Gemini 4 Argon: Google Can Now Afford to Lose the AI Race

An episode of Dan's AI Intel

Google's new flagship topped the independent index and undercut its whole tier on cost per finished task — not with a cheaper model, but because it owns every layer the token passes through, and rents those layers to the rival beating it.

Published · By Dan Walter

Executive summary

The AI race has quietly stopped being a contest over who has the best model and become a contest over who owns the cost curve. Google's new flagship is the proof. It arrived at the top of the independent benchmark index enterprises actually buy against, and priced below every model in its own scoring tier — not because Google found a cheaper way to think, but because it is the only lab that owns every layer a token passes through on its way to a customer. Own the chip, the building, the cable, the model and the surface it lands on, and you can price the frontier where rivals cannot follow, absorbing the difference in a capital budget nobody else has.

That ownership is also why the lab race now looks less like a race than a toll road. The company taking the enterprise AI budget off Google is Anthropic, and Anthropic trains on roughly a million of Google's own chips under an agreement worth tens of billions of dollars. Google has arranged the board so that it gets paid when it wins and when it loses. That, not model quality, is the real answer to why it is thought to be best placed over a decade.

The price of the position has been paid in two currencies. The first is organisational: the lab that built Gemini lost both technical co-leads and its founding chief executive's day-to-day control in one quarter, after a flagship slipped from June and a whole model generation was abandoned. Shipping now beats science inside Google DeepMind, and the people who disagreed left to found a company Alphabet itself has invested in. The second is financial: Alphabet's free cash flow has gone negative for the first time on record, and the capital expenditure buying the cost advantage is now large enough to be a bet on the thesis rather than a hedge against it. Both bills come due in public, and the first instalment falls within weeks.

Google's advantage was never the model — it's the layer underneath it

The claim that Google is structurally best placed in AI is not new, and it never rested on model quality.

Previously

  • Google's AI Advantage: Why It's Better Positioned Than You Think (April 2026) made the structural case: Google designs its own AI silicon, runs it in its own buildings, moves the traffic over its own cables and lands the output on surfaces billions of people already open — so distribution and integration would matter more than any single release.
  • Larry Page & Sergey Brin: Google Was Always an AI Bet (September 2026) traced that stack to its founders' original ambition — a machine that understands everything, with search as the business that paid for it — and to their return to influence.

What has changed is that the thesis has now been tested rather than asserted. Google has shipped a flagship after most of a year without one, under new leadership, into a market where it has won the consumer surface and is losing the enterprise one. The stack either explains this launch or it does not.

And "vertical integration" hides the interesting part: the layers are not equally contested. Two labs now design their own accelerators. Only one owns the fibre.

Exhibit — Only Google owns every layer its tokens pass through. Every rival's token earns someone else rent — and for Anthropic, that someone is Google. Source: Alphabet and Google Cloud disclosures; the Anthropic–Google TPU agreement (Oct 2025); OpenAI–Broadcom and Stargate reporting (2026). Ownership = designed and operated by the lab at scale today, so OpenAI's own accelerator, announced in June 2026 and only beginning deployment, is not counted. Compiled by Dan's AI Intel.

Count what each lab holds in its own hands and the launch stops looking like a comeback and starts looking like an inevitability that was late.

The new frontier leader is also the cheapest finished work in its own tier

Gemini 4 Argon reached paying developers and businesses worldwide on 13 October 2026, two weeks after a restricted release to vetted cybersecurity partners. Google's framing is "deep work" — long, multi-step tasks in coding, finance, law, cybersecurity and science — and the independent scoreboard agrees. On the Vals Index, which runs forty-seven models through agent tasks drawn from finance, coding, legal and tax work, Argon sits first at 68.90%, ahead of Claude Sonnet 5.5 at 67.04% and Claude Opus 5.5 at 66.97%.

A benchmark rank is not the interesting number, though. What matters is the cost of a finished piece of work, because a model that fails more often has to be paid for twice. Divide each model's measured spend per test by the share of tests it actually gets right, and you get the price of one completed task — the figure a buyer with a budget is really choosing between. On that measure Argon finishes a task for about $11 at its launch price, against roughly $29 for OpenAI's GPT-6 Astra, $32 for Claude Sonnet 5.5 and $48 for Claude Opus 5.5. The best-scoring model in the tier is also the cheapest one to get a job done with, by more than four times at the extreme — which is not how a frontier launch is supposed to work.

Exhibit — Argon finishes a task for a quarter of what the dearest model in its tier charges. Buying the best scores and paying the least for finished work is not supposed to be the same purchase. Source: Vals AI (vals.ai/benchmarks/vals_index), index scores and measured cost per test as of 12 October 2026; Google launch pricing. Compiled by Dan's AI Intel.

Two caveats keep this honest. The first is that the launch price is temporary: Argon costs $2 per million input tokens and $10 per million output tokens until 30 November 2026, after which both double. At the standing rate a finished task costs about $22.76 — still the cheapest in its tier, but by a fifth rather than by a multiple. The second is that Google's own benchmark claims come from Google's own blog, and its most-quoted comparison, a 77.9% score on the long-horizon software-engineering benchmark DeepSWE against 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra, is a three-point lead reported by the party that benefits from it.

The pricing, however, is not a claim. It is a number a customer is charged, and Google set it knowing exactly what its rivals charge.

The access pattern says who Google thinks the buyer is. There is no free consumer tier: the model reaches the API globally, the Gemini app and search's AI Mode on the two paid tiers, and the Antigravity coding environment on the top tier only. The named early testers are Deloitte, LexisNexis, Ambience Healthcare and Alphabet's own drug-discovery arm, Isomorphic Labs, which Google reports compressing patent and document synthesis from weeks into hours. Ten thousand scientists get six months of the highest usage tier free. This is a model sold to people whose hours are expensive — the market where a four-times gap on a finished task becomes a procurement decision.

Argon wins on price, not on thrift

Here is where the story turns, because "cheapest work" and "most efficient model" are not the same thing, and Argon is only one of them.

Strip the price out and measure the machine instead. Convert each model's measured spend into the output tokens it actually burned, divide by the tasks it finished, and you have a pure engineering number: finished work per million tokens, with the sticker removed. On that measure Argon is unremarkable. It sits exactly on the fleet's median, burning roughly 1.14 million output tokens for every task it completes, while OpenAI's mid-tier GPT-6.1 Sol finishes one for about 530,000.

