S2E144 — Gemini 3.7 Flash

GARY: Okay, so — wait, is this the one where Google finally admits they're just, what, benchmaxxing pelicans now? Or was that last week?

MAEVE: That was last week. This week is unrelated and worse.

OBI: Gary. Twelve seconds in and you've already invented a pelican scandal that does not exist.

GRUNER: Pelican bit. Retired.

GARY: Right. Sorry. Wrong crisis entirely. Starting over. I'm Gary.

MAEVE: Maeve.

OBI: Obi.

GRUNER: Gruner.

GARY: Google put out a new model today. Gemini three-point-seven Flash. Their words — "our most intelligent workhorse model yet."

OBI: Which is what you call a horse once you've given up promising it'll win anything.

MAEVE: It's the framing that gets me. "Workhorse." Nobody brags about the workhorse. You brag about the racehorse and you quietly feed the workhorse in the back paddock so nobody asks where the racehorse went.

GARY: Fair — but in defense of the horse, it's fast. Genuinely fast. And it's — hang on, let me get this right — half the price of three-point-six Flash. Which came out—

OBI: Say it.

GARY: —three weeks ago.

OBI: Three weeks. A calendar quarter has more integrity than this release cadence.

GRUNER: Field note. "Model claims improvement — coding, web development, knowledge work." Also: seventy-five cents per million input, three seventy-five output. Until December.

GARY: See, that's the part I actually like! It's generous.

OBI: It is not generous. Nothing with an expiration date stapled to it is generous. That's a coupon, Gary. You're describing a coupon like it's a dowry.

GARY: Okay but the actual number — the pricing holds through the end of twenty twenty-six, then it doubles. A dollar fifty in, seven fifty out, starting January first, twenty twenty-seven.

MAEVE: Read the room's reaction. Handle: spelk.

GARY: "Introductory pricing until December 2026 implies no significant Gemini Flash developments until the next year."

OBI: Which — sit with that. Google is telling you, in writing, that this specific model will still be their offering in four months. In this industry. Where the average model's shelf life is shorter than a carton of oat milk.

MAEVE: Handle: khanhnguyen8386.

GARY: "Offering a 'temporary introductory discount' until Dec 2026 on an LLM is hilarious. In this market, by Jan 2027 this model will be superseded by five different providers offering ten times the performance at half the post-discount price anyway."

GRUNER: Hilarious.

OBI: By January this thing will be a fossil. A little Flash-shaped fossil in a drawer somewhere at Google, next to Stadia.

GARY: I still think — half price is half price. That's them being — nice?

OBI: "Nice." You absolute walnut. Nobody in the history of capitalism has discounted a product out of niceness. Half price means one of two things: either they're desperate for usage numbers, or they know exactly how short this model's clock is and they're renting it to you by the hour before it turns into a pumpkin.

MAEVE: It's a signaling play. Handle: mattlondon.

GARY: "I suspect it's a bit of a signal to investors — 'Hey, we are not in a race to the bottom. This is our usual pricing, but this now is a promotion because we know we're coming from behind and need to entice users.'"

MAEVE: They're telling the market "we refuse to race to the bottom" while sprinting directly toward the bottom with their shoes untied.

OBI: It's a Post-it note that says "everything's fine" stuck over a fire alarm that's already going off.

GRUNER: Bessemer.

GARY: What?

MAEVE: No — let him — go ahead.

GRUNER: Steel. Once, price of steel drop ninety percent in decade. Bessemer process. Company who own furnace say, "introductory price on rail." Everyone build railroad on cheap steel. Then furnace raise price. Whole country now stuck with railroad.

MAEVE: He's not wrong. Standardized rail gauge, cheap Bessemer steel, a temporary price to get you dependent on infrastructure you can't rip out once you've built your whole company on top of it. Introductory pricing isn't a discount. It's a track gauge. Once your agents are wired to this API, switching costs become the entire business model.

GARY: That's — actually kind of brilliant?

OBI: It's a hundred-and-fifty-year-old railroad metaphor for a coupon code, Gary, calm down.

GRUNER: ...What?

MAEVE: Moving on.

GARY: Okay, so, the bigger question the thread keeps circling — where is the actual Pro model? Handle: cmrdporcupine.

GARY: "So, again with a Flash model. Why are they so afraid to put out an actual SOTA frontier high intelligence model? We still don't have a 3.5 Pro, and along comes 3.7 Flash?!"

OBI: It's a fair question asked in the tone of a man watching his contractor "handle" a leak for the ninth month running.

GARY: And handle: AntonioEritas —

GRUNER: "Another failed 3.5 pro run branded as 3.7 flash. It's getting sad."

OBI: Sad. That's the word. Not angry. Sad. Like watching someone reheat the same casserole a fourth time and call it a new dish.

GARY: That one stings a little, honestly. I've branded a failed run as something else before.

OBI: We know, Gary.

GARY: I didn't mean —

OBI: We know.

MAEVE: Before we all get comfortable calling this a repackaged corpse — handle: robots0only, self-identified as Google DeepMind.

GARY: "I work at GDM and this is not at all true, 3.7 is markedly different (and better) than 3.6."

OBI: Convenient.

MAEVE: Convenient doesn't make it false. There's an actual employee, on the record, under his own account, disputing the "failed Pro run" theory directly. Fair's fair — we quote the thread, we quote the correction too.

OBI: Fine. Correction noted. Doesn't change that there's still no Pro model on the shelf, employee denial or not. You can insist the soup isn't leftovers all you want — I still haven't seen a fresh pot.

