Rent the AI muscle, build the internal brain

Something shifted in AI this spring, and you can read it in three events that look unrelated and aren't.

In early June, both Anthropic and OpenAI filed confidentially to go public. Read what the filings sell. OpenAI leads with reach: about 900 million weekly active users, enterprise seats roughly doubled in a year, half a dozen revenue lines stitched into one story - and, under all of it, a forecast loss near $14 billion for the year. Anthropic leads with velocity: a run-rate that went from about $9B at the end of last year to roughly $47B by May, carried by enterprise adoption and by Claude writing code. Neither prospectus opens with "our next model will be the smartest." They open with adoption, dollars, distribution. The two companies that spent three years insisting only the next model mattered are now selling investors on everything except it.

Second: Sridhar Ramaswamy, the CEO of Snowflake, ran a public head-to-head - an open model, GLM-5.2, from a lab most enterprise buyers couldn't have named a year ago, against Claude Opus 4.7 on a real data-engineering benchmark. The open model scored 66 to Opus's 67, at about a fifth of the cost. When the CEO of a ninety-billion-dollar company runs that test himself and posts the result, the cost question has stopped being a footnote.

Third: the labs themselves started hedging. The model race didn't stop. In the same three weeks as the filings, both shipped their most powerful models yet - Claude Fable 5 from Anthropic, built for multi-day autonomous work, and OpenAI's GPT-5.6 line, previewed the morning I drafted this. But watch where the rest of the energy went. GPT-5.6's marquee feature is "ultra mode" - subagents that split the work in parallel - next to two cheaper tiers: Terra at half the price of 5.5, Luna cheaper still. The flagship's headline is orchestration and cost, not IQ. Anthropic put Claude inside Excel, Word, and PowerPoint, and made it something you @-mention in Slack like a coworker. The companies that built their whole identity on raw capability now spend just as hard on cost, orchestration, and access - alongside the frontier, not instead of it. The tell isn't that models stopped mattering. It's that the people who know models best are placing their other bets in plain sight.

Three signals, one message: raw frontier capability is no longer the only game. It's table stakes. If capability isn't the moat anymore, where did it go?

Why function was the game

For three years, the game was function. Make the model smarter and you win. That worked because of who does the judging. The people who evaluate frontier models - who run the benchmarks, write the threads, anoint the new "state of the art" - are developers. And coding is the one job in the enterprise that rewards pure function at almost no accessibility cost. A great model in a terminal is useful to an engineer the moment it lands. Nobody has to make it legible to a non-technical team first. And the scoreboard rewards exactly that: Claude Fable 5's June launch led on coding and agentic benchmarks - SWE-bench, FrontierCode, Terminal-Bench, OSWorld - while GPQA, the field's marquee science test, never made the table.

So the value landed where it was demanded, and stalled where more than function is needed. About 5% of public GitHub commits are now AI-authored; 41% of new code is AI-generated; Anthropic says 90-95% of its own product code is written by its own model. The money agrees. By a16z's tally, enterprise AI spending in early 2026 ran about $3B on coding alone - more than every other category combined, from legal to support to medical admin. The rest of the org has barely deepened. Gallup's workplace numbers show AI use climbing everywhere but staying shallow outside tech: plenty of "I tried it last week," little "I depend on it daily," especially in healthcare, retail, manufacturing, government. Benedict Evans calls it the capacity gap - the distance between adoption and dependence.

The gap has nothing to do with desire. These teams need the thing dev never did: access. Legibility, integration, a version that shows up where they already work - no prompt-engineering hobby required. Function alone was never going to reach them. And the function race is hitting diminishing returns anyway: a point or two of benchmark at five times the cost.

And the models are converging. Plot the frontier labs on one aggregate benchmark over three years and they bunch into the same tight band - OpenAI, Anthropic, Google, Meta, the Chinese labs - pulling each other up, no network effect holding any of them ahead. A capability everyone reaches and no one can fence is the definition of a commodity. So the moat moves. Three live bets on where it lands.

Bet one: access

The first bet is the labs' current focus, and the oldest play in enterprise software: distribution. Get embedded in the surfaces where work already happens, and let world-class capability be a commodity you happen to be best at.

That's what Slack and Office are. It's part of what the IPOs are for - war chests for a distribution land-grab, not another model race. It's the oldest playbook in business, and there's something funny about watching the most advanced companies on earth rediscover that being in the room beats being the smartest one in it.

There's a trap inside this bet, and the labs are walking in with their eyes open. The architecture everyone is selling - a frontier "advisor" agent supervising a swarm of cheaper workers - is the same architecture that teaches enterprises to spend less on frontier models. Put the expensive model in the advisor seat, push the token-heavy grunt work down to cheap or open models, and the bill to the people who sold you the advisor shrinks. The labs are training their best customers to need them for a thinner and thinner slice of the work. Distribution may be the only bet left, because the function bet is quietly eating itself.