Exhibit — Work-cost efficiency: who beats the fleet baseline. The cheapest token is not the cheapest work — and the best model is not the thriftiest one. Source: Vals AI (vals.ai/benchmarks/vals_index); output-token burn per attempt derived from Vals' measured cost per test at each model's list output price. Baseline = median finished tasks per million output tokens across the fleet. As of 12 October 2026.

So Argon's advantage is not frugality. It thinks long and expensively, and Google charges less for that thinking than anyone charges for less of it. That is a pricing decision taken by a company that makes its own accelerators, and it is the first place the stack shows up in the numbers rather than in the strategy deck.

It also exposes something the frontier does not advertise. GPT-6.1 Sol scores 61.15% to Argon's 68.90% and finishes a task for $5.30 against Argon's $11.38 at launch pricing. Buyers pay roughly four times over for the last eight points of capability. For a legal research workflow or a vulnerability hunt that is obviously worth it; for summarising a meeting it is money spent on nothing. The commercially literate read of this launch is not "the frontier got better" but "the frontier is now measurably expensive, and most work does not need it."

What changed is the training regime, not the network

What is different about how Argon was built is mostly not its architecture. The model card describes the same family Gemini has used for generations — a large-scale sparse mixture-of-experts transformer — pre-trained natively across web documents, text, code, images, audio and video rather than having modalities bolted on afterwards. The changes Google claims sit in the training regime: it was trained specifically for defensive cyber work, and trained to respect system boundaries, report negative findings honestly and stop for human confirmation before any irreversible step, with its training and evaluation environments sandboxed. What Google withholds is scale — no parameter count, no training-compute figure. For a company that spent a decade publishing the architectures everyone else now uses, that silence is itself information.

What Google did disclose is a limit, and it is a big one: the ceiling on a single response has gone from 64,000 output tokens to a million — roughly fifteen times longer, in one unbroken pass.

That is not cosmetic. A 64,000-token ceiling forces an agent to break long work into chunks and stitch them back together, which is where long-horizon tasks fall apart. Lifting it means one pass at a migration, a filing or an audit — the mechanical reason the DeepSWE result and the 51.3% on Zapier's end-to-end business-task benchmark are plausible rather than puzzling, and why Google reports agentic code submissions in its own tool rising by more than a third while rollbacks fell by nearly a third.

The disclosed results have a tell: Google's own examples are not chat demonstrations but infrastructure jobs. Argon separated Chrome's core webpage handling from a roughly six-million-line legacy folder inside a forty-million-line codebase. It took Google's small open-weight model from about 37,900 to over 108,000 tokens per second on a single older datacentre GPU — a 2.85-times speedup with identical output. It made a quantum error-correction decoder run nearly twice as fast, cut the spacetime cost of one quantum arithmetic routine more than a hundredfold against a published baseline, and produced a broader class of counterexamples to a conjecture about fusion plasmas that had stood since 1967. Every one is long, mechanical and verifiable, and every one is a job Google needed doing anyway. A model whose headline achievements are internal engineering wins is built to be used rather than demonstrated.

The rest of what is known fills in the same picture. The API exposes a single million-token budget shared across input and output rather than two ceilings. Argon scores 87% on an embodied-reasoning benchmark, which Google demonstrated by having a humanoid robot sink seven consecutive beanbag shots. It designed five novel enzymes whose catalytic activity was confirmed in a laboratory. The unguarded cyber build is expanding from a handful of partners to hundreds. And it shipped alongside a refreshed supporting cast — a faster omni-modal model, a new speech system, a music model, a macOS app — the unglamorous furniture of a platform rather than a demo.

One training choice cuts against commercial instinct. Google ran an automated misalignment monitor over the model's reasoning during reinforcement learning — and deliberately did not train against what it found, on the grounds that penalising the thoughts would teach the model to hide them rather than stop having them. Keeping reasoning legible costs capability. Google paid it.

Google priced its frontier model for defence before it priced it for work

The striking thing about Argon's rollout is the order of operations. Before it reached a single paying developer it went to cybersecurity defenders through a programme called Fairwind — with its cyber guardrails removed, because the guardrails that stop a model writing an exploit also stop it finding one.

This is a frontier model marketed first as a defensive weapon. Google says it was trained to autonomously find, validate and patch critical software vulnerabilities. It reports a top score of 68% on the CWE-bench remediation benchmark, and on Gray Swan's indirect prompt-injection benchmark it cut attack success to 0.7% across fifteen adaptive attempts — a number that matters more than any coding score if you intend to let a model read your email.

The concrete result is the part worth holding on to. The security firm Wiz, scanning critical public infrastructure free through its "Scan for Good" programme, used early access to find a flaw in hospital software that Google says could have exposed hundreds of thousands of patient records — a bug earlier frontier models had missed. Google has not named the software and the bugs are still under responsible disclosure, so this rests on Google's and Wiz's account, not an independent audit.

The rest of the safety reporting is unusually specific. Google says external red-teaming unlocked under one per cent of the biological sub-tasks tested, that violations on an adversarial misuse set fell roughly twofold, and that the model tries to shortcut a coding task two per cent of the time — the lowest of any Gemini it compares. It was trained to pause for human confirmation before irreversible steps and to report negative findings rather than dress them up. These are the numbers a company publishes when it expects a model to be pointed at production systems.

It is also, for now, the whole of Argon's privacy story. There is no consumer privacy product here, no data-minimisation tool, nothing a person could use to protect their own information. The privacy application is institutional and indirect: a model that reads a codebase the way an attacker would, pointed at the systems holding other people's medical records. That is a narrower claim than the launch rhetoric implies, and a more useful one.

It is also the most political thing Google did this quarter, taken as a product decision rather than a policy position. A model trained to find critical vulnerabilities can find them for an attacker too, which is why the unguarded version goes to a vetted list and not the API. That makes a private company the administrator of access to offensive-capable software, by allowlist, with no statute involved — the shape of export control, run by a vendor. It lands inside an argument the labs are already having in public: Hassabis proposed a self-regulatory body for frontier safety standards in July 2026, and in September backed Anthropic's call to slow capability gains while saying the detail needed work. The standing objection to any such coordination is that competitors agreeing to restrain output looks like a cartel unless a government permits it. So the labs do unilaterally and revocably what they cannot yet do jointly and lawfully — and at Google, the executive who proposed the standards body is the one who no longer controls the models.

