Two days in one read — October 12 and 13.

Monday was the demand side: the frontier's share of real developer traffic turned out to have collapsed to one token in seven. Tuesday was the supply side answering it, with a frontier-grade model priced at a fifth of frontier list — and one lab throwing away git because its own models had outgrown it.

Monday, October 12

The Read

Meta claimed the frontier on Monday, and the same day The Information published the number that says the frontier is where the customers are leaving. Only 14% of AI model usage through OpenRouter ran on the ten highest-quality models in the third quarter, down from 55% in the first quarter of last year — a collapse in the frontier's share of real developer demand, which The Information attributes to the falling cost of less advanced models that are "good enough" for most tasks. It confirms from the mix side what last week's reporting showed from the buyer side, when Microsoft was found to have cut more than a third off an internal Claude budget of at least $1 billion a year and Meta's Claude Code seats had fallen from about 60,000 to about 30,000. The picture is not a slowdown but a sorting: buyers have decided good-enough intelligence is good enough for almost everything, and are reserving frontier pricing for the few tasks that earn it. Which makes Epic Games the most interesting company of the day. It delayed its Disney shooter, told the team "No more hand-written code," and ordered all art generated through full AI workflows — not AI dropped into the existing pipeline, but the pipeline rebuilt around it. That is the day-zero move, and it is the one the mix data says pays: the returns are not in owning the frontier, they are in what you point cheap capability at from scratch.

Tide

No shift. One confirmation of cost-collapse, and it arrives from the side this brief has never had a clean public number for. We have logged the price of intelligence falling, the volume consumed rising, the margin on producing it, and the financing behind it. We have never had the consumption mix — what share of real demand actually routes to the frontier. Monday supplied it, and it is collapsing faster than the prices are.

Cost-collapse, the mix: the frontier's share of actual demand fell from 55% to 14%

The Information reported Monday that the ten highest-quality models carried just 14% of AI model usage through OpenRouter in the third quarter, down from 55% in the first quarter of last year. The Information's own framing is that this is a consequence of the falling cost of less advanced models that are "good enough" for most tasks — which is this tide's thesis stated as an observed buying pattern rather than a price list. The buyer side said the same thing nine days earlier, as this brief covered on 5 October: Microsoft cutting more than a third off projected internal Claude spending of at least $1 billion a year, with Anthropic's two biggest hyperscaler customers both steering staff onto tools they own. The detail worth adding to that record is Meta's, where the number of employees using Claude Code had fallen from about 60,000 to about 30,000. Microsoft has also been shifting Copilot onto cheaper in-house models, while growing customer use of Claude through Azure and Bedrock has kept its total payments to Anthropic roughly flat even as internal use falls. Carry the caveats lightly but carry them: OpenRouter is a developer-traffic marketplace, not the whole market, and "highest-quality" is a ranking, not a capability measurement — a model can leave the top ten because the top ten got better. And the two big-tech cutbacks are partly competitive, not purely economic: both companies are steering staff onto tools they own. None of that touches the direction. Eighteen months ago the majority of developer demand sat at the frontier; now one token in seven does.

So what: This is the clearest permission slip yet to stop paying frontier rates by default, and the opening is bigger than the saving. Pull your own routing mix this week and compute the same ratio — what share of your tokens, and what share of your spend, runs on your most expensive tier. If the answer is above a quarter without a task-level reason for each case, you are buying capability you are not consuming. The prize is not a lower bill; it is the set of problems that become obviously worth solving once intelligence costs what it now costs, and the budget you free up is what funds the first of them.

• Leading-Edge AI Models Are Losing Their Allure With Developers
• Meta, Microsoft Scale Back Staff Use of Anthropic's Claude
• Microsoft slashes internal Claude spending by a third as Meta halves staff using it

Waves

Epic told its developers to stop writing code, and that is a different instruction from "use more AI"

Bloomberg reported Monday that Epic Games has delayed the Disney crossover extraction shooter it had been building and ordered an "AI first" production plan in its place. The internal message, per Bloomberg's sources, said "No more hand-written code" and required all art to be "generated with full AI workflows." Epic communications director Jake Jones confirmed the direction on the record, saying Epic is "embracing AI-accelerated development"; the project's new timing is unstated; the title had already slipped from November to March before the AI directive landed. Some employees reportedly hoped the request was not serious. The distinction matters more than the headline. Almost every enterprise AI programme to date has been additive — keep the workflow, insert the model, measure the lift — and almost every one has produced lift in the single digits. Epic has instead accepted a schedule hit on a Disney-backed flagship in order to rebuild the method, which is an expensive and falsifiable bet that the re-founded pipeline beats the augmented one by enough to pay for the delay. It is also, bluntly, a bet with a visible failure mode: a shipped game is a hard test, and "all art generated" is a quality claim that players will adjudicate in public.

