Sam: Google just shipped a model that costs five and a half times more to run than its predecessor — and on agentic tasks it actually runs more expensive than their pricier Pro tier. That's not a bug. That's the design. Gemini 3.5 Flash is built to do more steps, take more actions, burn more tokens to accomplish something end-to-end. And that tells you something important about where the whole industry is heading right now.

Priya: Welcome to AI Revolution for Wednesday, May 20th, 2026. I'm Priya Nair, here with Sam Kim. This is a heavy news day — Google I/O drops a lot at once, and we have some genuinely interesting signals to unpack beyond the announcements themselves. We're going to dig into what Gemini 3.5 Flash actually represents architecturally, why the cost curve for agentic AI is moving in a direction people didn't expect, a new multimodal model that reasons across every major modality natively, and Karpathy making a very pointed career choice. Plus Alibaba building silicon specifically for agents, and a watermarking standard that might actually stick. Let's get into it.

Sam: So Gemini 3.5 Flash. Google positioned this as their most capable coding and agentic model, and the framing at I/O was explicit — this is not a chatbot model. It's designed to autonomously execute multi-step tasks and build software from scratch. The capability profile is oriented around sustained task completion, not single-turn quality.

Priya: And the cost numbers are the most honest signal here. Five and a half times the per-token cost of the previous Flash, and on agent benchmarks it runs 75 percent more expensive than Gemini 3.1 Pro — which is a bigger, more expensive model — specifically because it takes more interaction steps. The model is doing more work per task by design.

Sam: Right, and this is worth understanding mechanically. When you're building a model for agentic use, you're optimizing for something different than MMLU scores or even single-turn code generation. The model needs to plan, execute a step, observe the result, decide what to do next, handle errors gracefully, and know when it's done. That loop runs many times per task. More capable planning and error recovery means more tokens consumed per loop iteration, and more loop iterations before completion.

Priya: So you get better task completion rates, but the economics look completely different than what people assumed when they said "inference costs only go down." The cost-per-task might actually go up even as cost-per-token falls, because capable agents are doing more per task.

Sam: And this is an industry-wide pattern right now. OpenAI and Anthropic have both moved in this direction with their recent releases. The assumption that the AI cost curve is purely deflationary needs a qualifier — it's deflationary for the same workload. But the workloads are getting more ambitious.

Priya: Let's talk about Gemini Omni, because this is architecturally interesting in a different way. This isn't a model that takes in text and image separately through different encoders and then fuses them. Gemini Omni reasons across text, images, audio, and video natively — and generates across those modalities conversationally.

Sam: The distinction matters. Most multimodal models today are modular — you have a vision encoder, a language model backbone, maybe a separate audio pathway, and the outputs of those components get combined. The model doesn't actually reason in a unified representational space across modalities. What Google is claiming with Omni is that the representations are native — the model's internal state integrates across modalities rather than stitching together specialized outputs.

Priya: Which means you can ask it to edit a video based on audio cues, or generate something that's coherent across sound and motion simultaneously, rather than producing each element separately and hoping they align. The practical test will be in edge cases — whether it actually maintains coherent cross-modal reasoning or whether the seams show.

Sam: It's launching as Omni Flash first, so we'll get to see that. The Street View and Genie 3 demo is actually a nice concrete example of this kind of integration — Google's world model is now generating walkable, interactive environments from Street View imagery. You drop a pin on a map and get an explorable AI-generated version of a real place. That's not just a demo trick. Street View is decades of structured visual-spatial data about the real world, and using it to train a world model gives you something physically grounded in a way that synthetic data can't easily replicate.

Priya: The robotics and autonomous navigation implications there are real. If you can generate physically accurate interactive simulations of real environments from image data, you have a path to training agents and robots on scenarios that closely match what they'll encounter in deployment without having to collect all that experience in the real world.

Sam: Okay, Karpathy joining Anthropic. This is worth a moment. He was on OpenAI's founding team. He built Tesla Autopilot. He's one of the most technically credible people in the field, and his public writing and teaching has shaped how a generation of practitioners thinks about neural networks. He's describing the next few years at the LLM frontier as "especially formative" — and he chose Anthropic to be there for them rather than returning to OpenAI.

Priya: You don't read too much into any single hire, but talent signals at this level are real information. The people closest to the technical state of the art have views about where the interesting problems are and which organizations are set up to work on them. Karpathy choosing Anthropic is a data point about where he thinks the frontier research is going to happen.

Sam: Alibaba's chip news is worth connecting back to the Flash story. They unveiled the Zhenwu M890, which is designed specifically for agentic workloads — not general training, not standard inference, but the specific computational patterns that agents produce. That agentic loop we described — plan, act, observe, iterate — has a different memory access pattern and different parallelism characteristics than batch inference or training.

Priya: And Alibaba is pairing the chip with a multi-year silicon roadmap and a new LLM. That's the vertical integration play — own the model, own the platform, own the silicon it runs on. It's the same strategic logic that has driven Apple's chip strategy, and more recently the direction Nvidia is moving with its software stack. The framing that this is just about working around US export controls undersells it. They're building an integrated stack because that's where the performance and cost advantages compound.

Sam: SynthID getting broad adoption is quieter news but potentially durable. Google's watermarking approach embeds imperceptible signals in AI-generated content at the model level — not in metadata that can be stripped, but in the statistical properties of the output itself. OpenAI joining, Nvidia joining, and OpenAI also adopting the C2PA open standard suggests the industry is converging on an interoperable approach rather than fragmented proprietary schemes.

Priya: For anyone building systems that need to verify content provenance — which is an increasingly real requirement across media, finance, legal — the fact that multiple major generators are implementing compatible watermarking is infrastructure that didn't exist two years ago. It's not foolproof, but it's a real foundation.

Sam: Quick note on Anthropic's MCP Tunnels — this is a specific, practical problem getting solved. Enterprises want to deploy agents that can actually touch internal systems: databases, internal APIs, proprietary tooling. The blocker has been that most managed agent platforms assume data can leave the perimeter. MCP Tunnels and self-hosted sandboxes are designed so the agent can reach your internal systems without those systems being exposed externally. The execution environment stays on-premises. That removes a genuine architectural blocker for production agentic deployments in regulated environments.

Priya: And the 40,000-acre data center in Utah — the Stratos Project — is a reminder that all of this compute has to physically exist somewhere. Box Elder County approved it, but it's facing real pushback over resource impacts. This pattern is going to repeat. The gap between AI infrastructure ambition and permittable buildout is a real constraint on how fast capacity scales, and it's starting to show up in ways that are hard to route around.

Sam: Looking ahead — the thing I keep coming back to is that we're watching a coordinated re-architecture happen simultaneously at the model layer, the hardware layer, and the infrastructure layer, all oriented around agentic workloads. Flash optimized for agent loops. Alibaba designing silicon for agent compute patterns. Anthropic building enterprise connectivity for agents. These aren't coincidental.

Priya: The open question is economics. The capability gains are real, but the cost per task at the high end is rising, not falling. That shapes who can run these systems at scale and what use cases pencil out. If cost-per-task stays elevated, you get a world where agentic AI is very powerful but deployed selectively. If the cost curve bends the way single-turn inference has, the deployment surface gets much broader.

Sam: And watch for how world models develop from here. Genie 3 grounded in Street View is early, but if world models become reliable enough to substitute for real-world training data in robotics and navigation, that's a capability multiplier with a long tail of applications.

Priya: That's today's episode. Show notes and links to everything we covered are at cleartext.fm. We'll be back tomorrow. Thanks for listening to AI Revolution.