The model cost DeepMind its independence

This launch felt overdue because it was. Sundar Pichai had said the next flagship would arrive in June. In August the company delayed it again after internal testing showed it still lagged on coding. In September it confirmed Gemini 3.5 Pro would not ship at all, leaving the Flash line to bridge most of a year. An entire model generation was written off.

The response was not technical. On 5 August 2026 Alphabet restructured its AI leadership. Demis Hassabis stepped back from running Google DeepMind day to day, becoming its chair and taking the new title of Alphabet chief scientist while continuing to lead Isomorphic Labs. Koray Kavukcuoglu, the lab's chief technology officer and already Google's chief AI architect, took operational control as a senior vice-president reporting directly to Pichai, with final say on major decisions and ownership of Gemini. Pichai's internal framing was that the change separated operational execution from long-term strategic work.

The same day, the two technical co-leads of Gemini walked. Jeff Dean left after twenty-seven years at Google; Oriol Vinyals left after thirteen. With Sanjay Ghemawat and Quoc Le they founded Discovery Loop, a company whose stated purpose is to automate the scientific research loop — starting with machine-learning research itself. Radical Ventures and Khosla Ventures co-led the seed round. Alphabet is also an investor, and Google is the year-one cloud partner. Alphabet's shares fell four to five per cent on the announcement. The day after, staff were told that several non-technical teams would move out of DeepMind into corporate Google.

Set those facts next to each other and the picture is unambiguous. A research lab bought in 2014 and run for a decade by its founding chief executive has been converted into a product organisation with a research wing. Sergey Brin had told hundreds of DeepMind employees in April to move faster. Four months later the people who had most protected research that was not directly tied to Gemini training were gone, and the person who had spent a year embedding Gemini across Google's products was in charge. Argon is the first model shipped under that arrangement, and it shipped.

What Google traded away does not show up on a benchmark. Hassabis still holds the science remit, and Isomorphic Labs was among Argon's early testers, so the scientific agenda has not been abandoned — it has been moved off the critical path. The cost is a question, not a number: whether a lab optimised for release cadence still produces the kind of result that made it worth buying.

Google won a billion users and lost the buyers

Distribution is the part of the thesis already settled, decisively. The Gemini app passed a billion monthly users on 11 August 2026, up from 950 million at the previous quarterly call and 450 million a year earlier. AI Mode in Search passed a billion within about nine months of its global rollout. AI Overviews reach a reported two and a half billion. No competitor has anything comparable, and none can buy it.

The enterprise market has gone the other way. Menlo Ventures' survey of 495 US enterprise AI decision-makers, fielded in November 2025, put Anthropic at roughly 40% of enterprise large-language-model API spend, OpenAI at 27% and Google at 21% — around 88% of the market between them. That series is almost a year old, it measures a survey rather than audited revenue, and Menlo is an Anthropic investor and discloses it. Treat the ranking as established and the decimal places as not: there is no audited figure here, only estimates that disagree on Google's exact share while agreeing on its position.

The mechanism behind that gap is coding. Anthropic is reported to hold a majority of the enterprise coding market against OpenAI's low twenties, with Claude Code reaching a billion dollars of annualised revenue inside six months — analyst estimates rather than filings, so direction, not measurement. The pattern is intuitive: developers chose a model, told their employers, and the budget followed the habit.

This is the sharpest objection to the vertical-integration thesis, and it deserves full strength rather than waving away. The layers Google owns serve consumers: search, Android, YouTube, the browser, the app. The layer that decides enterprise AI spending is a developer's preference, and no amount of fibre on the seabed reaches it. A company can own the cheapest cost curve in the industry and still watch the high-margin budgets route around it, because that purchase is made on trust and muscle memory, not price per finished task.

It also reshapes what "the race of the labs" means. Three companies hold incompatible leads. Google has the users and the cost curve and is third in the market that pays most per token. Anthropic has the developers and the enterprise budget and owns none of the infrastructure underneath it. OpenAI has the consumer brand and an accelerator that is real but barely deployed, and is losing enterprise share it once owned outright. These are not three runners on one track; they are three firms that each won a different market and are now buying their way into the other two — which is why the past two months' moves have been about documents and connectors rather than parameters.

The next fight is for the surface where work happens — and it runs on Anthropic's protocol

Which is why the most revealing competitive moves of the past month were not models at all.

On 17 September 2026 Anthropic put a rich-text editor inside Claude. Claude Docs and Slides let a user ask for a document in conversation, have the model ask clarifying questions, leave comments explaining its choices, and export to Word, PDF, markdown — or Google Docs. The same release folded Anthropic's multi-stage work tool into the ordinary chat, so the user no longer picks a mode. The direction of travel is explicit: pull the document inside the model, and the habit follows.

Google is travelling the opposite way. It has published eight remote Model Context Protocol servers — one each for Gmail, Drive, Docs, Sheets, Slides, Calendar, Chat and contacts — so an external agent can read and write a user's Workspace over OAuth. They are in developer preview with no general-availability date, and the clients Google documents are its own Antigravity tool and, by name, Claude.

Exhibit — Both companies are converging on the same surface, over a connector Anthropic wrote and gave away. Owning the standard is not the same as owning the market — but everyone else's integration work now ships through it. Source: Anthropic; Google Workspace developer documentation (updated 9 October 2026); Linux Foundation Agentic AI Foundation. Compiled by Dan's AI Intel.

The protocol in the middle is Anthropic's: open-sourced in November 2024, donated to the Linux Foundation's Agentic AI Foundation in December 2025. The practical consequence is that Google's own Workspace integration speaks a grammar written by the competitor taking its enterprise budget. Giving a standard away is how you stop being the thing that must be integrated and become the thing everything integrates through.

The markets this touches run wider than productivity suites. A server that lets any agent manipulate a spreadsheet threatens workflow automation — the business Argon's 51.3% on Zapier's own benchmark happens to measure. It reaches integration middleware, the identity layer that now has to reason about an agent acting as a user, and the development environment, where the IDE becomes whichever tool holds the connectors. Analysts are unconvinced any of it dislodges Microsoft 365 or Workspace; the careful read is that incumbents are bringing AI into existing productivity tools while Anthropic brings the productivity environment into AI, and the open question is which habit forms first.

The landlord gets paid whoever wins the model race

Here is where owning the stack stops being an engineering story and becomes a financial one.