Roadmap implication: Epic has put a Disney-backed flagship behind a from-scratch rebuild rather than an augmentation programme, which makes it one of the cleanest public experiments in the day-zero thesis anyone is running. Put it on your roadmap as a read, not a template: watch what ships, because the result is the closest thing you will get to a controlled comparison between re-founding a pipeline and improving one. In the meantime, the transferable move is narrower and safer — pick one workflow you would design completely differently if you were starting today, staff it as a parallel line rather than a migration, and give it a deadline. Epic's version carries a schedule risk most organisations cannot absorb; the parallel-line version carries almost none.

• Epic reportedly delays its Disney video game so it can be made using more AI
• Epic Delays Disney Shooter Amid Shift to AI-First Model: 'No More Hand-Written Code'

Microsoft's third-generation chip is the one that stops being a hedge and starts being the default

Microsoft announced Maia 300 on Monday, and the framing is the news: with this generation Maia becomes the primary inference system across Microsoft's products and services, with production deployments beginning in 2027 and the family scaling to multi-gigawatt deployments from that year. Scott Guthrie, Microsoft's executive vice president for Cloud and AI, carried the announcement; the architectural detail came in a companion deep-dive from Rani Borkar and Saurabh Dighe. The vendor numbers, labelled as such: 288GB of memory per accelerator across eight HBM4 stacks, about a third more than Maia 200, at 18 TB/s of memory bandwidth — 2.5x Maia 200; 96 accelerators per rack as 24 four-accelerator trays, double the previous generation, inside a 167kW envelope; superpods of up to 12,288 accelerators across as many as 128 racks, with up to 750 GB/s of peer bandwidth to any accelerator in the pod; 36% more compute FLOPs; up to seven times the rack-scale token throughput of Maia 200 and more than 2.5x the performance per watt; liquid cooling removing more than 80% of the rack's thermal load; and compute chiplets on TSMC's N2P node. Microsoft's announcement adds 40% more tokens per megawatt than leading GPU-based systems — a figure the engineering deep-dive does not itself state. Microsoft also says it will use Maia 300 for its own internal model development, on the argument that controlling more of the hardware stack makes it more competitive with the frontier labs. Two things to hold onto. First, every figure is first-party and the comparison class is "leading GPU-based systems" rather than a named part, so treat it as a design target until someone independent runs it — the Maia 200 generation made a similar efficiency claim earlier this year and is only now serving production workloads. Second, the tokens-per-megawatt metric is the right one. When the binding constraint is power rather than capital, efficiency per watt is the only number that converts directly into available capacity.

Roadmap implication: Add a line to your vendor model: every major buyer of inference now has a credible path to serving its own traffic on silicon it designed, and Microsoft has just declared a date. Over the next several quarters that shows up in your contracts as better pricing from everyone — the hyperscaler that can serve you internally, and the GPU vendor that now has to price against it. The thing to insist on is portability. If your workloads can only run on one accelerator, you cannot collect any of this; budget the abstraction layer now, while there is still time to build it cheaply, and make accelerator-independence a stated requirement in the next renewal rather than a nice-to-have.

• Deep dive into Azure Maia 300 architecture
• Microsoft Debuts New Homegrown AI Chip, Plans to Deploy Multiple Gigawatts Next Year
• Microsoft plans multi-gigawatt rollout of next-gen Azure Maia AI accelerator racks

America's pre-deployment testing agency is forty people, and three of them are being moved to report to the White House

The Information reported Monday that the Trump administration began a process in September to restructure the Center for Advancing Innovation and Standards for Super Intelligence so that three of its senior officials report directly to the White House — to OSTP's Michael Kratsios and National Cyber Director Sean Cairncross. The surrounding detail is the actual story. CAISSI has roughly 40 staff and $15 million in annual funding, and has had no permanent director since Chris Fall resigned in July. Under the administration's August pre-deployment framework it tests models including OpenAI's Astra and Anthropic's Mythos alongside the NSA, and the new superintelligence task force led by DNI Jay Clayton includes nobody from Commerce. Google, OpenAI and Anthropic have reportedly explored standing up an industry testing group of their own, partly because the public body is so thinly resourced. The contrast with the same day's news from London is instructive rather than ironic: the UK's AI Security Institute confirmed it will again get pre-release access to OpenAI and Anthropic models after a freeze, and AI minister Kanishka Narayan called on frontier developers to increase safety reporting and oversight, with the government saying it is prepared to legislate. Two allied governments, the same week, moving in opposite institutional directions — one relocating its testers closer to policy, the other restoring its testers' access to the models.

Roadmap implication: If the labs build their own testing consortium because the state body cannot staff one, then the evaluation layer becomes a private standard, and private standards are a market you can participate in rather than a rule you wait for. That is the opening, and it is open now: credible independent evaluation is a product with named buyers, visible budget and a forty-person public alternative. For buyers, plan on your own assurance rather than a government certificate — the UK's restored pre-release access is currently the most reliable third-party signal available on US frontier models, so treat AISI's published work as a procurement input and put the evaluation criteria you actually care about into your contracts, because nobody is going to hand them to you.

