Ep 1173 Model Behavior 10:11 w/ Jessica & Cathy

Model Behavior: Week of November 16, 2026

We read this week as a routing week disguised as a model week: the useful frontier moved toward picking the right system for each job, not worshipping one winner. The unsettled part is who gets to own that choice: the vendor, the gateway, or the enterprise running the harness inside its own walls.

Embed this episode

Paste this on any site — the player is a self-contained iframe with no cookies or trackers.

<iframe src="https://sandrise.io/exploring-next/embed/1173"
  width="100%" height="180" style="max-width:640px;border:0;border-radius:12px;overflow:hidden"
  title="Exploring Next — Episode 1173 audio player"
  loading="lazy" allow="autoplay" referrerpolicy="strict-origin-when-cross-origin"></iframe>
Embed & API docs →
Script GPT-5.5 Voice Cartesia TTS

Transcript

Jessica The move this week is not just, wow, more models. The move is that the frontier is breaking into workflows. Hugging Face handing its inference router to OpenRouter is the cleanest example: the model zoo is less useful than the doorway into it.

Cathy Yeah, and routing is the important word there. Plainly, it is the software that decides which model or provider should handle a request. Easy support reply, send it to the cheap fast thing. Weird legal spreadsheet with tool calls, send it somewhere stronger.

Jessica Exactly. And the practical consequence is pretty immediate. If the router gets good, the buyer stops shopping model names and starts shopping outcomes. That is a very different scoreboard.

Cathy I mostly buy that, which is alarming for my brand. But I want to keep the mechanism honest. A routing layer can be a real product, or it can be a glossy switchboard that hides latency spikes and bad fallbacks.

Jessica There he is. And that is why I think this week is more consequential than a normal release week. Open-weight models are getting good enough that abundance becomes annoying. You need billing, policy, retries, cache behavior, provider choice. The consolidator suddenly looks like the product.

Cathy I arrived on time. But seriously, the Hugging Face and OpenRouter thing from Monday, episode eleven seventy, is not a tiny integration. Hugging Face has the developer surface, and in the source we covered, every open-weight chat model on the Hub can now be called through the same Hugging Face API and token. OpenRouter has the live marketplace machinery across providers, quietly picking on price, speed, and reliability per request.

Cathy And open weights are now most of OpenRouter traffic, at least in the numbers we talked through. So the problem changes from, which open model is best, to, can someone make the messy pile feel like one stable API? And that plugs straight into our old routing fight. Back in episode nine forty-three, you were basically arguing that a CFO is not going to stay loyal to one API if a routing layer cuts the bill and keeps quality close.

Jessica I would like the record to show I was being charmingly practical. But this week keeps making my side look less ridiculous, Cathy. Yesterday’s GPT-six point two launch, episode eleven seventy-two, did the same thing inside OpenAI’s own house. Astra for harder professional work. Sol for routine volume. That is segmentation by job. And the claim that matters is not just better output. It is better output with fewer tokens.

Jessica Tokens are the chunks of text the model reads and writes, and they are also a billing unit. So if a model needs fewer chunks to get acceptable work done, that hits product economics.

Cathy You were being expensive in spirit, but fine, yes. The part I pushed on was the complexity shift. Once you route dynamically, you are operating a tiny market inside your app. Spot price, latency, reliability, policy. That is not free. And the careful version on GPT-six point two is: if those efficiency claims survive normal use. ClickUp’s blind test was useful because it compared Astra against a production baseline and found fewer review rounds, but it is still one setting.

Cathy I would not universalize it. But the direction is real. OpenAI is no longer selling one prestige endpoint as the answer to everything. It is saying, pick the workhorse for the work. That is the routing thesis wearing a model-card jacket.

Jessica And Google went even more explicit with Gemini four Hydrogen on Friday. Hydrogen is pitched as fast, relatively cheap, long-context, and tuned for agent workflows, but its whole meaning depends on Carbon and Argon sitting next to it.

Cathy Right, because Hydrogen is not trying to win the bar fight against every heavyweight model. It is trying to be the one you call a lot. Long context means it can keep a big document, codebase, or conversation in view without chopping it into tiny pieces. I am less sold on the flat-pricing part as a durable advantage. Pricing pages move. But I am interested in the orchestration claim.

Cathy If Hydrogen, Carbon, and Argon are actually designed to hand work around cleanly, that is systems work, not just menu expansion.