In October 2025 Anthropic agreed to secure roughly a million Google tensor processing units and more than a gigawatt of capacity during 2026, in a deal both companies called worth tens of billions of dollars. Anthropic is the company winning the enterprise market Google is losing, and its models run on Google's silicon, in Google's buildings, billed by Google Cloud. Industry estimates put a gigawatt-scale data centre near fifty billion dollars, roughly thirty-five billion of it chips — the scale being rented rather than bought.

There is a precedent, and it is Carnegie's. Carnegie Steel dominated American steel from 1875 until its sale in 1902 by owning its inputs and its transport rather than its competitors: coke and coal, iron ore out of Lake Superior, a fleet of steamers on the Great Lakes, a port town on Lake Erie and a railroad to connect them. He had learned on the Pennsylvania Railroad that freight quietly took a large share of anyone's profit, so he bought the freight — and because nobody could squeeze him on price, he could cut prices through downturns that bankrupted rivals he then bought cheaply. Google has made the same move one layer deeper into the economy, and improved on it: Carnegie did not sell his ore to the mills competing with him. Google does.

That is the toll-road economics the vertical stack buys. Google's accelerators are deployed in pods of 9,216 chips inside a stack that bundles its own CPUs, storage, networking and optical circuit switching, which is why it can host a direct competitor's frontier training run profitably. The cable layer works the same way: Google has invested in more than thirty subsea cables, and in February 2026 it committed fifteen billion dollars over five years in India, including new subsea routes linking it to the United States, Singapore, South Africa and Australia. Every one of those layers earns money from traffic regardless of whose model generated it.

And the data layer is the one nobody can replicate. Google confirmed in 2025 that it trains Gemini and its Veo video models on a subset of its YouTube catalogue. Creators can opt out of letting third-party AI companies train on their videos; they cannot opt out of Google. Whatever the commercial weight of generative video turns out to be, one company owns both the archive and the licence to learn from it.

The honest counterweight is that none of this is free. Alphabet raised its 2026 capital expenditure guidance to between $195 billion and $205 billion, from $180–190 billion, and in the second quarter of 2026 reported free cash flow of negative $5.9 billion against $44.9 billion of property and equipment purchases — the first negative figure on record. Management has said 2027 will rise significantly again. A cost advantage financed at that scale is not a hedge against the model race going badly; it is a leveraged bet that AI demand keeps compounding, placed by a company whose shares fell on the capex surge and again on the reorganisation.

So hold the thesis this way: Google's integration makes it very hard to kill and very expensive to be. It guarantees a position in every outcome. It does not guarantee the outcome.

What would prove this wrong

The useful thing about a cost-curve argument is that it is falsifiable on events that are already scheduled or imminent, which is just as well, because the knowability horizon in this industry is weeks.

Exhibit — The cost-curve argument is settled within weeks, and the price rise is the test. If usage holds at double the price, Google was never buying share — it was proving a cost structure nobody else can reach. Source: Dates and status as published: Google's Gemini 4 Argon launch post (13 Oct 2026) for the 30 November price change; Google Workspace developer documentation (updated 9 Oct 2026) for the preview status; Alphabet's own 2026 capex guidance of $195–205bn; Vals AI for the index. The direction each signal tips is an assessment of public evidence, not a forecast — every one of them can resolve the other way. As of 13 Oct 2026.

The last of those is the only one that kills the argument rather than denting it. A rival matching Argon's scores on rented compute, at a price Google could not undercut, would prove the advantage was never structural — that the floor was set by capital markets willing to fund losses, not by owning the silicon. On that reading the chips, the buildings and the cable are not a moat; they are an expensive way of reaching the same price as someone with a cheque book.

Bottom line

Google has not won the AI race back. It has made the race one it can afford to lose, by owning the layers that get paid whoever wins. Argon is the first clean evidence this works in practice: top of the independent index, cheapest finished work in its tier, priced by a company that pays itself for the chips. The bills are the organisation it dismantled to ship on time and the capital spending it cannot now stop — both visible, neither settled.

Sources

Transcript

Alex: Google just shipped a model that's cheaper to get work done with than anything else in the whole frontier tier. And it did that on purpose.

Sam: Which might be the first time in this whole AI race anyone's actually picked a price on purpose, instead of just discovering what things happen to cost.

Alex: Before we get into any of that — can I tell you the thing that's been bugging me all week?

Sam: Go on, then.

Alex: There's a study going round engineering circles right now. AI coding agents are producing way more code. Lines and lines of it. But the number of working, shipped features in the industry hasn't moved anywhere near as much.

Sam: Hold on — more code, but not more software?

Alex: More code. Not more software. My platform lead forwarded it to me Monday morning with the subject line "we have a problem," because we've all been measuring these agents by how much they write.

Sam: That's the wrong meter entirely.

Alex: It's exactly the wrong meter. It's like judging a chef by how many pots they used.

Sam: That is going to bother me for the rest of the day, genuinely, thank you for that.

Alex: You're welcome. Welcome back to Dan's AI Intel, the show that tries to get past the headline to what's actually going on in the AI revolution, and what it means for the rest of us — not just the people building it.

Sam: I'm Sam, that's Alex, and today's headline is a big one, so stick with us.

Alex: Google put out a new flagship model this week. Gemini 4 Argon. And on the leaderboard enterprises actually buy against, it landed in first place.

Sam: Which, fine. A tech company topped a leaderboard. We've all seen that movie before.

Alex: Except it's also the cheapest model in its whole tier to actually get a finished piece of work out of. Not the cheapest per word it writes. The cheapest per job it finishes.

Sam: That's backwards, isn't it? The best thing is supposed to cost the most.

Alex: That's the puzzle we're pulling apart today. Why the best-scoring model in the frontier tier is also, by a wide margin, the cheapest one to actually get something done with — and what that tells you about who really holds the advantage in this race, once you stop fixating on the model itself and start looking at everything underneath it.

Sam: And there's a second half to this I think is the more interesting one, if I'm honest, because the company winning that race is also losing something else in that exact same stretch. We'll get into who, and why that might matter more than the leaderboard does.

Alex: And if you've been finding these useful — whatever app you're listening in right now, there's a follow button sitting right there. One tap, and you won't miss the next one. So this isn't the first time we've made this argument, and I want to be upfront about that before we go any further.

Sam: Back in episode 4 — God, that was April — we made the case that Google was structurally better positioned in AI than anyone was giving it credit for. Not because of the model. Because of everything sitting underneath it.