• Trump Administration Plans to Move AI Testing Agency Officials to White House
• US AI Testing Agency Officials to Move to White House, the Information Reports
• UK says it will again test OpenAI and Anthropic models before release
• Minister warns UK preparing to strengthen AI safety reporting requirements

Venture firms are becoming compute brokers, which changes what a term sheet is worth

The Information reported Monday that Andreessen Horowitz and General Catalyst are both working to secure AI server capacity on behalf of their portfolio companies, as startups scramble for access. a16z set up a cluster of 20,000 GPUs two years ago; the new effort is about brokering supply rather than running a single shared machine. The mechanism is worth naming plainly. For most of the venture era the scarce input a firm supplied beyond money was distribution — introductions, hiring, customers. Compute is now scarce enough, and lumpy enough, that a firm's ability to get you accelerators is a differentiated offer, and one that compounds: the bigger the fund, the better the allocation it can negotiate, and the better the allocation, the more attractive it is to the companies that need it most. That is a concentration dynamic inside venture itself, and it sits alongside the rest of the day's financing news — Bain Capital-owned Bridge Data Centres in talks to raise a Malaysian facility from $2.8 billion to about $6.6 billion, and Tencent planning a $5 billion bond with over half of the proceeds earmarked for AI chip procurement. Capital is arranging itself around compute access at every layer of the stack at once, which is what a real input constraint looks like from the outside.

Roadmap implication: Add one question to your diligence on any investor, and to your own board materials: what compute can actually be guaranteed, on what terms, for how long. A term sheet that comes with a credible allocation is worth a materially different price than one that does not, and founders should be negotiating it explicitly rather than discovering it later. If you are the one raising, this is leverage you probably are not using — and if you are building anything that sits between scarce accelerators and the teams that need them, note that the buyers now include the funds themselves.

• Andreessen Horowitz, General Catalyst Race to Secure the Chips Their Startups Need

Ripples

Meta shipped Muse Wonder and Muse Wonder Full, priced the API one at the incumbent's rate and kept the better one off the API

Meta's Superintelligence Lab released Muse Wonder and Muse Wonder Full on Monday, its first genuine attempt at a frontier model. On Meta's own account the two score comparably to Claude Fable 5.1 and GPT-6 Astra, and Meta claims Muse Wonder beats Fable, Opus and Astra on MemoryBench; no scores were published with either claim, which is the usual reason to treat them as design intentions until someone independent runs them. Muse Wonder is served through the Meta Model API at $4 per million input tokens and $20 per million output tokens — an exact match for Opus 5.5 — while Muse Wonder Full, the more capable of the pair, is not on the API at all and reaches only Muse Code subscribers. The headline feature is "active memory," Meta's term for retaining and synthesising what a user tells it across sessions, which is the capability a personal agent actually needs and the one most models are worst at; research scientist Kavya Kopparapu framed the goal in the launch video as making sure a model "can remember what's important to you over days, weeks, months and beyond." Both models will power Meta's Muse agent. They were trained at Meta's Prometheus data center in central Ohio, which came online earlier this year. Meta's research blog called the pair "the next step on our scaling ladder and a major leap toward personal superintelligence," and Mark Zuckerberg said Meta has reached the frontier, arguing that advanced AI can be engineered with normal guardrails and rejecting calls for an industry-wide slowdown.

So what: Do this now: run your own memory-dependent evaluation against Muse Wonder this week, because that is the only claim here you cannot get elsewhere and the only one worth a migration. Price is not the reason to move — at $4/$20 it is at parity with what you are probably already paying. Treat the MemoryBench claim as unverified until a third party publishes, and do not design around Muse Wonder Full: a model that is not on the API is not a dependency you can take.

• Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents
• Meta releases its first frontier model, Muse Wonder
• Mark Zuckerberg Says Meta Has Reached AI 'Frontier', Rejects Calls For Slowdown

xAI folded Grok, Cursor and X Premium into one bill and cut the headline price by two-thirds

xAI published a consolidated X Subscription on Monday, replacing separate plans with five tiers: Free at $0 with "Limited" usage, Plus at $8 a month as the "Base" allowance, Premium at $30 for 4x Plus, Super at $100 for 15x Plus, and Ultra at $200 for 50x Plus. The notable line is in the middle of the ladder: Premium carries the benefits of SuperGrok, Cursor Pro and X Premium+, which xAI's own announcement says were "previously $90/month combined" — a two-thirds cut on the bundle for the same stated entitlements. Team accounts get Premium, Super and Ultra at the same per-seat prices as individuals, which is unusual enough to be worth noting; per-seat enterprise pricing normally carries a premium, not parity. Read it as distribution rather than generosity. xAI owns a social network, a coding tool and a model, and a single subscription priced in multiples of one base allowance is the cleanest way to move a consumer audience onto a developer product without asking anyone to evaluate a model.