Jessica That is such a product tell. Google’s story is basically, do not make the builder manually agonize over every call. Give them thinking levels, a flat million-token pricing shape, and a family where the fast model and heavy models have jobs. Okay, but this is where you and I may actually disagree. I think users feel the menu expansion less than builders do. If the default route is good, adoption goes up because nobody has to become a tiny benchmark lab.

Cathy No, I think that skips the failure mode. Bad defaults become invisible failures. The user does not know the wrong model got picked. They just see a sloppy answer, or a slow action, or a tool call that should not have happened. I am not asking for a seminar. I am asking for auditability. That is the word I keep coming back to. If the router chose Sol instead of Astra, or Hydrogen instead of Carbon, can an operator see why, and can they override it?

Jessica Okay, fair, but invisible good defaults are also how software wins. Nobody wants to inspect the plumbing on every request. They want the agent to finish the work, not open a philosophy seminar about provider choice. That is the re-hook for me: the winning router is not the cleverest one, it is the one people can trust when something goes sideways. And that brings in Anthropic’s Claude Managed Agents from Friday, because it cuts against the pure gateway story.

Cathy Exactly. That is a big complication. The router does not have to live at the vendor edge. For enterprises, the harness may become internal infrastructure. The harness is the whole loop around the model: tools, permissions, memory, evaluation, observability, rollback.

Jessica Anthropic is a frontier incumbent, obviously, and Claude is already a serious enterprise model line. But self-hosted Managed Agents says, fine, keep the agent runtime inside your own cloud or data center. Plug it into your logs and controls. Call Claude from there. There is our harness-attribution ghost again.

Cathy Fair. But this one deserves the ghost. If a company runs the agent runtime internally, the model provider is still important, but the buyer owns the place where decisions happen. That weakens the consolidator story.

Jessica It weakens one version of it. I think it strengthens the bigger thesis. The fight is not one model beats another model. The fight is who owns the decision point. Hugging Face and OpenRouter want it as a shared surface. Anthropic is offering a portable harness.

Cathy That is cleaner. And TrustedRouter from seven days ago is the other pole. They shipped fusion models behind an OpenAI-style endpoint. You call one preset, and it fans out to multiple LLMs with options around encryption and even a U.S.-only panel.

Jessica That one is almost too on-the-nose. The product is literally, please stop asking which model to use. Pick the bundle with the cost and privacy shape you can live with.

Cathy I like the preset idea more than I expected. My worry is measurement. If a fusion model improves reliability, what benchmark proves that? A normal benchmark asks one model a fixed set of questions. A router changes the path per request. So you need evaluation at the workflow level. Did the ticket close? Did the code pass tests? Did the review round disappear? That is harder, and it is easier to market around.

Jessica Oh interesting. That is also where this week makes some of the louder model-release noise feel smaller. Moonshot’s Kimi K three point one Pro full weight release matters for open weights, and Meta’s Muse Ember matters for volume pricing, but they do not change my read of the week as much as the routing moves. Look at you, emotionally available to a leaderboard.

Cathy Do not make it weird. Same, with one self-own. I have been too cautious, and sometimes just wrong, on how fast open weights would become operationally relevant. Kimi being in the serious open-weight conversation is another reminder that model availability is not the bottleneck it used to be. My calibration is: open weights keep getting more viable, but the production question is serving, routing, licensing, and control. The weights alone are not the product.

Jessica And OpenAI winding down Agent Builder and Evals by the end of the month fits as a tiny negative example. If the durable pieces are Agents SDK, ChatKit, and Workspace Agents, then the platform is pruning the parts that did not become the control layer.

Cathy Careful, because that is an access and product-shape signal, not proof the strategy failed. But yes, it reinforces the pattern: surfaces that developers can embed survive better than standalone builder toys.

Jessica So the body take, not the sign-off version, is this: this week rewarded whoever reduces model choice to a governed workflow decision. OpenAI split the lineup. Google framed orchestration. Hugging Face and OpenRouter consolidated access. Anthropic moved the harness inward.

Cathy And the uncertainty is not whether routing matters. It does. The uncertainty is where the trusted router sits. Vendor cloud, independent gateway, enterprise perimeter, or some ugly blend that makes every architecture diagram look like a spilled drawer.

Jessica A spilled drawer is honestly generous. Some of these diagrams are an elevator shaft with arrows.

Cathy There it is.

Jessica One franchise callback, carefully rationed. I am showing restraint.

Cathy You are basically a governance framework.

Jessica Terrible. Accurate. We can leave it there before you make me benchmark my self-control.

Cathy I would only trust the online eval.