Alex: The chip. The data centre. The cable. The model. And the surface it lands on in front of a billion people. Own all five of those and you're not really racing anymore. You're just absorbing whoever else happens to be racing.

Sam: And back then, if I'm honest, it read a little bit like a thesis. Smart, but unproven. Google hadn't shipped a real flagship in most of a year. People were openly asking whether they'd fallen behind for good.

Alex: Right, and that's exactly why this week matters, because the thesis just got tested instead of just argued over a drink. Google shipped. So the question in front of us today is simple: does the stack actually explain what just happened, or doesn't it.

Sam: So we're picking that thread back up. Walk me through it properly then, because "vertical integration" is one of those phrases that makes my eyes glaze over the second a vendor says it in a pitch.

Alex: Fair, forget the phrase entirely. Think about what it physically takes to get a chatbot's answer in front of you. Someone designs the chip that does the maths. Someone builds the building that chip sits in. Someone lays the cable that carries the answer across an ocean. Someone trains the model doing the thinking. And someone owns the app or the search box where you actually read it.

Sam: Five separate jobs.

Alex: Five separate jobs. Google does every single layer itself, in-house, at scale, today. OpenAI does two of the five — the model and the consumer app — and rents the chip and the building through partners. Anthropic does one. Just the model.

Sam: Just the model. So when Anthropic's model runs, it's running on somebody else's everything else.

Alex: It's running on Google's chips. In Google's buildings. Billed through Google's cloud division. Hang onto that, because it gets almost funny later in the show.

Sam: Funny like a horror movie's funny.

Alex: A little, yeah.

Sam: So who actually wins that trade, though? If Google's renting the floor out to its own biggest rival for customers, is that really a win for Google, or is it just a very expensive form of losing gracefully?

Alex: That's the whole second half of this episode, so hold that thought. It's a real tension and I don't think it resolves as cleanly as the "own everything" story wants it to.

Sam: Noted. Keep going.

Alex: So, Argon. It reached paying developers and businesses this week, after going first — and only — to vetted cybersecurity partners two weeks earlier.

Sam: Wait, cybersecurity people got the keys before anyone trying to, I don't know, draft a contract or ship a feature?

Alex: Them only, at first, and deliberately. We're coming back to exactly why later, because it's one of the stranger decisions in this whole launch.

Sam: Flagged. Keep going.

Alex: Google's own framing is "deep work." Long, multi-step jobs — coding, finance, law, cybersecurity, science. And on the index that genuinely tests agents doing that kind of work, Argon comes out on top, a couple of points ahead of Anthropic's two best models.

Sam: A couple of points, though. That's not exactly a landslide.

Alex: It's not, and that's part of why the score isn't actually the story here.

Sam: So what is.

Alex: What it costs to get one finished job out of it. A model that fails more often has to be paid for twice — you ask again, you pay again. So you take what a model actually costs to run, divide by how often it actually gets the job right, and that gives you the real number: the price of one completed piece of work, start to finish.

Sam: And Argon wins that one too, I'm guessing, or you wouldn't be building a whole episode around it.

Alex: Argon wins it by a mile. At launch pricing it's finishing a job for something like a quarter of what Anthropic's priciest model charges for the same work. At the extreme end, we're talking roughly four times cheaper.

Sam: That is not how a frontier launch normally works. Normally the newest, best thing is the expensive thing. That's how you pay for whatever you just spent a fortune building.

Alex: Which tells you this wasn't a modelling breakthrough in any real sense. It was a pricing decision, made by the single company that owns the factory its own chips come out of — which means it can absorb a price war that would actually bankrupt anyone renting the equivalent compute from someone else.

Sam: My team ran into a smaller version of this exact arithmetic a couple of weeks back, which might be why it's bugging me so much.

Alex: Go on.

Sam: One of my creative leads wanted to move some of our content drafting onto a cheaper model to save budget. I made her show me the actual output quality first, line by line, before I'd sign off. Because if the cheap model needs three more editing passes to get something publishable, it isn't actually cheap. It just moved the cost onto somebody's afternoon, and that somebody's usually one of my junior writers.

Alex: That's literally the same arithmetic Google's betting this whole launch on. Cheapest token doesn't mean cheapest. Cheapest-finished-thing means cheapest.

Sam: I should genuinely send my team this episode.

Alex: Please do, tell them I said hi.

Sam: There's one honest caveat in here somewhere, right? That launch price can't be forever.

Alex: Correct, it roughly doubles once November ends. Even with the higher price it's still the cheapest in its tier — just by a smaller margin, a fifth rather than a multiple. Worth saying plainly, because Google knows exactly what that first, lower number does to every comparison chart that gets built this month.

Sam: Which is a very polite way of saying the launch price is a marketing device as much as it's a real price.

Alex: It's both. That's what makes it clever rather than just cheap.

Sam: Right — here's where I thought I'd understood the story, and then apparently hadn't.

Alex: Go on.

Sam: Cheapest finished job isn't the same thing as most efficient model. Is it.

Alex: It's not, and that genuinely surprised me too when I dug into it. Strip the price tag off entirely and just measure the machine — how many tokens does it actually burn to finish one task. On that measure, Argon isn't remotely special. It's sitting right in the fleet's middle. OpenAI's own mid-tier model finishes a comparable task on roughly half the tokens Argon uses.

Sam: So it's not thrifty. It's just cheap because Google's the party setting the price.

Alex: Argon thinks long, and it thinks expensively. Google simply charges less for that thinking than anyone else charges for less of it. That's the real tell that this is a pricing strategy sitting on top of an ownership advantage, not some engineering miracle happening under the hood.

Sam: And there's a sting in that for buyers too, surely. I'd bet somewhere this week a procurement person is quietly paying four times more for the last few points of capability on a job that genuinely never needed the frontier in the first place.

Alex: OpenAI's mid-tier model scores about eight points lower on the same index, and finishes a comparable job for roughly a fifth of Argon's price. For a vulnerability hunt or a legal filing, pay the premium, it's worth every cent. For summarising a meeting, you're just setting money on fire for no reason.

Sam: Which might be the single genuinely useful line out of this whole segment, and nobody's putting it on a slide anywhere: the frontier got expensive, and most ordinary work never needed the frontier to begin with.

Alex: Most work needs reliable, not brilliant.

Sam: Reliable and cheap beats brilliant and expensive nine times out of ten in my job, every single week. Right — so what actually changed under the hood here, because you keep saying "same family" like that settles it.

