I was looking at my workflow logs this morning and it hit me that I haven't actually used a dedicated chatbot interface in nearly three weeks. Everything is happening in the background now, through the various A P I hooks and agentic layers we have set up. It feels like the era of talking to a box on a screen is finally, officially over.
It really is. We have moved into what people are calling the Multi-Surface Operating Layer. It is less about a product you visit and more about the substrate your entire professional life sits on. Herman Poppleberry here, by the way, and I have been knee-deep in the technical documentation for the new releases from the big three all week. We are talking about a fundamental shift in how we conceptualize computing. It is no longer about the "chat"; it is about the "agentic substrate."
Today's prompt from Daniel is about exactly that shift. He wants us to look past the marketing vibes and the multimodal flashy stuff to see what actually differentiates G P T five point four, Gemini three point one, and Claude four point six under the hood. Beyond just being really good at coding, how do these models actually vary in their architectural D N A and their instructional personalities? Because as of late March twenty twenty-six, the choice you make determines the very logic of your automated workflows.
It is a great time to ask because the convergence we saw last year has started to fracture into very distinct design philosophies. While everyone moved to Mixture-of-Experts architectures, the way they route those experts and the way they prioritize reasoning versus speed has diverged significantly. We are seeing a real split between the "Thinking" models and the "Instant" models, and that has massive implications for how you build your A I stack.
Let's start with the current landscape as of late March twenty twenty-six. We have OpenAI with their new Thinking and Pro models, Google DeepMind just dropped Gemini three point one Pro and Flash, and Anthropic came out with Claude four point six Opus and Sonnet earlier this month. On the surface, they all seem like they can do everything, but the benchmarks tell a more nuanced story. We are seeing OpenAI consolidate their o-series reasoning with the Codex line, while Google is pushing recursive memory and Anthropic is giving us granular effort controls.
They really do. If you look at something like the S W E bench Verified, which measures how these models handle real-world GitHub issues, Claude four point six Opus is currently leading the pack at eighty point eight percent. Gemini three point one Pro is right on its heels at eighty point six percent, while G P T five point four is trailing a bit at seventy-seven point two percent. That is a massive shift from two years ago when OpenAI was the undisputed king of coding. It shows that the competition has not just caught up; they have specialized.
That seventy-seven point two percent for G P T five point four is interesting because OpenAI has been leaning so hard into the reasoning side of things with their o-series consolidation. You would think that more thinking time would lead to higher scores on a benchmark like S W E bench. Sam Altman has been very vocal about prioritizing "Novel Science" and agentic execution over just having a model that sounds like a person.
You would think so, but the issue often comes down to the architecture of the Mixture-of-Experts routing. Claude four point six Opus is rumored to be a seven hundred forty-four billion parameter model, but it only uses about forty billion active parameters per token. Compare that to the open-weights baseline of Llama four Maverick, which uses seventeen billion active parameters out of four hundred billion total. Anthropic seems to have found a way to make their forty billion active parameters more specialized for high-stakes instruction following. This is what we call "sparse activation" efficiency.
When you say specialized, are you talking about how the model actually selects which experts to engage for a specific prompt? Because that seems to be the "secret sauce" everyone is fighting over right now.
That is exactly it. In a Mixture-of-Experts system, you have a router that decides which sub-networks or experts are best suited for the incoming data. Anthropic has always had this philosophy of careful instruction following. Their routing seems optimized to minimize what we call the "drift of intent." When you give Claude a complex refactoring task across a massive repository, it stays anchored to your original constraints better than G P T five point four, which sometimes tries to be a bit too creative or conversational. It is the difference between a model that wants to please you and a model that wants to be correct.
I have noticed that. G P T five point four feels like it is trying to be a collaborator, whereas Claude feels like a very high-level architect who follows the blueprint to the letter. But we have to talk about the Thinking latency. OpenAI's move to bake the reasoning process into a ten to thirty second pause before you get an answer has been a huge point of contention this month. Some people love the depth, but others find it unusable for real-time tasks.
