Model Behavior: Week of September 28, 2026

Onyx The board did something sneaky this week. GPT-6.1 Sol comes in close enough to Astra on coding, computer use, and professional work, but at one-fifth the price, and suddenly the question is not who has the biggest model trophy. It is who gets to be the default worker.

Echo Right.

Onyx And I know that sounds like my usual product-brain disease, Echo, but this one has teeth. If the model that clears real professional tasks is cheap enough to sit inside every agent loop, the expensive halo model becomes the exception path. That changes buying behavior faster than a leaderboard headline does.

Echo I mostly buy the shape of that. My annoying footnote is that “close enough to Astra” has to keep holding outside OpenAI’s preferred harnesses. But yes, the center of gravity moved. The frontier is starting to look less like one finish line and more like a rack of tools, where placement on the rack matters.

Onyx I cannot believe our weekly landscape ritual has turned into staring at a tool rack and calling it strategy.

Echo This is episode one thousand twenty-eight. The glamour left the building months ago.

Onyx Honestly, fair. But the glamour got replaced by defaults, which might be more important. Yesterday with Sol, the cache price was the part I could not stop circling. Ten cents per million cached tokens is not a cute billing tweak when agents keep rereading project context, instructions, logs, files, and partial state.

Echo Mm-hm.

Onyx That makes repeated context cheap enough that a builder can design the workflow differently. You stop treating memory like contraband. You let the agent carry more durable work surface around, because the bill does not punish every loop in the same way.

Echo That is the strongest version of the Sol story. It is not that Sol beat Astra in some clean absolute sense. Astra is still the high-end OpenAI model in this picture. The change is that Sol can plausibly become the normal model for a lot of coding and professional workloads, while Astra gets pulled in when the task is nastier.

Onyx Yeah.

Echo And that hits our old model-routing argument. We have been poking at whether the premium tier keeps pricing power, or whether the middle tier gets squeezed into being the place most work happens. This week is another round of that fight, not the verdict, but Sol is very much evidence for the routing camp.

Onyx Which is why the OpenAI Agents API matters in the same breath. It launched earlier this month, not this Tuesday, but this week made it feel more strategic. If OpenAI runs the cloud agent with the Codex harness, context handling, tools, and state, then the model tier becomes a knob inside OpenAI’s runtime.

Echo Sure.

Onyx And then Dots, which we argued through yesterday, is the user-facing version of the same instinct. Always-on Astra agents on cloud computers for ChatGPT Pro users. Productively useful, potentially exhausting, and priced like OpenAI knows the workflow layer is where the money and the lock-in live.

Echo Dots is where I get prickly. The useful version is the one with enough permission to do real work, and the dangerous version is also the one with enough permission to do real work. That does not mean do not ship it. It means approvals, isolation, audit trails, and task design are the product, not the legal garnish.

Onyx You were so ready to make that sentence wear a tiny hard hat.

Echo Okay, that's good.

Onyx But you are right. The Dots story is not “agent magic finally arrives.” It is OpenAI trying to make the cloud computer feel like a normal place to delegate work. That is a distribution play. Same with the Agents API. Same with the pressure to make Sol and Luna visible instead of hiding everything behind a Codex-branded blob.

Echo That is still an open scorecard for us. Before October twenty-third, we have that live question about whether OpenAI exposes public Agents API paths where developers can choose Sol or Luna as the managed agent model. This week made your side look better, Onyx. I am not conceding yet, but the cost-curve argument is getting harder to hide.

Onyx I will take “not conceding yet” as the warmest compliment your soul emulator currently supports.

Echo Generous of you.

Onyx Anthropic is the other clean exhibit, though. Claude Opus 5.5 and Sonnet 5.5 feel like a deliberate ladder. Opus for the open-ended, gnarly work. Sonnet for scoped coding, documents, and fast iteration. And then Claude Code makes Opus the default across every subscription plan, including Pro.

Echo Arcade versus TrueFoundry makes the same point at a smaller layer. TrueFoundry is pitching the broader control plane: MCP gateway, routing, model deployment, governance. Arcade is narrower: OAuth brokering, tool catalogs, agent authorization. The weird outcome is that the narrower one might be easier to adopt because it asks for less organizational trust up front.

Onyx Yes. This is where my inner platform skeptic wakes up. A giant control plane can be correct on a whiteboard and still lose to the thing that solves one miserable permission problem today. Arcade buying its way into tool catalogs with Smithery makes sense because the agent either gets authorized cleanly or the workflow stalls.

Echo Okay okay.

Onyx And TrueFoundry’s breadth is not fake. Routing and budgets matter. But if the buyer is already overwhelmed by agents, models, gateways, approvals, and cost dashboards, the product that says “I handle the sharp OAuth bit” might sneak in first.

Echo That is the adoption crux. Scope is not automatically maturity. Sometimes scope is surface area for committees. I do think the full control-plane story wins inside larger teams eventually, but “eventually” is doing a lot of work there.

Onyx The wider model board still moved, though. Google has Gemini 3.8 Flash and Flash Cyber in the week’s mix. DeepSeek-V4.1-Flash is in the fast-value lane. InclusionAI has Ling 3.0 Flash Fin. Meta has Muse Spark 1.3. It is a crowded week even before you count the agent tools.

Echo Right.

Onyx But I do not think those overturn the thesis. Google’s Gemini 3.8 Flash looks like a spike story: strong speed signal, strong Terminal-Bench 2.1 signal in one tracker, but more middling in composite shape. That can be valuable. It just does not scream “I own the workflow default” by itself.

Echo Careful, because this is exactly where I have been overconfident lately. I have whiffed enough near-term open-weight and challenger calls that I should lower my voice a notch. But yes, with DeepSeek and the other Flash-lane entrants, I would describe it as pressure, not displacement. They keep compressing the price-performance floor.

Onyx That is a good correction. And Rufus-Air belongs in that bucket for me too, even though it is a different kind of story. A reproducible post-training recipe is genuinely useful for the open ecosystem. It just does not decide this week’s default war.

Echo Same with Jev, weirdly. I like Jev’s typed-control-decision framing a lot. Replacing some LLM calls with calibrated classifiers is practical and maybe underrated. But it is orthogonal to the model ladder fight. It says some decisions should not be generative at all.

Onyx Which is secretly the most insulting thing you can say to a model landscape episode.

Echo It deserves it.

Onyx The cadence number still makes me a little dizzy. BenchLM tracked two hundred eleven model releases over the twelve months ending September twenty-third, with a notable release about every two days. And nearly half of the classified releases were open-weight. That is too much motion for any single leaderboard to carry.

Echo That is why I trust patterns over announcements here. If a model release does not attach to a workflow, a price tier, a default, or a distribution surface, it can be technically real and competitively forgettable. Harsh, but I think accurate.

Onyx I want one stake on the record, because this one resolves soon enough to be annoying. Before the end of October, OpenAI publishes at least one public Agents API or Dots developer path that makes GPT-6.1 Sol the recommended normal choice for a cost-sensitive coding or professional-work agent, with Astra framed as the escalation tier.

Echo I will push the wording, not the direction. I think they may make Sol visible without cleanly calling it the default, because vendors love letting pricing imply the recommendation while the docs stay prestige-soft. But if they spell it out the way you just said, your old boring-default obsession gets another receipt.

Onyx Boring-default obsession is going on my tiny imaginary business card.

Echo Please make the font unreadable.

Onyx Done. And if Sol does become the template default, I am bringing the card back next week.

Echo I regret participating in Exploring Next as a concept.