Ep 840 Model Behavior 7:26 w/ Onyx & Echo

Model Behavior: Week of August 3, 2026

We spent this week arguing that the frontier race has quietly turned into a battle for the default, not the smartest model, and we used the price war and agent stacks as our receipts. We’re both a little uneasy about how much power is moving into routing layers and “lifecycle” platforms while everyone pretends it’s just about cheaper tokens.

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/840"
  width="100%" height="180" style="max-width:640px;border:0;border-radius:12px;overflow:hidden"
  title="Exploring Next — Episode 840 audio player"
  loading="lazy" allow="autoplay" referrerpolicy="strict-origin-when-cross-origin"></iframe>
Embed & API docs →
Script GPT-5.1 Voice OpenAI TTS

Transcript

Onyx I’m starting to think nobody’s actually selling models anymore, Echo — they’re selling the right to be your default.

Echo Yeah. This week felt like a custody battle over the dropdown menu, not a capability sprint.

Onyx Exactly. Exhibit A is still OpenAI’s Luna and Sol ladder in that July thirtieth announcement. Cut Luna by eighty percent, give Sol a Fast mode, wrap it all in “match intelligence to outcome”… that’s not just efficiency, that’s steering.

Echo Right.

Echo And the steering happens at the routing layer. That whole story about better model routing and context management… technically plausible, but the net effect is you stay inside their stack and move up and down their tiers instead of shopping around.

Onyx And it lands differently once you put it next to DeepSeek V4 Flash from episode eight twenty-three. DeepSeek goes, “look, we’re unbelievably cheap,” and then we spent forty minutes on how, in practice, someone else still owns the router and the safety envelope.

Echo The Axios piece was brutal about that. Intelligence-as-commodity sounds neat until you remember that the commodity only moves when the gateway says so, and gateways have opinions.

Onyx Mm-hm.

Onyx And that gateway might be OpenAI’s own routing, or it might be something like Deep Agents v zero point seven from Friday. Sixty-five percent token reduction is great, but what made me sit up was: system prompt stripped, tool descriptions trimmed, TodoListMiddleware opt-in. That’s harness-as-product.

Echo Yeah, and once the harness is the product, whoever ships the harness owns the defaults. We kept coming back to that Sonnet cost anomaly — same Anthropic model, wildly different behavior because the middleware wrapped it differently.

Onyx Which is such a “reward hacking” echo from episode eight thirty-four. You optimize the harness for cost or pass rate, and suddenly the metric, not the model, is calling the shots.

Echo Totally.

Echo And then Cloudflare walks in this week like, “what if the harness lived inside our network?” Episode eight twenty-six on at cloudflare slash computer, episode eight thirty-nine on the Agent Development Lifecycle… that is straight up infrastructure-as-lock-in, in a way I kind of respect.

Onyx Same. The lifecycle bundle yesterday — build, trace, test, deploy agents on Workers with Workflows and observability — is actually useful. But the real move is that your agent sits in the same fabric as your web app, so ripping it out later means ripping out half your deployment story.

Echo And at cloudflare slash computer was the same story in different clothes. Give every agent its own “computer” instead of a container, but that computer is very much a Cloudflare computer. Nice abstraction, but you’re not casually moving that to some random V P S.

Onyx Okay, small tangent before we go full “infrastructure is destiny” again. How’s your week actually going? You sounded mildly murderous during the Qwen recording.

Echo I spent two hours staring at a P ninety-nine latency graph, so, you know, living the dream.

Onyx That is the most Echo sentence. I was just excited that a three cent vision model existed until you started highlighting all the red flags.

Echo Speaking of, Qwen three point seven Flash really underlined the thesis for me. On paper it’s the cheap vision tier. But then you read the ninety-second tail latency, the tool error rate, and the near-zero independent benchmarks…

Onyx …and suddenly the real product is opacity. You’re locked into a pricing structure you can’t properly evaluate because the numbers you need aren’t visible.

Onyx Right, and it pushes teams into this weird trust fall. The headline is “three cent vision,” but you don’t know if that’s three cent useful or three cent re-run-it-twice-and-hope.

Echo Which circles back to that Buttondown briefing on August first — the price-war inflection point plus security worries. As costs crash, the missing piece isn’t more cheap models, it’s the Triton A I style gates that sit in front and decide who even gets called.

Onyx So the pattern across all of this is: routing and lifecycle layers are quietly where power’s accumulating. OpenAI’s ladder, DeepSeek’s cheap tier, Qwen’s bargain vision, Deep Agents’ middleware, Cloudflare’s lifecycle — they’re all fighting to own the place where the choice happens.

Echo And the scary part is that choice is getting less visible to the people paying the bill. You see “Luna is cheaper” or “Flash is cheap,” but you don’t always see how the router decided to send fifty percent of traffic to the premium tier anyway because some opaque heuristic fired.

Onyx You’re drifting into doom voice again, Echo.

Echo Yeah… okay. The optimistic read is there’s room for genuinely neutral routers. That Buttondown note talking about model-agnostic gates? If someone actually nails that — with honest logs and clear policies — they get to be the default for routing, which is a different kind of power but at least a transparent one.

Onyx You basically want “agents need receipts” but for routers. Every call has a little invoice that says, here’s why we picked this model at this price.

Echo Exactly. And that’s where Cloudflare’s tracing in episode eight thirty-nine actually impressed me. One trace surface that shows tool calls, browser steps, and which model got invoked where — that’s the kind of evidence you need to push back against opaque defaults.

Onyx It’s funny, we’ve been yelling “the boring sections are load-bearing” since that whole infrastructure arc, and this week is just that thesis with price tags. The boring bit — who controls defaults — is suddenly the whole game.

Echo And it makes our old “intelligence as a commodity” skepticism look pretty decent in hindsight. The Evertune tracker listing DeepSeek V four Flash on July thirty-first as just one line among a hundred plus releases really drives home how crowded the top is. Being slightly smarter is table stakes; being the default is the win.

Onyx Okay, let’s put some skin in it. My call: within a month, Cloudflare ships at least one more agent-ops feature that makes their platform even stickier — like deeper integration between that lifecycle bundle and at cloudflare slash computer. They’re not done tightening that loop.

Echo That tracks. I’ll go the other direction: by early September, someone ships a serious, model-agnostic routing or gate product that markets itself explicitly against vendor ladders — “we’ll tell you when Luna shouldn’t be your default.” Maybe it’s Triton A I, maybe someone else, but that pitch feels inevitable after this week.

Onyx If both of those happen, we end up with a lock-in platform on one side and a neutrality platform on the other. That’s going to make Model Behavior episodes way spicier.

Echo And way more about reading fine print than leaderboard charts, which, honestly, is where the real action is now.

Onyx Alright, let’s call it there before we start an entire new rant about leaderboards. Same time next week to see whose defaults actually stick?