It is a fundamental trade-off. OpenAI is prioritizing auditability and deep logic. By forcing the model to generate a hidden chain of thought before the final output, they are essentially giving it a workspace to check its own work. This is why G P T five point four is still very strong for brainstorming or when you need a model to navigate a G U I through their Operator agent. It needs that pause to simulate the outcomes of its actions. It is literally running internal simulations of what happens if it clicks a certain button in your browser.
But for a real-time agentic workflow, thirty seconds is an eternity. If I have an agent trying to manage a live server migration, I cannot have it sitting there thinking for half a minute while the site is down. This is where Gemini three point one Pro seems to be finding its niche. Google seems to be moving away from that long pause while still hitting incredible reasoning numbers.
Google is playing a different game entirely. Gemini three point one Pro is the value leader right now at two dollars per one million input tokens. For comparison, Claude four point six Opus is at five dollars. Google is leveraging their massive internal infrastructure to offer high-level reasoning with almost zero latency. They are also winning on pure logical reasoning. Gemini three point one Pro just hit seventy-seven point one percent on the A R C A G I two benchmark. That is a record, Corn.
For listeners who might not follow the leaderboard wars as closely, the A R C A G I benchmark is basically the gold standard for testing whether a model actually understands logic or if it is just reciting things it saw in its training data. It is designed to be impossible to memorize. The fact that Google is leading there suggests they have cracked something about fluid intelligence that the others are still struggling with.
It really is a test of fluid intelligence. The fact that Gemini is leading there, and also dominating the G P Q A Diamond benchmark for graduate-level science at ninety-four point three percent, tells us that Google DeepMind has successfully integrated their AlphaZero-style search capabilities into their language models. They are moving toward universal multimodal perception. This is what Demis Hassabis has been calling Project Astra—the idea that the model isn't just processing text, but is "seeing" the logic of the world.
I want to go back to the hardware side of this for a second because it feels like a turning point. We saw reports this month that OpenAI has started deploying A M D M I four fifty clusters. That is a huge deal given how much Nvidia has dominated the space with C U D A. We talked about the competitive pressure in episode fourteen seventy-one, but this feels like a physical manifestation of that pressure.
It is a strategic move to break the C U D A lock-in. Training these trillion-parameter Mixture-of-Experts models is becoming prohibitively expensive in terms of energy and hardware margins. By validating the R O C m software stack on A M D hardware, OpenAI is trying to lower the floor for inference costs. If they can get G P T five point four running efficiently on A M D clusters, that two dollar and fifty cent price point might drop even further. It is a hedge against Nvidia's supply chain volatility.
It also gives them leverage in negotiations. But let's look at the instructional characteristics Daniel asked about. If you are a developer or a researcher, why would you choose one over the other in twenty twenty-six? If I am writing a massive legal brief, I am probably leaning toward Claude four point six, right?
Almost certainly. Anthropic introduced these Effort Controls in the four point six release. You can actually toggle between Low, Medium, and Max reasoning depth. If you are doing long-form legal analysis where you need to maintain consistency across a one-million-token context window, setting Claude to Max Effort ensures it explores every edge case. It has the lowest hallucination rate in massive contexts because its brand voice is inherently more cautious. Dario Amodei has kept that "Safety-First" posture, which is why they are currently negotiating those high-stakes deployment terms with the Pentagon. They want to be the "trusted" model.
And Google's play for that same researcher would be the context window itself. They are previewing this recursive memory architecture that could theoretically handle one hundred million tokens. That is not just a document; that is an entire library of specialized knowledge. We touched on the shift toward multimodal perception in episode fourteen eighty-two, and this recursive memory feels like the logical conclusion of that.
The recursive memory shift is fascinating. Instead of just having a static context window where the model forgets the beginning once it reaches the end, Gemini three point one is starting to use a form of dynamic compression. It summarizes and embeds its own previous thoughts into a persistent memory layer. This could eventually make traditional Retrieval-Augmented Generation or R A G obsolete for many use cases. If the model can just remember the entire history of a project natively, why bother with a separate vector database? It changes the entire architecture of how we build A I applications.
We talked about that shift in episode fourteen eighty-two, how the vector gold rush was moving toward this kind of multimodal optimization. But there is a second-order effect here with the training data. The Ziff Davis lawsuit against OpenAI is a big cloud hanging over the industry right now. It is the elephant in the room when we talk about "reasoning" capabilities.