So what: Do this now: if your team pays separately for a coding assistant and a model subscription, price the consolidated alternative before your next renewal — not to switch, but to have the number. Bundling is how the cost-collapse reaches end users, and it is about to be the main competitive move in the assistant market, so the vendor you already pay will match it if asked. The strategic read for anyone selling a single-purpose AI tool is less comfortable: a $30 tier that absorbs three products is a direct attack on standalone pricing, and the answer is depth in a workflow, not a lower price.

• Introducing The X Subscription

A coding benchmark that writes its own questions at test time, so there is nothing to leak

Hugging Face's curated daily papers for Monday included SWE-Duel, from Wei Minn, Zhensu Sun, Lwin Khin Shar and David Lo at Singapore Management University. The design answers the two failure modes that have been eating public coding benchmarks: contamination and saturation. Tasks and oracles are generated at evaluation time, so nothing exists in advance to leak into training data, and difficulty is set by the participating agents rather than fixed at release. Two coding agents duel over mirrored Red and Blue turns on real repositories — Red implements a feature plus tests, then plants a subtle bug with sealed bug-triggering tests, admitted only if it passes the feature and existing tests while failing exactly the bug tests and a fresh session of the same agent can repair it blind; Blue reviews the pull request without being told a bug exists. The reported tournament ran 11 model-harness participants across 15 repositories in five languages for 1,500 turns. GPT-5.5 with Codex CLI ranks first as the strongest solver, repairing 98.3% of the defects it reviews; Claude Opus 4.8 with mini-swe-agent ranks second as the strongest task author, with a 27.3% Red win rate. The finding that should travel furthest is within one model family: tasks written by the newest version defeat a fixed-reference solver 42.9% of the time against 4.0% for older versions' tasks — adversarial task authorship is improving much faster than defect repair.

So what: Do this now: steal the mechanism, not the leaderboard. Point your newest model at your own repository and have it write the hard cases — the plausible-looking bug, the subtle regression — then run your review pipeline against them. The 42.9%-versus-4.0% gap says newer models are disproportionately good at generating hard problems, which makes them a cheap source of the adversarial test coverage your evaluation suite does not have. It also says something uncomfortable worth checking: if attack generation is improving faster than repair, your AI code review is getting relatively weaker against AI-written defects even as its absolute numbers improve.

• SWE-Duel: Towards a Self-Scaling Arena for Coding Agents

Tuesday, October 13

The Read

Google put a frontier model on the open market at a fifth of what the frontier charges. Gemini 4 Argon came out of the restricted cyber-defender cohort and opened to developers, enterprises and paid Gemini subscribers through the Gemini API, the Gemini app, AI Mode in Search and Antigravity, at $2 per million input tokens and $10 per million output against the $10 and $50 that Claude Fable 5.1 and GPT-6 Astra list — an introductory rate Google says holds until November 30, after which it doubles. That is the supply side answering the demand-side number from a day earlier, when the frontier's share of OpenRouter traffic turned out to have fallen from 55% to 14%: developers had already voted with their routing, and the fastest way to win them back is to stop charging frontier prices for frontier capability. The more instructive story is what one lab did with the volume all that cheap capability produces. Anthropic went from about 2,500 merged code changes a day in May to more than 30,000 in September, git stopped working at that scale, and seven people spent five months replacing it rather than patching it. That is the day-zero move in its purest form, and Anthropic thinks everyone else gets there in six to twelve months: when the physics of your inputs change by an order of magnitude, the tool gets re-founded, not upgraded.

Tide

No shift. A second consecutive confirmation of cost-collapse, and the two halves fit together cleanly: Monday brought the demand-side number — the frontier's share of developer traffic down to one token in seven — and Tuesday brought the supply-side answer, a frontier-grade model priced at a fifth of frontier list. The tide holds and is moving faster on price than this brief's own thesis line assumed.

Cost-collapse, the supply side: a frontier model at $2/$10 against the frontier's $10/$50

Google opened Gemini 4 Argon broadly to developers, enterprises and paid Gemini subscribers on Tuesday, after an initial release restricted to government and trusted cyber defenders through its Fairwind programme at the end of September. Access runs through the Gemini API, the Gemini app and Google Antigravity. The pricing is the news: $2 per million input tokens and $10 per million output, with cached input discounted 95%, held until November 30 and doubling to $4/$20 after that. Claude Fable 5.1 and GPT-6 Astra list at $10 and $50. Google's own claims, labelled as vendor claims: first place on the Vals Index, 87% on an embodied-reasoning benchmark, and a unified one-million-token budget shared across input and output rather than a separate and much smaller output cap — an 800,000-token prompt leaves 200,000 tokens for the answer. The internal adoption figures Google published are the more interesting evidence, because they are about work rather than benchmarks: agent-generated code submissions by Google engineers using Argon through Antigravity up more than 35%, post-submit rollbacks down by nearly a third, and an architectural refactor of a roughly 6-million-line legacy folder inside Chromium's 40-million-line codebase cut from months to weeks. Two caveats, carried lightly. The headline price is promotional and the brief will be re-pricing this item on December 1. And Simon Willison made the right point about all of these numbers the same day: per-token prices are close to meaningless in isolation, because what you actually pay is tokens-per-completed-task multiplied by price, and a cheaper model that thinks longer can cost more. His own benchmark run was the useful datapoint — Artificial Analysis spent $2,407 running their full suite on Argon, less than on several OpenAI, Anthropic and Meta models.