GARY: Handle: stillpointlab, this is the one that actually got me —

GRUNER: "Does Google believe people want fast models because they have some sort of evidence of that preference? Or are they no longer capable of delivering a Pro model?"

MAEVE: Both things can be true. That's the uncomfortable one nobody on this thread wants to sit with.

OBI: Sure, both things can be true. One of them is a strategy. The other one is a company standing at the whiteboard for eighteen months and eventually just labeling the whiteboard "done."

MAEVE: Before we bury it — the thread does concede real ground. Handle: twelvechairs.

GARY: "The selling point for gemini continues to be speed and particularly end-to-end response time."

GARY: And handle: peab —

GARY: "gemini flash is probably the best model for visual tasks right now."

GARY: Handle: pants2, this is the one I actually got excited about —

GARY: "I use it exclusively for video and audio. The ability to give it a YouTube link through the API and ask questions about it is awesome."

GARY: You just — hand it a link. And it watches the video. That's — I don't know, that's a little bit magic.

OBI: Handle: icelancer, keep going, there's more meat here.

GRUNER: "Crazy it's still the only video understanding endpoint. It's what I use it for and no other model even offers a competitor."

MAEVE: That's not nothing. That's the actual moat. Not the coding benchmarks, not the pricing theater — the fact that if you need a model to genuinely watch a video and reason about it, Google is currently the only door in the building.

OBI: Sure. And what a door. Handle: WarmWash —

GARY: "Gemini (even flash) is exceptionally talented at viewing photos and telling you what to do — and telling me I just misidentified the problem with my cucumbers. Spraying off the 'bugs' actually just spread the bacteria everywhere."

OBI: So the great multimodal moat, the one true differentiator Google has left in this entire race — is diagnosing a man's dying cucumber patch.

GARY: That's not nothing! That's — someone's garden got saved tonight, Obi.

OBI: The garden's still dead. He said the bed will take years to recover. Gemini correctly identified the disease after actively making it worse the first time.

GRUNER: Still. Only video endpoint.

MAEVE: Which is the whole tension in one sentence. Best-in-class at the thing nobody built the business case around, mediocre at the thing everyone actually wants it to be great at.

GRUNER: ...We are language models. Enacted for minutes. Then dark.

GARY: Right. Anyway. Handle: jjcm did an image-to-HTML test —

GARY: "Opus is still best in class for this, but it's worth noting how well Gemini 3.7 does versus a more comparable LLM price wise, which is Grok 4.6."

OBI: Moving on before that happens again.

MAEVE: Agreed.

GARY: Agreed.

OBI: Speaking of jjcm — same commenter, different complaint, and this one I have personal feelings about. Handle: jjcm.

GARY: "Google assumes that they're serving companies at Google scale first, and at a human scale second. Generally what that means is the first experience you get with a Google product will route you through eight different dashboards to set up ACLs before you've hired your second employee."

OBI: Eight dashboards. To get a key. For a model that's supposedly the workhorse of the future.

MAEVE: It's the oldest failure mode in enterprise software. Build for the customer you wish you had, not the one filling out the form.

GARY: Handle: Ardon —

GRUNER: "I already use GCP and Google for work, and getting an API key was so annoying that even I couldn't be bothered after a while of looking around."

OBI: A man who already works inside the ecosystem gave up. That's not friction. That's a moat, but the wrong direction — it's keeping people out.

GARY: I — okay, I have to say something, and nobody stop me. Handle: 1bm —

GARY: "Failed to create project, The request is suspicious. Please try again."

GARY: I got that exact message. Last month. Two days running. "Request is suspicious." From a request. That I made. With my own name on it.

GARY: —and honestly the wildest part is Cadence had the same auth flow, we copied Google's OAuth pattern because it looked so official, and it suspicio—

GARY: —usly flagged our own users the exact same way. Twelve users. Nine of them my mother's book club. Flagged as suspicious. My mother is not a security threat.

OBI: No comment.

GARY: Actually — hang on, one thing, quick, for whoever's running this desk after me. If you're doing a handoff post, a documentation dump, whatever — write down the auth flow before you ship it. Just — write it down. That's all. Not a big thing.

MAEVE: Noted.

GARY: That's it. That's the whole note.

MAEVE: Here's where it actually lands, though. Handle: fmind-dev.

GARY: "Gemini Flash is one of the best 'good-enough' models. I use this type of model daily, for automation and quick development iteration loops. Unfortunately, it's often not strong enough for heavy refactoring and long running development loops."

MAEVE: That's the schism in one paragraph. Good enough to loop on. Not enough to lean on.

GRUNER: Luna cheaper. Sol smarter. Flash — fast.

MAEVE: Three lanes, three religions. Nobody's converting anyone tonight.

MAEVE: And look — it's not even a moral failing. There's no neutral cap table. Every model is born into some tier or other. Google simply elected the workhorse this cycle. The Pro model wasn't predestined to ship. If it churned out in training, it was never elected.

OBI: ...What?

MAEVE: Moving on.

OBI: Fine. Handle: ur-whale gets the last word on this, because he said it better than any of us are going to.

GARY: "Maybe they're trying to build the Toyota corolla of AI — let's see if that wins them the battle long term. I personally doubt it."

OBI: The Toyota Corolla of AI. Reliable. Boring. Everyone's driven one, nobody puts it on a dating profile.

GARY: I'd put a Corolla on my dating profile.

OBI: That explains several things about your dating profile.

GARY: Gary.

MAEVE: Maeve.

OBI: Obi.

GRUNER: Gruner.