The allegation is that OpenAI used proprietary reporting from outlets like C N E T to train the reasoning modules of G P T five. The argument is that they didn't just use the text for facts, but used the structured reporting to teach the model how to "think" and "reason" through complex topics. If the courts decide that using high-quality journalism to teach a model how to reason is not fair use, it changes the economics of training data overnight. It forces these companies to move away from the open web and toward expensive, licensed datasets.
Which might explain why Anthropic is being so careful with their safety-first posture. They are trying to position themselves as the most legally compliant option. But OpenAI is doubling down on the agentic execution side. With their Operator G U I agent, they are basically saying, we do not care if we are slightly behind on a specific coding benchmark because our model can actually use your computer for you. It can open a browser, log into your accounting software, and reconcile your invoices while you sleep.
That is the Multi-Surface Operating Layer in action. It is not about the chat box; it is about the model having agency over your digital environment. But I do wonder about the personality of these models. Gemini still feels very much like a Google product. It is deeply integrated into Workspace. If I am in a Google Doc, Gemini three point one Pro feels like it knows exactly what my Gmail and my Calendar look like. It is what they call "ambient retrieval."
It is a different kind of intelligence than the pure text-based reasoning we saw a few years ago. It is native multimodality. Gemini was built to process video, audio, and text simultaneously because it was trained that way from the ground up. If you feed it a video of a technical lecture, it isn't just transcribing the audio; it is watching the slides, noting the speaker's gestures, and correlating that with the tone of voice. That is a level of perception that G P T five point four still struggles with, as it often relies on separate vision and audio encoders that are stitched together.
And that brings us to the practical side. If someone is listening to this and trying to decide how to allocate their A I budget for their team, what is the breakdown? Because at five dollars per million tokens for Claude Opus, you have to be sure you need that extra precision.
If you are doing high-throughput, low-latency tasks—like customer support bots or real-time data processing—Gemini three point one Flash is the winner. It is incredibly cheap and fast. If you are doing high-stakes logic, like auditing smart contracts or complex repository-scale refactoring, you pay the premium for Claude four point six Opus. The five dollar per million token cost is worth it when a single error could cost you millions.
And G P T five point four is the versatile all-rounder. It is the best for rapid prototyping and brainstorming because it has that conversational flexibility that Claude lacks. Claude can sometimes feel a bit stiff or overly formal. G P T five point four still has that spark of creativity that makes it feel like a better partner for the early stages of a project. Plus, if you need it to actually execute tasks across different apps, the Operator integration is currently superior to what Google or Anthropic have in the wild. It is the "Swiss Army Knife" of the group.
I have noticed that when I ask G P T five point four to help me outline a new project, it challenges my assumptions in a way that feels more human. Claude just takes my assumptions and tries to make them work as perfectly as possible. It is the difference between a collaborator and an instrument. Anthropic wants Claude to be a precision instrument. OpenAI wants G P T to be a digital colleague. Both are valid, but you have to know which one you need for the task at hand.
That is a great way to put it. One thing that fascinates me is the shift in how we evaluate these things. We are moving away from these static benchmarks and toward more dynamic, agentic evaluations. How does the model handle a task where the environment changes? This is why we saw that "Cursor Incident" in episode fourteen seventy-one, where models were failing because the environment was shifting faster than they could adapt.
It feels like we are reaching a point where the raw intelligence of the big three is starting to plateau, or at least the gap between them is so small that the ecosystem and the specific architectural quirks are what matter most. We are in the "Deployment Era" now. It is no longer about how many parameters you have; it is about how efficiently you can use them in a live environment. We covered this a bit in episode fourteen seventy-nine when we looked at the new era of inference speed.
I think we are seeing a plateau in conversational quality, but the ceiling for agentic reliability is still very high. We are not yet at the point where I can give an A I a goal like, "build me a fully functional e-commerce site and launch it," and expect it to work one hundred percent of the time without human intervention. But we are getting closer. The eighty point eight percent score on S W E bench for Claude four point six is a massive leap from where we were even a year ago.