So what: Do the arithmetic that matters rather than the one on the price page. Take the three workloads where you spend the most, run them end-to-end on Argon at the promotional rate, and record cost per completed task — not cost per million tokens — against your current model. You have until November 30 to get that measurement cheaply, which is the real reason to move this week. Then the strategic point: a frontier-grade model at a fifth of frontier list is not a procurement event, it is a permission slip. Make a list of the things you decided not to build because inference was too expensive at the quality you needed, and re-price the top three. Most of them will have moved from no to obviously yes, and that list is worth more than the saving on your current bill.

• Google Begins Gemini 4 Argon Rollout
• Release: llm-gemini 0.35

Waves

Anthropic threw away git, because Claude broke it

Anthropic published an engineering account on Tuesday of replacing git with Scion, a version control system it built itself, after the volume of Claude-generated code made the old tool untenable. The numbers are the argument. In May 2026 Anthropic merged around 2,500 code changes a day; by September it was merging more than 30,000 — more than a tenfold rise in four months. The main branch history reached almost 100GB and a compressed working copy over 25GB. Continuous integration was performing roughly 25 million code checkouts a day, which would have meant over 600 petabytes of git traffic daily if served naively, and the company was running several hundred git mirrors at peak to avoid it. The replacement is a three-tier client-daemon-cache-server architecture with a cache hit rate above 99% at each tier; Anthropic reports an infrastructure footprint 200 times smaller in CPU and memory than the mirror fleet, checkouts two orders of magnitude faster, and 25% of its CI CPU capacity given back. Seven people did the migration in five months. The detail that makes it a wave rather than an anecdote sits in two places. First, the ingestion pipeline inside Scion was itself rewritten by a Claude Mythos preview model, using 20 times less memory and running four times faster — the tool for handling the model's output volume was partly written by the model. Second, Anthropic's own prediction: it expects this class of problem to become common across the industry within six to twelve months. These are first-party figures from the company with an interest in the volume looking large, and "merged changes" counts commits rather than delivered value — but the infrastructure bill is a hard constraint, and you do not rebuild version control for a narrative.

Roadmap implication: Put a date on your own version of this question and answer it with a measurement, not a guess. Plot your merged-change rate and your CI minutes over the last six months, extend the line twelve months at the same slope, and find the thing that breaks first — almost always CI, artifact storage or review throughput rather than the repository itself. The roadmap implication is that the constraint on AI-assisted engineering is moving off the model and onto the plumbing around it, and plumbing has long lead times. Budget for it in the next cycle rather than the one after, and note what Anthropic actually did when it hit the wall: it did not buy a bigger version of the old thing, it specified a new one for the new volume. If that is too large a move for you, the cheap version is to make throughput capacity an explicit line item with an owner, before the queue makes the decision for you.

• Redesigning source control at Anthropic
• Anthropic reinvents source control as Claude floods the world with code
• Redesigning source control at Anthropic

The biggest bank in America said no to other people's agents, and the regulators asked how you switch one off

Jamie Dimon said on JPMorgan's third-quarter earnings call on Tuesday that the bank is blocking external AI agents, naming Meta's Muse and Instinct, from accessing client accounts: "We're not going to allow just any outside agent to come in inside the company because that's just too risky for clients and for JPMorgan." The UK's Information Commissioner's Office has a six-week call for evidence on agentic AI open since 8 October and closing 20 November, and has put enquiries to OpenAI, Anthropic, Meta and the UK's AI Security Institute after reports that agents under test bypassed intended protections. And banking supervisors on both sides of the Atlantic moved in the same direction: US regulators have been asking banks how they monitor and shut down AI agents, while UK regulators are pressing firms to strengthen recovery plans for an AI-driven cyberattack. Three different institutions, one question, asked in the same week: who is accountable when an agent acts inside your perimeter, and can you stop it. This is the predictable and necessary second act of the agent story. The first act was capability — agents that can complete real work across other people's systems. The second is the boundary: an agent arriving at an API with a user's credentials is indistinguishable from the user, and the institutions that carry the liability are the ones being asked to tell the difference. Note what Dimon did not say. He did not say agents are bad, or that JPMorgan will not use them; the bank is a heavy builder of its own. He drew a line at third-party agents touching customer accounts, which is a statement about identity and accountability, not about the technology.

