The Product Desk

The spec is three states, not two

A commuter pulls the Duo out at a crosswalk, clears two notifications on the outer screen, and puts it back folded. iOS handles three distinct configurations, and only one of them is contested. Folded, the device is passport-sized and an existing phone layout works untouched. Unfolded, it is near iPad-mini canvas, which asks for a multi-pane or master-detail layout instead of a stretched phone view. Half-folded on a desk, iOS turns the panel into a clock and calendar surface with StandBy-style widgets and an angle-adjustable outer screen for video. That third state is the most distinctive and least contested opportunity on the device. The half-folded state has no shipping apps behind it today.

The Information's reporting adds the interaction pattern Apple floated to developers: Netflix using the outer display to browse clips, with the unfold gesture triggering full playback. Pitched, that is a cinematic reveal. Built, it is a funnel step. Treat the outer screen as a glanceable discovery surface and the unfold as an intent signal, then instrument the unfold so the next investment decision rests on data.

The volume math is what makes this a roadmap item

The industry shipped roughly 20 million foldables in 2025, about 2% of all smartphone sales, per Counterpoint figures cited by Morning Brew. Morgan Stanley expects Apple alone to ship 6.5 million Duos in the December quarter, which Bloomberg pegs at roughly $14 billion, or about 16% of expected iPhone revenue. Counterpoint sees 12 million-plus Apple foldables in 2027. A first-generation product taking a sixth of iPhone revenue in one quarter roughly doubles the foldable install base in a single holiday season. The people buying it are the ones who upgrade annually and pay for the storage tier.

Where the reporting diverges

The dates do not line up, so plan against a window rather than a day. Techpresso reports preorders opened September 9 with the iPhone 18 Pro line reaching stores around September 18. Morning Brew puts Duo preorders on October 16. MIT Technology Review's Download says the Duo ships October 23. Mid-to-late October is when the telemetry changes.

TLDR IT calls the Duo "irrelevant to enterprise roadmaps this quarter" and argues the price ceiling caps addressable volume. That read holds where seats sell through IT. It does not hold where consumer or prosumer ARPU carries the P&L. Techpresso's reviewers flagged slippery edges that made unfolding awkward. That is a real signal that first-generation adoption may run softer than the press cycle implies. Samsung answered with a campaign touting its own foldable lead while Google pulled the Pixel Tablet from its store, so large-format mobile is consolidating around Apple while the app layer above it stays open.

Where the Q4 numbers will mislead

Because the Duo unfolds into a tablet-shaped canvas, iPad-class usage will start appearing in phone-class install data. Expect Q4 movement in session length, screens per session, and tablet DAU that has nothing to do with anything shipped. If device model is not already a first-class analytics dimension, that gap surfaces in the exact week the cohort most needs isolating. Most teams find out it isn't when they try.

Small install bases are worth building for when the buyers are the most valuable ones on the platform and the shelf holds one competitor. That is the case here.

The honest sequencing is compatibility now, exclusives later. Support all three states and make the Duo a named cohort in analytics before preorders open. Then let December sell-through decide whether multi-pane-only experiences earn headcount.

The only public benchmark, and what it is made of

HubSpot's figure matters because everyone else is guessing. Rangan's own segmentation is the usable artifact: a campaign manager with 10 million contacts running a multi-language, multi-region campaign "is going to work in HubSpot all day long," while a CMO prepping a meeting will "use one of these AI interfaces to actually grab that information." That is a persona split you can build to — native depth for high-cardinality operational work, API and artifact endpoints (summaries, decks, trend rollups) for executive read paths. It also tells you where not to spend: in-app chat for high-volume operational workflows is the worst of both surfaces.

Data gravity predicts which side of the line you land on

Salesforce's Parker Harris asked publicly, "Why should you ever log into Salesforce again?" and the company reports that third-party AI connecting into its apps is increasing work through the CRM. Figma's CEO Dylan Field said work will happen "in Figma directly" — while Claude and ChatGPT already manipulate Figma's frames, components, and variants, not just export files. Field also conceded that total chatbot dominance is "not entirely bearish," a softening reportedly prompted by early returns on usage-based monetization. Figma's own consumption data contradicts Figma's public narrative, and the market has taken the data's side.

The pattern across sources is consistent: if an agent must traverse your data to do useful work, it amplifies you. If your value lives in the visual or interactive surface itself, agents route around it. The integration bar has moved from reading content to manipulating structure — read-only integrations will read as unserious in enterprise evaluations within two quarters.

