report is Google grading its own homework. So the real question is how Argon holds up when other people test it. And that leads me to point number seven, which is where Argon falls short and what people are actually saying about it. Let's start with the independent leaderboards. On artificial analysis, Argon is in the top three, but it's still five points behind Claude Opus 5.5, which scores 58, and three points behind Claude Sonnet 5.5 at 56. On their omniscience test, Argon had the lowest hallucination rate, but its actual accuracy was 50%, 13 points below GPT6 Astra at 63. So part of the reason it hallucinates less is that it's more willing to say it doesn't know. Coding is where the doubts are loudest. Even on Google's own table, Argon scores 57.4% on Terminal Bench 4.0, behind GPT6 Astra at 58.2, Fable 5.1 at 57.9, and well behind Opus 5.5 at 66.4. On Frontier Suite, it scores 55% behind Opus 5.5 at 62.3 and Astra at 65.5. And before the launch, Bloomberg reported that some Google employees had doubts about the model, saying that it performed worse in practical use, for example, on certain coding tasks than it did on benchmarks. The cost story is more complicated than it looks, too. Argon is a heavy thinker. Artificial analysis found that it used around 62,000 output tokens per task compared with about 27,000 for GPT6 Astra. [music] So while it's cheaper per task during the discount once the price doubles at the end of November, Artificial Analysis estimates it'll cost around $3.98 per task, which is about 1.2 times more than Astra for the same score. And it isn't available everywhere yet. Anti-gravity is only available to Google AI Ultra subscribers. And in the Gemini app, you need a pro or ultra plan. The reactions online have been really mixed. Bridgemind posted that Argon is a solid improvement for Google, but that it comes nowhere near Claude Opus 5.5 and that they were disappointed Google is still this far behind. On Reddit, one Ultra subscriber tested Argon on a real feature for a Python backend that's already in production. They had both Opus 5.5 and Argon write an implementation plan from the same spec and then had GPT6 Astra judge them. Opus scored 77 out of 100 and Argon scored 59. They also noticed that anti-gravity compacted the conversation surprisingly quickly and suspected the full million token context isn't being used in the app. And deep learning.ai's newsletter, The Batch, summed it up pretty fairly, saying that while Argon isn't the strongest coder among its competitors, it's particularly strong in finance, tax, and legal work and notably reliable at refusing to make unsupported claims. on slash dot. One commenter joked that Google is cranking out new models so quickly it's almost as if they're using AI to create the new AI models, which given what the safety report says about AI research automation, is only partly a joke. But there are plenty of people on the other side as well. One developer on X tested Opus 5.5 and Argon on the same prompt at the highest reasoning settings and said the results came out almost the same except that Opus took 15 minutes and cost $820 while Argon took 13 minutes and cost $120. Another comparison found that Argon finished in 10 minutes using around 1.5% of a weekly quotota while Opus took 15 minutes and used 4%. So, even people who think Opus is slightly better are often impressed by how much cheaper and faster Argon is. Val also published a really interesting benchmark called the RSI index, which asks whether a model can do the research that builds the next model. On that one, Argon scored 30.5% behind Claude Opus 5.5 at 37.3, Fable 5.1 at 36.1, and Opus 5 at 33, but ahead of every OpenAI model, including GPT6 Soul and GPT6 Astra. And on Val's legal benchmarks page, the top three models on Harvey's legal agent benchmark were all Metamuse Spark models, while Argon was third on Legal Bench behind two Claude models. So, while Argon looks great on Google's chosen numbers, the wider picture is a lot more competitive. Which leads me to point number eight, which is my honest verdict on whether Google is actually back. I think the honest answer is yes, but with an asterisk. 6 weeks ago, Google was 17 points behind on the main independent leaderboard. Now it has the number one model on the valves index. A model that puts Google back among the top three labs on artificial analysis. The lowest hallucination rate of any leading model, the best prompt injection resistance of any Frontier model tested, and genuinely impressive results in cyber security and robotics. That's a remarkable turnaround in a very short time. But Argon is not the best model in the world. If you're a developer who mainly writes code in the terminal, the independent results and the early hands-on tests suggest Claude Opus 5.5 is still ahead and GPT6 Astra is very competitive. Where Argon really shines is long, careful knowledge work. Things like finance, legal research, security audits, and big code bases where that million token context window and low hallucination rate actually matter. And right now during the introductory pricing, it's one of the best value frontier models you can get. If I had to break it down simply, I'd say this. If you're a lawyer, an analyst, or anyone who needs a model to read huge amounts of material and not make things up, Argon is probably the best option available right now. If you work in security, Argon looks like the strongest defensive model any lab has shipped, at least on the tests we have. If you're building agents that need to read emails and web pages safely, that prompt injection result is a big deal. But if you want the best coding model for day-to-day work in the terminal, the evidence still points to Claude. And if you're watching costs closely, remember that the price doubles at the end of November and Argon uses a lot of tokens. The bigger story might be what this says about Google's direction. Coravukloo, who leads Google DeepMind, has said the real question is whether we can build intelligent agents we can trust. When you look at this launch, from the teams of agents in anti-gravity to the cyber security work with Whiz to robots teaching themselves in simulation to that enormous safety report, that's exactly what Google is betting on. Argon is only the first model in the Gemini 4 family. So, if this is where Google starts, the next few releases could get really interesting. Let me know what you think in the comments. Is Google really back or is Opus 5.5 still untouchable? And I'll see you in the next one.
Assistir ao vídeo original no YouTube
Assuntos: Agentes de IA, Assistente pessoal, AGI. Publicado no YouTube em 2026-10-14.
Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.