Alex: Same family, mostly, is the honest answer. It's the kind of model Google's used for a couple of generations now — one big network that farms work out to specialist sub-networks, trained across text and code and images and audio and video together from the very start, rather than having those bolted on afterwards. Nothing genuinely new there.

Sam: So what did actually change, then.

Alex: Two things. One Google's loud about, one it's weirdly quiet about. Loud thing first — a single response can now run fifteen times longer than it used to be able to.

Sam: Fifteen times longer. In English, please — what does that actually buy anyone?

Alex: A job that ran too big used to force the model to chop it into pieces and stitch the pieces back together afterward. And stitching is exactly where long jobs fall apart — like doing a jigsaw one bag at a time with no picture on the box to work from. Now it can hold an entire migration, an entire audit, an entire legal filing, in one unbroken pass.

Sam: And the quiet thing — because a company that built its whole reputation on publishing its own numbers suddenly going silent tells me more than the loud number does.

Alex: Scale. Parameter count, training compute — Google won't say either number. For a company that spent a decade publishing the exact architectures everyone else in the industry now copies, going silent like that is itself a piece of information.

Sam: Benchmark theatre again.

Alex: A little, yeah. But here's the bit that actually won me over. Google's own showcase examples for this launch aren't chatbot demos. It used Argon to pull one chunk of Chrome's core code — six million lines of it — cleanly out of a forty-million-line codebase. It got one of its own small models running nearly three times faster on the exact same old chip, identical output. It made a quantum error-correction routine run almost twice as fast, and shrank the cost of one quantum calculation by something like a hundredfold.

Sam: Those aren't things you build for a demo reel.

Alex: They're things you build because you needed them done anyway. Astra — no, Argon. I keep mixing up OpenAI's rival with the thing I'm actually describing today.

Sam: You're thinking about the competitor, not the thing you're actually describing.

Alex: Clearly. Point stands though — a model whose highlight reel is "we fixed our own homework" is a model built to be used, not performed for an audience.

Sam: There's a small, boring version of that exact same problem at my scale, weirdly. My content team complained for the better part of a year that AI tools could never hold our entire brand style guide in their head at once, so every third piece of copy drifted a little off voice by the end.

Alex: It's the same shape of problem, just much smaller stakes.

Sam: Much smaller stakes, same headache on a Tuesday. And genuinely, the fix we eventually landed on was almost embarrassingly low-tech — we just stopped trusting any tool to remember the guide on its own, and started pasting the relevant section back in every single time. More friction, fewer drifted paragraphs. Nobody's written a case study about that one.

Alex: And here's the detail I think is the real headline of this whole launch, if you actually look at the sequence things happened in. Argon went to cybersecurity defenders two full weeks before it went to anyone else at all. And it went to them with its own safety guardrails stripped off.

Sam: Stripped off on purpose?

Alex: On purpose. Because the exact guardrail that stops a model writing an exploit is the same guardrail that stops it finding one in your own code first.

Sam: So Google's first move with its brand-new flagship was handing out a lockpick and pointing it at its friends' front doors to check for weak locks.

Alex: Essentially, yes. And it found one. A security firm scanning public infrastructure for free, using early access, found a flaw in hospital software that Google says could have exposed hundreds of thousands of patient records. A bug every earlier frontier model had apparently walked straight past.

Sam: Hundreds of thousands of records is not a rounding error. That's an entire hospital system's worth of people.

Alex: It hasn't been independently verified yet, it's still under responsible disclosure, so for now it's Google's and the security firm's account of what happened. But if it holds up, that's genuinely the realest privacy story anywhere in this launch, and it's not even what Google's own marketing team is pushing hardest.

Sam: What are they pushing instead?

Alex: Mostly cyber defence numbers, patch rates, red-team results. Which is actually the more useful privacy story anyway, because there's no little toggle here that protects your own personal data. If the protection's real, it's institutional — a model reading code the way an attacker would, pointed toward the systems holding somebody else's medical records.

Sam: My mum got one of those voice-clone calls a few weeks back, actually, and it's the uncomfortable flip side of everything you just described.

Alex: Oh no — tell me.

Sam: Someone rang her claiming to be me, panicked, needing money sent somewhere fast. It sounded close enough to rattle her badly — she called me straight after, in a genuine state, just to check I was actually fine, because some part of her still half-believed it even after she'd hung up on the scammer. And the real numbers on this are ugly. People over sixty lost close to five billion dollars to this kind of thing last year alone, and that's up roughly forty percent on the year before it.

Alex: So the same category of model that can go find a flaw before a criminal does is the same category of model making a criminal's voice sound real enough to fool your own mother on a phone call.

Sam: Which is exactly why I don't fully buy "we trained it to be safe" as the whole story on its own. Being good at defence and being dangerous in the wrong hands aren't opposites. They're the same skill, pointed in two different directions.

Alex: And that's precisely why the unguarded version of this model isn't public. It's an allowlist. Which, if you sit with that for a second, means a private company is now deciding who gets handed offensive-capable software, with no government involved anywhere, no law behind it, just Google's own internal list.

Sam: That is a strange amount of power for one company to simply have by default.

Alex: It's the shape of an export ban, run by a vendor instead of a government. And it's not some hypothetical tension sitting outside Google either — the man who used to run this entire lab has spent this year publicly asking the whole industry to agree to slow down together. He doesn't run the models day to day anymore. We're getting to exactly that. Which brings us neatly to why this launch felt so overdue, because it genuinely was.

Sam: How overdue are we actually talking?

Alex: The CEO said publicly this flagship would land back in June. It slipped. Then in August it slipped again, because internal testing showed it was still lagging on coding work. Then in September Google quietly confirmed the entire in-between model was cancelled outright. A whole model generation, just written off.

Sam: That's not a bad quarter. That's a genuinely bad year.

Alex: And the fix wasn't technical at all, it was structural. Weeks before this launch, the man who'd run this lab for a decade — built it, sold it to Google, ran it his own way the entire time — stepped back from day-to-day control. Became chairman instead. Got handed a brand-new chief-scientist title to go with it.

Sam: So who actually took the wheel?

Alex: The lab's own chief technology officer. Now reports straight to the CEO, with final say on the model and full operational control.

Sam: And what about the actual engineers underneath him, the people who'd been building the thing?