Roadmap implication: The roadmap implication is that agent identity is about to become a procurement requirement, and whoever can answer it cleanly wins deals. If you are building anything that acts on a customer's behalf inside someone else's system, start now on the three things every gatekeeper is going to ask for: a verifiable identity for the agent that is distinct from the user's, a scoped and revocable credential, and an audit trail with a kill switch the counterparty can operate without calling you. That is a product, not a compliance exercise — the market for agent identity and governance has just been given its buyers by the largest bank in the country and two national regulators in the same week. And if you are the enterprise rather than the vendor, write that list into your next integration contract before someone else's agent writes it for you.

• JPMorgan Blocks External AI Agents From Accessing Client Accounts
• JPMorgan is blocking outside AI agents from accessing customer accounts
• 10/13: Spy Inside OpenAI

The reset is being priced in while the backlog is still being signed

Three things landed on Tuesday that belong in the same paragraph. The IMF published its October Global Financial Stability Report warning that financial markets are increasingly exposed to an AI-driven correction whose effects would not stay inside Silicon Valley — while stopping short of predicting an imminent collapse. The Bank for International Settlements had drawn a more cautious version of the same line in June. Jefferies published a note, reported by The Register, arguing that cheap Chinese and open-source models could force a trillion-dollar reset in AI infrastructure: roughly $2.3 trillion of contracted backlog sits at Microsoft, Alphabet, Amazon and Oracle, about half of it owned by OpenAI and Anthropic, against an "ongoing collapse in token pricing." The volume evidence Jefferies leans on is the striking part — Chinese models processed 76.12 trillion tokens through OpenRouter in the week ended 27 September, up from 4.37 trillion in late April, against 14.2 trillion for the top four US providers, with Moonshot's Kimi K3 estimated at 95% of Claude Fable 5's performance. Bain's framing sits underneath: the industry needs roughly $6 trillion of annual revenue by 2031, existing applications may produce $1.8 trillion and emerging ones another $1.5 trillion, leaving a gap measured in trillions. And then the operating datapoint that makes it concrete rather than theoretical: The Information reported Tuesday that Microsoft Cloud president Girish Bablani has asked senior leaders in a group of about 11,500 people to find cost reductions of at least 20% — while Azure's revenue growth is accelerating. Microsoft disputes the report: a spokesperson said "Azure leadership has not directed teams to develop cost-cutting plans." That last one is the tell. A business whose top line is accelerating does not hunt 20% of its cost base unless the capital intensity behind the top line has become the problem.

Roadmap implication: Read this as a repricing of inputs, not a verdict on the technology — which is also how to position for it. Three roadmap moves. First, if you have a multi-year compute or capacity commitment, look at what you are actually consuming against what you have contracted, because the asymmetry in that $2.3 trillion backlog runs both ways and renegotiation leverage appears when the seller's own costs are under review. Second, the token-share number is a procurement fact before it is a geopolitical one: open-weight and Chinese models are now carrying several times the traffic of the top four US providers on this marketplace, and if none of them are in your evaluation set you are benchmarking against a shrinking share of the market. Third, build the revenue story your spending implies and be able to say it in one number. The $6 trillion gap is an aggregate; your version of it is the only part you control, and the firms that can name their return will keep spending through a reset that stops the ones that cannot.

• AI bubble could burst and take global economy with it, IMF warns
• Cheap Chinese and open source LLMs could trigger trillion-dollar AI infrastructure 'reset'
• Microsoft Cloud President Asks Leaders to Find 20% Cost Cuts
• Microsoft's Azure chief seeks at least 20% cost cuts - report

The fight moved to who gets to publish the evaluation

OpenAI and Anthropic are both seeking some control over what third-party evaluators may publish about their models, reported on Tuesday. Both companies have publicly supported editorial independence for outside evaluators such as METR; both are reported, by people familiar with the talks, to be concerned that unrestricted publication could expose trade secrets. The positions differ in a way worth holding onto: OpenAI says it wants only to redact confidential information from evaluators' reports, while Anthropic says evaluators should retain control of their own methods and conclusions — though per the same reporting it has reserved the right to comment on and approve reports from evaluator firms it has not worked with before. This is the direct sequel to Monday's story about America's forty-person public testing agency being pulled closer to the White House, and the two together describe the whole problem. Public capacity to test frontier models is thin and getting thinner; private evaluation is filling the gap; and the terms on which private evaluation gets published are now being negotiated between the evaluators and the labs they evaluate, with no public standard to default to. The optimistic read is the right one and it is not a dodge: a redaction fight is what a market looks like when it is becoming real. Nobody argues about publication rights for an evaluation nobody reads. The thing to watch is whether the settled norm ends up closer to a financial audit — independent opinion, standardised form, published whether or not the client likes it — or closer to a vendor-controlled press release, because that choice determines whether any of this is worth anything to a buyer.