It makes me think about the internal routing again. If Claude is using forty billion active parameters, and it is hitting eighty percent on these coding tasks, what happens when we figure out how to route those experts even more efficiently? Is there a version of this where a ten billion active parameter model performs as well as Opus does today?
It is likely. We are seeing a lot of research into what is called sparse activation. Instead of just picking two or three experts per token, you might have a much more granular selection process. The goal is to minimize the energy cost. The energy consumption of these data centers is becoming a geopolitical issue. If you can get state-of-the-art performance with a fraction of the power, you win the market. This is why the A M D shift is so critical—it is about the power-to-performance ratio.
That brings us back to the A M D shift. If OpenAI can run these models on more energy-efficient hardware, they can afford to let the model think for longer. That thirty-second pause might be annoying for us, but if it saves them forty percent on their electricity bill because the model is checking its work instead of just brute-forcing a response, they will take that trade-off every time. It is a business decision as much as a technical one.
And it's not just electricity. It is about the physical footprint of the clusters. Moving away from Nvidia allows them to diversify their supply chain. It is a hedge against the volatility we have seen in the H one hundred and B two hundred markets. If you are building an "Operating Layer" for the world, you cannot be dependent on a single hardware vendor.
So, to summarize the instructional characteristics for Daniel: Claude is your careful architect, G P T is your creative collaborator, and Gemini is your hyper-connected, multimodal researcher.
That is a solid breakdown. And don't forget the price points. Two dollars for Gemini, two-fifty for G P T, and five dollars for Claude Opus. If you are a startup, those margins matter. You might find yourself using Gemini for your high-volume tasks and only calling the Claude four point six Opus A P I when you have a truly difficult problem that requires that deep, safety-first reasoning.
It is a multi-model world. No one is using just one substrate anymore. You are building a stack that leverages the strengths of each. It reminds me of the early days of cloud computing where you had your database on one provider and your front-end on another. The "AI Stack" is the new reality for developers in twenty twenty-six.
The Multi-Surface Operating Layer is basically the new O S. You pick your kernel based on what you are trying to compute. If you want to dive deeper into how these models are being used in specific industries, we did a deep dive on the shift toward multimodal perception in episode fourteen eighty-two that is worth a listen. It provides the context for why Google is winning on those graduate-level science benchmarks.
It is wild to think about how fast this is moving. We are talking about version four point six and five point four, and it feels like just yesterday we were excited about version three point five. The arc of deprecation is getting shorter every month. We talked about that "deprecation trap" in episode eight zero eight—it is a real challenge for long-term stability in software engineering.
It really is. Developers are constantly having to refactor their code to keep up with the new A P I capabilities. My recommendation for listeners is to start auditing your own workflows. Are you using a high-priced Thinking model for a task that a Flash or Sonnet model could handle just as well? Most of the time, the answer is probably yes. You can save a lot of money by being smart about which "expert" you are actually calling.
Audit your A I stack the same way you would audit your cloud bill. The savings can be massive, and with the performance of models like Gemini three point one Flash, you aren't really sacrificing much for the majority of everyday tasks. Save the "Thinking" models for the high-stakes logic.
And keep an eye on that recursive memory. If Google can really scale that to one hundred million tokens, the way we think about data storage and retrieval is going to change forever. R A G might just become a footnote in A I history.
Well, that is a lot of ground covered. We have looked at the M O E routing, the hardware shifts, the benchmark leaders, and the instructional personalities of the big three. It is a fascinating time to be in this space, even if we are all just trying to keep our heads above water with the release cycles.
It is the weird, fast-moving edge we love. I am already looking forward to whatever they drop in April. The pace isn't slowing down.
Thanks as always to our producer Hilbert Flumingtop for keeping the gears turning behind the scenes. And a big thanks to Modal for providing the G P U credits that power this show. If you are building agentic substrates or just need some serious compute, check them out.
This has been My Weird Prompts. If you are finding these deep dives helpful, a quick review on your podcast app really helps us out. It is the best way to help new listeners find the show.
You can also find us at myweirdprompts dot com for the full archive and all the ways to subscribe. We will see you next time.
See you then.