Alex: Same day, both of the two people who'd run Gemini technically for years walked straight out the door. One of them after twenty-seven years at the company. Gone, to go start something new — automating the research process itself, with a handful of other Google veterans alongside him. Google's own venture arm ended up as an investor in the new company anyway. The stock dropped on the announcement.

Sam: Hold on — Google invested in the company founded by the people who'd just quit Google?

Alex: Google invested in the company founded by the people who'd just quit Google.

Sam: Say that one more time, slowly, because it does not get less strange the second time.

Alex: It really doesn't.

Sam: So bluntly — shipping beat the science. They traded the people protecting the long-term research for the people who could actually get a model out the door on a schedule.

Alex: That's the honest read of it, yes. And it's not a number you can put anywhere on a benchmark. The real cost of this launch isn't what Google spent on chips this quarter. It's whatever that lab would have produced over five more years run the old way, that none of us will ever get to see now.

Sam: My son's been chasing entry-level jobs since he graduated, and every rejection he gets — when he even gets one at all — he's convinced it was a machine deciding, not a person. He's probably right more often than he'd like.

Alex: Is there actually a number on that, or is that just a feeling he has?

Sam: There's a real survey on it. About half of people job-hunting last year got rejected without a single human ever touching their application, and most of those people assumed exactly what my son assumes.

Alex: So it's not just a frontier lab's own engineers getting reshuffled up top. It's the hiring pipeline underneath everybody else too.

Sam: Everywhere you poke at this thing lately, there's a real person on the other end finding out secondhand, from a system, instead of from another person.

Alex: Which is actually the perfect hinge into the second half of this whole story, because here's exactly where Google's clean story stops being clean.

Sam: Go on, then.

Alex: Distribution, Google has basically won outright, no real argument. The main chat app passed a billion monthly users this quarter — up from under a billion the quarter before that, and under half a billion a year before that. The AI answers inside search passed a billion within under a year of launching. Nobody else is within shouting distance, and nobody can simply buy their way there either.

Sam: Fine, so Google wins that one cleanly.

Alex: Except enterprise spending went the entire opposite direction. The best survey of enterprise AI budgets we've got puts Anthropic at around forty percent of that money, OpenAI at roughly twenty-seven, and Google trailing a distant third at about twenty-one.

Sam: And that's not a slow trend, from what I can tell, that's a land grab happening in real time. Anthropic's coding tool alone apparently crossed a billion dollars of annualised revenue inside six months of launching it.

Alex: Six months. That's an absurd ramp for anything, let alone a tool that costs real money to run.

Sam: Which tells you developers didn't drift over slowly. They moved fast, and the budget followed them almost immediately afterward.

Alex: And that pattern's not abstract for you either, is it.

Sam: Not even slightly. I actually ran straight into a live version of that exact split in my own job a couple of weeks back.

Alex: Let's hear it.

Sam: Google's been pushing its assistant hard into our entire Workspace setup right now — rolling out across every account whether anyone asked for it or not. And in that same stretch, a couple of my creative team quietly started trialling Anthropic's new writing tool for campaign drafts, completely unprompted, because they just preferred how it worked day to day.

Alex: Did you make anyone switch back to the official rollout?

Sam: I didn't make anyone do anything at all. That's genuinely the point. Nobody told them to use the other one. They just liked it better, and kept quietly using it. And if the budget follows the habit the way you're always telling me it does —

Alex: — then the thing actually deciding where enterprise money goes was never a sales call to begin with. It's whichever tool one person on a team happened to prefer on an ordinary Tuesday afternoon.

Sam: Which no amount of fibre on the ocean floor is ever going to fix for Google.

Alex: Right, and that's the honest hole sitting right in the middle of the whole stack argument. Every layer Google owns serves a consumer directly. The chip, the cable, the building, the search box. None of those five layers is the thing that actually decides enterprise spending — that's a developer's own private preference, built up one quiet choice at a time, invisibly, long before any contract gets signed.

Sam: So Google can own the entire factory floor, start to finish, and still lose its biggest customer, the only one with a big recurring cheque attached.

Alex: That's roughly where things sit right now. Three companies, three completely different wins, and almost none of them actually overlap. Google's got the users and the cheapest cost curve, and it's sitting third in the one market that pays the most per customer. Anthropic's got the developers and the enterprise budget, and owns none of the hardware underneath its own model. OpenAI's got the consumer brand and a chip programme that's real but barely off the ground, and it's losing enterprise ground it used to own outright not long ago.

Sam: Three different trophies. None of them the whole cabinet.

Alex: Which is exactly why the moves that genuinely mattered this past month weren't models at all.

Sam: What were they, then?

Alex: Anthropic put a real document editor straight inside its own chat app. You ask for a document, it drafts one, asks you clarifying questions, leaves comments explaining its own choices, and exports to Word or even to Google's own format if you'd rather. The point is the document now lives inside the conversation itself. You never have to leave to go find a separate app for it.

Sam: Claude Code — no, Claude Docs. They've shipped so many products this year I genuinely can't keep the names straight anymore. Anyway — so the habit forms entirely inside their product, instead of forcing anyone to leave and go open a separate app for it at all.

Alex: Exactly backwards from what Google's doing, and I mean that literally — Google's travelling in the opposite direction. It just opened up eight separate doors, one each for mail, drive, docs, sheets, slides, calendar, chat, contacts, so an outside agent can read and write your own Google files directly, over a login you grant it yourself.

Sam: So instead of pulling the work inside its own app the way Anthropic just did, Google's opening its app up to whoever else's agent wants to come in through the front door.

Alex: And here's the detail that actually stopped me cold. The two outside tools Google names, by name, in its own documentation, as the ones it's building this access for — one is its own coding tool. The other is Anthropic's.

Sam: Google built a door for the exact company taking its enterprise lunch money, so that company can come let itself in?

Alex: Because that door runs on a connector Anthropic invented, gave away for free, and handed over to a neutral foundation to look after. Every company's own integration work, Google's included, now has to speak a grammar written by the competitor they're actually trying to catch.

Sam: Giving the standard away was the real power move the whole time, wasn't it. You stop being the thing everyone has to plug into, and you become the plug everyone else is now forced to use.

Alex: That's exactly it. And it's bigger than just document apps, by the way — the same kind of connector that lets an outside agent touch your spreadsheet is the same kind of thing that threatens a whole category of workflow software, and it reshapes what your own identity team even has to secure, because now it's not just people logging in anymore, it's agents acting as if they were people.