Bet two: orchestration as a commodity

The second bet says the layer above the models - the routing, the stitching of cheap models into frontier-grade output - becomes the commodity, and the value flows through it.

OpenRouter, fresh off a $113M raise, made the bet concrete this month. Its Fusion product takes your prompt, fans it across several models at once, has a judge model score the answers, and returns the synthesis - "mixture of agents" in a single API call. The pitch is flagship quality at half the cost, with no orchestration code on your end. The counterintuitive part: the blend can beat the best single model. In OpenRouter's own numbers, a Fable 5 plus GPT-5.5 panel scores about 69% to solo Fable 5's 65%. The arbitrage Ramaswamy ran by hand is now a checkbox.

And it's generalizing past any single API. Factory.ai sells autonomous "droids" to engineering teams and routes every task through a shared model router that learns which model is worth paying for. Hatz.ai sells the same logic down-market to small businesses and MSPs: dozens of models behind one interface, every message auto-routed to the cheapest one that can do the job. Different buyers, one promise - stop picking models, let the layer pick for you, optimize for spend.

If this bet is right, nobody holds the orchestration layer, because it's a feature anyone can rent. The cheap-model-plus-clever-routing trick stops being a craft and becomes plumbing. The moat doesn't live here. It passes through.

Bet three: orchestration as the moat

This is the bet I've spent the most time inside, because my hands have been in exactly this layer for months. The commodity story is true for the easy half and wrong for the hard half.

Garry Tan at YC frames the split as brain versus muscle. The muscle is execution: chaining model calls, routing, running tools, catching errors - orchestration for a single workstream. That half is easy now and getting easier. Fusion is the muscle, productized. You can rent it.

The brain is the other thing. Memory that persists across sessions and across agents. The organization's accumulated context. The synthesis layer that knows what was decided last week and why. That half is hard, scattered across a hundred half-finished repos and threads, and almost none of it comes out of the box. The distance between the two is wider than anyone selling you orchestration wants to admit.

The cleanest evidence I have is my own. I ran the same multi-agent job three times - same task, same models, the same raw intelligence on the bench each run. The first run threw 237 agents and 24.6 million tokens at it over about three hours and produced nothing: the naive fan-out melted into a server-side rate-limit cascade and collapsed before it finished. A later run did the identical job with 18 agents and 1.6 million tokens in 27 minutes. It worked. Same models. Fifteen times less spend. The variable was never the agents' intelligence. It was the discipline around them - dedupe before the expensive step, throttle the fan-out, persist the work before you verify it, fail fast instead of hammering an overloaded API. None of that lives in the model. All of it lives in the harness.

You can watch this play out in public. Bridgewater - the hedge fund - wanted models to triage research the way its analysts do: which document matters, where the boilerplate ends, what's worth a human's time. No frontier model cleared the bar. Gemini, Claude, and GPT sat near a coin flip cold, and climbed only to the mid-70s with expert-written prompts - short of the 80% an analyst would trust. Newer releases barely moved the number per dollar. So Bridgewater stopped shopping for a better brain and built one. It fine-tuned an open model, Qwen, on its own experts' judgment, and beat every frontier model it tested at about a fourteenth of the cost. Thinking Machines, the lab that ran the training, calls it differentiated intelligence. The base model was rented. The judgment was theirs.

The reasoning itself - the thing the labs sell, the thing GLM-5.2 now matches at a fifth of the price - is commoditizing. Short of someone cracking AGI, a different subject, the frontier "advantage" keeps shrinking from quarters to months to weeks, until only a genuine generational leap - a Mythos, not a point release - moves the zeitgeist at all.

The strongest tell that the brain is the moat: the companies that have built one don't sell it. Ramp will tell you its internal agent now writes more than half of its merged pull requests. It will tell you how, in broad strokes, and never in copyable prompts or diagrams, and it will not open-source a line. They folded hundreds of scattered agents into one framework with thousands of skills, and that framework is a competitive secret and a recruiting magnet, not a product. That's not stinginess. That's a company that knows exactly which layer the moat sits on. It isn't the model.

So where did the moat go?

Three bets: access, orchestration-as-commodity, orchestration-as-moat. From the operator's chair, only the third holds. The labs will win plenty on distribution. The arbitrage layer will keep commoditizing, Fusion-style, until the cheap-models trick is just plumbing. And underneath both, the durable edge is migrating to the brain - the memory, the accumulated context, the discipline, the thousand small rules that keep a swarm from melting under its own fan-out - the one layer nobody ships in a box, because the few who can build it have every reason not to. So the moat didn't vanish. It moved.

That's the layer I'm working in now. If you've got tips, learnings, war stories, send them - the brains that move the needle aren't for sale yet.