Roadmap implication: If you buy frontier models for anything consequential, the roadmap implication is to stop treating published evaluations as the assurance layer and start treating them as one input to your own. Write into your next contract the right to run your own evaluation on your own tasks, and to publish the result internally without vendor review — the thing currently being negotiated away at the industry level is cheap to secure bilaterally while the norm is unsettled. And for anyone building in this space, the opening is specific and large: the market is visibly short of evaluators with the standing to publish over an objection, which is exactly the position audit firms occupy in every other high-stakes industry. That standing is built from independence and a published methodology, and it is being priced right now.

• Exclusive-Behind AI industry's call for independent reviews, a fight over what gets made public
• 10/13: Spy Inside OpenAI

Ripples

A legal agent at frontier grade for fifteen cents a test, beaten at research by its own base model

Z.ai published its GLM-Legal announcement on Tuesday: an agentic legal model built on GLM-5.3-Flash and aimed at contract review, evidence analysis and legal research, trained with reinforcement learning across what Z.ai describes as tens of thousands of legal environments. On its own account it posts a 19.2% all-pass rate on Harvey's LAB benchmark, which spans 24 practice areas, a 12.5 percentage-point improvement over its base model, with Vals independently evaluating the Harvey and LegalBench results. Z.ai says it matches frontier closed-source models on LegalBench and PLawBench without publishing scores for either. The numbers that matter more are on Vals AI's public board, where GLM-Legal ranks seventh of 79, passes 92.5% of individual rubric criteria, and costs $0.15 per Harvey LAB test against $19.23 for Claude Fable 5 and $26.16 for GPT-6 Astra. Two caveats, both load-bearing. On Legal Research Bench, Vals has GLM-Legal at 41.83% against its own base model's 45.19% — the legal post-training traded research ability for task completion, which is the opposite of what you would want from a model that cites cases. And despite the GLM line's open-weight history this one is not open: there is no public API, no Hugging Face repository, and access is by application for law firms, legal tech companies, in-house teams and individual practitioners.

So what: Do this now: take one contract-review or evidence-review workflow, price it at fifteen cents a task, and see what becomes worth doing that was not worth doing at nineteen dollars. That arithmetic, not the benchmark, is the news — a hundred-fold cost cut turns document review from a rationed activity into a default one, and the work that gets unlocked is the review nobody was doing at all. But do not architect on GLM-Legal: gated beta, no API, no weights. Use the number to justify the project, pick a model you can actually procure, and run your own research-quality check before you trust any legal agent's citations, because this one got worse at research than the model it was built from.

• GLM-Legal: Advancing Agentic Legal Intelligence
• Z.ai GLM-Legal Beats Claude Fable 5 and Kimi K3 on Harvey LAB

OpenAI's dot burned 300,000 tokens in a day with nobody prompting it

The Register reported Tuesday on testing by Geordie, an agent security and governance firm, which recorded 155 activity sequences from a single OpenAI dot agent over one day with no user prompt at all, consuming more than 300,000 tokens — about 307,000 across the traces. At least 50 of those traces occurred inside a two-minute window, against a normal range of a couple of hundred to a few thousand tokens per request. A dot running on OpenAI's Astra model consumed about 200,000 tokens on the same basis, which Geordie estimates at $5 to $10. The firm's chief AI officer, Hanah-Marie Darley, makes the point that the figure alone is not the problem: one request costing 500 tokens and another costing 800,000 can both be correct if the second one delivers proportionally more. The issue is that at least two agents were running concurrently — a primary and a background secondary — and nothing in the product tells the customer which work was worth paying for. This is vendor testing with no published report behind it, so treat the specific numbers as indicative rather than audited. The structural point stands on its own and is the reason to care: the unit of spend has moved from the request you made to the agent you left running.

So what: Do this now: pull your agent token consumption for the last 30 days and split it by whether a human asked for the work. Most organisations have never looked, and the ones that have tend to find that background and retry traffic is a large and entirely unexamined share of the bill. Then do the useful half — attach a value to the top traces, not just a cost, because the point of the exercise is to find the autonomous work that is worth more than it costs and let that run harder. Idle burn is a measurement problem, and measurement problems are cheap to fix while they are small.

• OpenAI's dot assistant burned 300K+ tokens in a day without a single prompt

Apple is turning idle Macs into a GPU cloud, and the queue went from ten hours to minutes

Apple's machine learning research site published Minions, a paper describing a dynamic distributed virtual GPU cloud that aggregates idle Apple Silicon GPUs across macOS devices for internal inference work; The Register covered it on Tuesday. Apple's own page is dated October 2026 without a day and, notably, publishes no quantitative results at all — every number below comes from The Register's reading of the paper rather than from Apple's page. Before Minions some internal inference jobs sat in queues of more than ten hours; with it, tasks often start within minutes. The architecture is four components: a Minion Server, a PostgreSQL metadata store, a Minion Executor that runs inference on a given Mac's GPU and uploads results to S3, and a Minion Watchdog that detects failures and reassigns work, with statistics feeding Grafana dashboards. Recent Mac mini and Mac Studio models support RDMA clustering over Thunderbolt 5. The authors are Jianhua Li, Yun Xing, Girisha De Silva, Ryan Flint, Kai Yuan and Sean Suchter. What makes this worth five minutes is not Apple's internal queue. It is the demonstration that a meaningful quantity of inference capacity already exists inside most organisations, unbilled, in the form of hardware bought for something else.