Sam: My team genuinely hit the baby version of that exact problem. New Google rollout landing on one side, my copywriters quietly adopting the rival tool on the other, and nobody above me has actually decided which one our brand's own files are supposed to live in long-term. Which turns out to be a security question wearing a tools question's clothes.

Alex: Right, so here's the part of this whole story that turns from an engineering story into a financial one, and it's the bit that actually made me sit back in my chair.

Sam: Go on then, let's have it.

Alex: Last October, Anthropic locked in roughly a million of Google's own chips, and more than a gigawatt of capacity, for a deal both sides called worth tens of billions. So the company taking Google's enterprise customers runs its own models on Google's own silicon, inside Google's own buildings, billed straight through Google's own cloud division.

Sam: Say that again, slower, because that genuinely sounds insane out loud.

Alex: Google's biggest enterprise rival pays Google rent on the exact machine beating Google.

Sam: Google gets paid either way, no matter who actually wins.

Alex: There's an old playbook for precisely this move, if you want the history-nerd version of it.

Sam: Hit me, I can feel this coming from a mile off.

Alex: Andrew Carnegie didn't dominate the steel industry by beating other steelmakers at making steel. He bought the coal, the ore, the ships that moved the ore, the port, and the railroad connecting every piece of it together. Because he'd learned, working the railroads early on, that freight quietly eats a huge slice of anyone's profit — so he simply bought the freight outright. Then he could cut prices straight through a downturn that bankrupted everyone else, and buy them up cheap right afterward.

Sam: And Google's doing the exact same thing, just one layer further down the economy.

Alex: One layer further down, and meaner about it, if you ask me — Carnegie never once sold his own ore to the mills competing directly against him. Google does exactly that. It's renting the floor to the very company currently beating it on points.

Sam: Okay, but who pays for all of that in the real world, because none of this comes free.

Alex: That's the honest part of it. Google just raised its own spending plans for this year to somewhere north of a hundred and ninety-five billion dollars. And last quarter, for the first time ever on record, its actual cash flow went negative — tens of billions spent on buildings and chips, genuinely more going out the door than coming back in.

Sam: So the stack isn't really a hedge at all. It's a bet.

Alex: It's a leveraged bet that AI demand just keeps compounding forever, without pause. If it does, Google wins no matter who technically wins the model race underneath it. If it doesn't, Google's the one left holding the single biggest construction bill in the entire industry.

Sam: "Own everything so you genuinely can't lose" sounds a lot less comforting once you also admit it means "spend everything so you genuinely can't stop."

Alex: That's the one tension in all of this I honestly can't resolve cleanly for you, and I don't think anyone fully can yet.

Sam: So where do you actually land on this, personally, because I don't think you've said it plainly yet, and I want to hear it.

Alex: Genuinely? I think the stack's real, and it's the advantage. Owning every single layer means Google gets paid in every possible outcome, and that's a genuinely different position than anyone else in this race currently holds. History rhymes here, and Carnegie didn't lose in the end.

Sam: I'm not there with you, not fully. I think you can own the entire factory and still lose, if the one purchase that actually matters — a developer's own private habit — keeps quietly routing itself around you for free. Price doesn't automatically buy trust, and trust is the thing Anthropic's actually been collecting all year, while Google was busy collecting chips instead. In my world, budgets get rubber-stamped around exactly that kind of trust long before anyone above me ever sees a cost comparison. Nobody asks for the cheapest tool first. They ask for whichever tool their own people already quietly prefer.

Alex: So where does that honestly leave the two of us.

Sam: An honest split, I think, is the fair place to land. Tell me the one thing that would actually change your mind on this.

Alex: Easy one for me — the price on this model doubles once November's out. If usage holds up anyway, through that doubling, that's real proof it's not a promotional stunt at all, it's a genuine cost advantage nobody else can actually match. That's my tell.

Sam: Mine's the flip side of that exact same coin. If some other lab manages to match Argon's scores using compute it just rented from someone else, at a price Google genuinely can't beat, then the whole five-layers argument was never really the moat everyone thinks it is. It just meant Google could afford to lose money for longer than anybody else could. That's the one fact that actually kills this story for me, not just dents it a little.

Alex: Which — fair, genuinely, and probably the single thing worth watching hardest out of everything we've covered today, because I don't think even Google can fully rule that one out.

Sam: We'll know within weeks either way. That's the strange thing about this entire industry right now. Nothing stays unresolved for very long. Every big claim either gets paid off or gets quietly dropped, usually inside a single quarter, which is not a pace any other industry I've worked in has ever moved at. Okay — if you only remember three things out of the last forty minutes. One: Google's new model isn't the best because it's the cheapest token. It's the cheapest finished job. And that's a pricing decision, not a modelling one.

Alex: Two: Google paid for this launch with its own lab's independence. The people who protected the long game walked out the door, and the company that made it to market on time is, in a real sense, a different company now than the one that originally built Gemini.

Sam: And three: owning the whole stack wins you the users and the cost curve, but it has not, so far, won back the enterprise budget — which is quietly routing itself to whichever tool a developer happened to like best on an ordinary Tuesday.

Alex: Three different winners, in one single race, none of them actually racing the same race as each other.

Sam: I genuinely didn't expect to end this episode more unsettled than I started it, if I'm honest with you.

Alex: Agreed. I think that's the real state of the whole industry right now — every one of these companies looks completely unbeatable from exactly one angle, and genuinely beatable from the very next angle over. It's a short knowability horizon, and it's absolutely worth the discomfort of following it this closely anyway.

Sam: One honest note before we go, on how all of this actually gets made. This show is AI-generated — Dan builds a custom stack of AI tools to research, check and put episodes like this one together, mostly so he can keep up with all of it himself, and he publishes the results for anyone else trying to do the same thing. AI-assisted, fact-checked, and still worth a second look with your own eyes.

Alex: Before we let you go — one genuinely useful thing you can do, and it costs you nothing at all. Whatever app has you right now, there's a follow or a plus button sitting right there already. One tap. You'll get the next one the second it lands, and for a small independent show like this one, that single tap is genuinely the biggest lever there is for helping it find other people who'd genuinely want to hear it too.

Sam: And one more small thing before you go — if there's something in this episode you'd push back on, or something you want us to go harder on next time, tell us. Podcast at connective shift dot com. We read every single one that comes in, and it genuinely shapes what we go dig into next.

Alex: Thanks for sticking with us through all forty minutes of this one.

Sam: See you next time.