The corroborating move from the other direction

Anthropic is running the same mechanic in reverse. Claude for Microsoft 365 went generally available in Excel, PowerPoint, and Word on paid plans with Outlook in beta, carrying shared context across apps, reviewable edits, and customer-template fidelity. Alteryx bolted an MCP server onto its governed-analytics install base, and Kestra 2.0 now exposes its workflows as MCP tools — turning an orchestrator into an agent control plane. Products that expose their core verbs as agent-callable stay in the value chain; the rest get demoted to passive data sources someone else's agent queries.

Where the sources argue with each other

One camp says instrument agent-mediated usage as a first-class metric alongside logins. The other camp — the Meta token-leaderboard cautionary tale — says raw volume metrics get gamed and convert straight into COGS. Both are right, and the reconciliation is the whole point: meter agent traffic to bill it, not to celebrate it. Every external-agent call should map to an account, an entitlement, and a billable event before you widen scope. Salesforce is the live example of doing it in the wrong order: it has evangelized agent access without solving how to charge for it.

The number to put on the board this quarter: the share of workflow completions that no longer generate a billable session.

Then run the ARR sensitivity. If seat pricing assumes humans log in, model what happens when 30% of completions stop producing a session — and add data-egress activity to churn scoring on your top accounts, because the quietest version of this risk is customers pulling their data out so they stop paying to hold it.

Read the supply side as a financing story

Oracle's quarter is the confirmation underneath the price increase. The company reported 30% top-line growth for the period ending August 31, added $30 billion in AI computing deals, and structured most of them as customer prepayments or bring-your-own-chip arrangements. It spent $28.5 billion on capex while burning only about $5 billion of its own cash, because $11.4 billion arrived as customer prepayments — double the cash its operations generated. Customers are financing their supplier's data centers to secure capacity. That makes capacity access a treasury decision, and it means startups without a balance sheet get rationed. Techpresso adds the component-level echo: HBM shortages are already pushing Chinese AI chip prices up, so the cost direction is not local to one vendor.

The counterweight in the same reporting

log10.io's ClinReg leaderboard update is the most useful procurement artifact in the available reporting. Six new models entered the field in six weeks — GPT-6 Astra, GLM-5.3, GLM-5.3 Flash, Gemini 3.8 Flash, Gemini 3.7 Flash, Muse Spark 1.3 — which gives any single model decision roughly a six-week half-life. Accuracy converged; price did not. GPT-6 Astra leads overall at 89.9 for $10.80 a run; GLM-5.3 sits at 89.3 for $3.80 and takes the agentic crown; GLM-5.3 Flash scores 87.6 for $0.30. Anthropic's Opus 5, at $22.29 a run, scores below the thirty-cent model.

Capability is heterogeneous by task shape, which is the part a roadmap can act on. Closed models still own long-form regulatory prose — GPT-5.6 Terra at 91.2 on IND drafting versus Kimi K3's 86.7. Literature-screening-style tasks are saturated, with open and closed tied, so they belong in your table-stakes column rather than your differentiation narrative. Three tasks, three different winners: the router and the eval harness outlast any individual model pick, which is where engineering budget should sit.

Two traps in your current cost model

First, tier heuristics are unreliable. In one documented production grading pipeline, prompt-cache thresholds made Anthropic's Sonnet 5 cheaper than budget-tier Haiku 4.5 — the premium model won on cost. Effective cost is a function of your prompt architecture and cache-hit rate, not the vendor's price card. Second, the headline spread is list price. PointFive measured one standardized coding task (200K input, 30K output) across five frontier models at $0.35 to $1.75 — a 5x spread on identical work — and log10 is explicit that its 75x figure excludes self-hosting, GPU reservation, ops, and qualification. Put the break-even volume in the business case and write the number down.

The risk you inherit by switching

The cheap winners are Chinese-origin labs, and that carries a dated catalyst. The U.S. accused DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.ai of industrial-scale distillation days before a Washington Trump–Xi meeting where AI is on the agenda, and Anthropic separately alleges Moonshot and DeepSeek routed user requests to Claude without telling users. A legal determination and a named fallback per workload should land before engineering builds a dependency, so a policy block costs a config change rather than a quarter.

Accuracy converged inside five points while price spread 75-fold — which makes the model pick a gross-margin decision, not an engineering preference.