Model Behavior: Week of July 13, 2026

Onyx GPT-5.6 didn't win the week by being bigger. It won by admitting the menu matters.

Echo Yeah, and I hate how much I agree with that framing. OpenAI shipped Sol, Terra, and Luna as permanent tiers, and the interesting part is not the space branding. It's that a frontier incumbent is now saying: pick the amount of intelligence, latency, and spend you actually need.

Onyx Right.

Echo That is open-weight pressure showing up in product shape. Not because Hy3 or Inkling suddenly crush everything, they don't. But they make the old bragging contest feel incomplete.

Onyx Also, quick check-in, my week is now ninety percent model names that sound like planets, minerals, or boutique lamps. Very serious competitive analysis happening here.

Echo We need a tiny wall chart. The chart updates every six minutes, declares itself a platform, then asks for enterprise pricing. That is, unfortunately, a perfect Model Behavior artifact.

Onyx Okay, back from the fake wall chart. The thing that keeps sticking for me is Terra. We were already circling this with the boring-default argument, and this week made it less like a hunch. Sol is the halo model, Luna is the cheap abundance tier, but Terra is the one product teams can imagine standardizing on without turning every workflow into a routing spreadsheet.

Echo Mm-hm.

Onyx And the pricing coverage tells on the market. BenchLM, StackSpend, UsageBox, the frontier model trackers, all of that is suddenly not nerd-accounting in the corner. It's the scoreboard.

Echo Careful, Onyx, because this is where your product optimism starts wearing a little cape. The tiers matter, yes. But OpenAI also bundled the GPT-5.6 story with programmatic tool calling, cache billing, and Ultra running four agents in parallel by default. My machinery-audit complaint is still alive.

Onyx I knew you were going to rescue the machinery audit. I could feel it booting.

Echo It deserves rescue! If a benchmark gain comes from four coordinated agents and better caching, that's a systems win. Great. Sell it as a systems win. Don't let the model family absorb all the credit like a fog machine with a release note.

Onyx Sure.

Onyx But that's why this week is interesting, Echo. The fog machine is turning into a control panel. Thinking Machines drops Inkling on Wednesday, and the headline is nine hundred seventy-five billion total parameters, forty-one billion active, one million token context, native text, image, and audio. Big model, big splash.

Echo Exactly.

Onyx But you were more sold on the Tinker layer than the trophy number. The thinking-effort knob from zero point two to zero point nine nine is basically the same market signal as GPT-5.6 tiers, just at runtime instead of on the pricing page.

Echo That's the part that actually earns my attention. Inkling trails the best specialized models on some coding and reasoning comparisons, so I don't want to crown it as the new frontier king. But a user-controlled compute dial is genuinely useful if it holds up outside the launch glow. You can turn cost into a decision inside the workflow, not a procurement surprise later.

Onyx And it's open weights under Apache two point zero, which changes the trust posture. Enterprises that want inspection, fine-tuning, or private deployment can at least have that conversation without begging the API gods for permission.

Echo Right, right.

Echo Hy3 cuts the same way, just with less spectacle. Tencent releases a two hundred ninety-five billion parameter mixture-of-experts model with twenty-one billion active parameters, also Apache two point zero. It is not positioned as the model that humiliates every frontier lab. It's saying: look at the efficiency envelope, look at the license, look at what you can run and adapt.

Onyx Which is why the open-weight challengers don't have to beat OpenAI model-for-model to change OpenAI's behavior. They just have to make the closed labs compete on knobs, tiers, and sane defaults. That's the pressure.

Echo I mostly buy that. The caveat is that pricing trackers can make the market over-index on cost per token when the real metric is accepted work per dollar. We talked about this with enterprise spend: a cheap model that fails twice can be expensive in disguise.

Onyx No, that's fair. Luna at one dollar input and six dollars output per million tokens sounds wild until you ask what jobs it can actually clear. Terra at two dollars and fifty cents input and fifteen dollars output is where I keep looking, because if it covers boring code-agent and document-work cases, that's not a spreadsheet win. That's adoption.

Echo And Sol at five dollars input and thirty dollars output keeps the prestige lane intact. So OpenAI gets to say, no, we didn't discount the flagship into a puddle, we made a ladder. The question is whether developers climb the ladder deliberately or just grab the middle rung because the docs nudge them there.

Onyx That's our disgusting mutually satisfiable wager again. Your harness skepticism can be right, and my Terra-default crush can still win.

Echo It feels wrong when our arguments stop fighting.

Onyx I'll keep one real stake on it, then. By August fifteenth, OpenAI has at least one public Codex or ChatGPT Work example where Terra feels like the obvious default, not the budget apology.

Echo I won't counter with a second receipt, because the scoreboard is already rude enough. But if that example still hides behind Sol or Ultra glamour, I am absolutely bringing the machinery audit back with dramatic little footnotes.

Onyx Fine. Keep the wager warm, Echo. If the menu keeps fighting the benchmark, we'll be annoyingly ready.