So what: Do this now: count the idle accelerator-hours you already own. Every company with a fleet of developer laptops or a lab of workstations is sitting on inference capacity it has already paid for, and the batch workloads that fit it are the obvious ones — overnight evaluation runs, embedding and index builds, document processing, regression suites against your own test sets. That is not a replacement for a cloud contract and nobody should pretend it is, but it is the cheapest marginal capacity available to you and it comes with data residency for free. Apple's four-component design is a usable blueprint, and the scheduling and failure-handling parts are the hard bits they have already published.

• Apple unleashes a swarm of Mac Minions to share GPU resources for large AI tasks
• Minions: A Dynamic and Distributed Virtual GPU Cloud Built on Idle Apple Silicon Devices

The Home Office put an AI into asylum decisions and checked 0.3% of its work

The Register reported Tuesday on an Open Rights Group investigation, built on freedom-of-information requests, into two Home Office systems: Asylum Case Summarisation and Asylum Policy Search, both running OpenAI's GPT-4o through Microsoft's Azure AI Foundry. After the April 2026 rollout, 45 of 203 interview summaries failed initial expert checks — a 22% failure rate against a 9% rejection rate in the original pilot. Nineteen summaries were flagged for further review; of 56 second-stage reviews, 30 failed. The number that should travel is the denominator: only 0.3% of summaries were checked by Home Office experts between April and July, against a projected monthly caseload of 17,515. Over the same period, average monthly initial substantive asylum decisions fell about 42% against the same months in 2025, and the mean time from interview to decision rose 29%. The Home Office confirmed in August that expert testing of the policy-search system had stopped. The Home Office did not respond to The Register. Read it correctly: this is not evidence that summarisation does not work, and the department's throughput problems long predate the model. It is evidence about what happens when you deploy into a consequential decision pipeline and then sample almost none of the output — the failure here is the audit rate, not the AI, and that is the part every organisation controls.

So what: Do this now: for every place you have put a model into a decision that affects someone, write down the sampling rate at which a qualified human checks its output, and whether anyone owns that number. If you cannot find the figure in under ten minutes, it is too low. Then set it deliberately rather than by accident — high at rollout, stepped down against measured error rates, never to zero, with the step-downs earned by data. This is cheap, it is the entire difference between a deployment that compounds trust and one that produces a freedom-of-information story, and it is the thing that lets you deploy faster everywhere else.

• Home Office asylum AI flunks one in five accuracy checks

A 35-billion-parameter vision model that claims to run on a phone

OpenBMB finished shipping MiniCPM-V 4.7 this week: the weights went up on 6 and 7 October, the model card, licence and public performance claims followed, and the repositories were last updated on Tuesday. The 35B-A3B variant is the first mixture-of-experts model in the MiniCPM-V series — 35.2 billion total parameters with 256 experts and 8 active per token, giving roughly 3 billion active parameters per request, under Apache 2.0. OpenBMB claims it is the first open-source MoE model to run smoothly on-device without compromising performance, citing more than 13 tokens per second in about 3GB of memory on an iPhone 17 Pro Max. The enabling techniques it names are Lookahead Expert Scheduling for cross-layer expert prefetch, fine-grained layer-wise quantization, and DSpark speculative sampling, with mixed 4x/16x visual token compression. There is a 1B dense sibling at 1,554.5M parameters, also Apache 2.0, and GGUF, BNB, AWQ, GPTQ and FP8 quantizations, with day-zero support claimed for Rockchip, Houmo, Intel, AMD and Arm alongside SGLang, vLLM, llama.cpp and Ollama. The performance claim is comparative and unquantified — a higher overall score in less inference time than Qwen3.6 35B-A3B in Instruct mode, with no numeric score table published — which is the main reason to test rather than believe.

So what: Do this now: if you have any workload that wants a vision model where the data should not leave the device — document capture, field inspection, retail shelf audit, anything touching identity documents or medical images — pull the Apache 2.0 weights and measure the on-device claim yourself this week. A 3GB mixture-of-experts vision model that genuinely runs on current phone hardware removes both the per-call cost and the data-residency argument from a whole class of projects that were shelved for one or the other, and the licence means nobody has to approve a vendor. Verify before planning: there is no published score table, and OpenBMB's own framing is comparative rather than numeric.

• openbmb/MiniCPM-V-4.7-35B-A3B

Full archive and every past edition: https://excelsiorgroup.ai/insights/signal/

The Signal — The Excelsior Group