Newsletter · · Ashutosh Agarwal

The Labs Stop Selling You a Model and Start Selling You the Work - Platform Watch - Week of July 24, 2026

Startups and venture newsletter for the week of July 24, 2026. The Platform Watch edition on how OpenAI and Anthropic pivoted from renting model access to selling finished work outright with ChatGPT Work, Claude Cowork and OpenAI Presence, and what that platform shift exposes and defends for founders building on top of them.

Platform Watch

Week of July 24, 2026: The Labs Stop Selling You a Model and Start Selling You the Work

For two years the deal was simple: the labs rented you intelligence, and you built the product. This week that deal changed. Anthropic passed OpenAI to become the most valuable startup on Earth, and it got there on the back of a product, not a model. Both labs shipped software agents that don't answer your questions, they do your job: finished spreadsheets, slide decks, websites, customer-service calls. The model was always the thing you rented. Now the labs are renting out the labor itself, and that lands right on top of a lot of startups.

This Week's Platform Move: Anthropic Becomes the World's Most Valuable Startup by Selling Work, Not Tokens

If you build on top of a foundation model (the big AI systems from OpenAI, Anthropic, and Google), here is the one thing to take from this week. The center of gravity at the labs has shifted from "here is a smart model, go build something" to "here is a worker, let us do the task for you." That is a different business, and it competes with a different set of companies: yours.

The headline is that Anthropic has overtaken OpenAI as the most valuable startup in the world, and the reason is telling: it did it largely on the strength of Claude Code, a product, rather than on raw model bragging rights. A rundown of the enterprise numbers laid out this week: Anthropic now holds 41% of paying enterprise firms versus OpenAI's 39.5%, and its momentum is concentrated in fresh money: 73% of first-time enterprise buyer spend. In the specific market that matters most as a leading indicator, coding, Anthropic holds a dominant 54% share against OpenAI's 21%. The logic offered for why coding is the tell: "A developer doesn't care if your CEO is famous. They care if the model can debug a thousand lines of Python without hallucinating a fake library" (Elon Musk Podcast, "Anthropic overtakes OpenAI as most valuable startup," July 23, 2026).

And this was the week both labs stopped pretending they only sell models. OpenAI shipped ChatGPT Work and Anthropic has Claude Cowork, desktop agents built to run in the background for hours and hand you finished output. As the same podcast put it: "The shift is from a conversational assistant to an agentic workspace. So instead of answering questions, ChatGPT Work is designed to run in the background for hours. Generating finished spreadsheets, slide decks, and websites. It's a fundamental pivot from retrieving information to executing labor" (Elon Musk Podcast, "Anthropic overtakes OpenAI as most valuable startup," July 23, 2026). The two are built differently in a way founders should note: ChatGPT Work is "a cloud-oriented orchestrator" that reads from connected apps like Slack and Google Drive but "struggles to write back to those external systems," while Claude Cowork "operates directly on the user's local computer" and reads and writes files natively, which is why developers heavily favor Claude Cowork. The labs claim 10 million combined active users across the two, though the hosts flagged that number as soft: usage-limit resets inflate it, and "the distinction between an active user and a curious observer is critical when you're evaluating product market fit."

Then, two days later, OpenAI walked straight into another startup category. It launched OpenAI Presence, "an enterprise AI platform that lets companies build, deploy, and manage real-time voice and chat agents for customer support, sales, HR, and IT." Everyday AI's Jordan Wilson has been waiting for this: "I've thought for years that eventually one of the first big disruptions in AI would be customer service. But the real-time models weren't narrow enough, or fast enough, or nearly smart enough. Now they are. And the path for disruption is almost impossible to ignore." His read on the strategic shift is the whole Platform Watch thesis in one line: "While Anthropic has recently been grabbing headlines by pulling models from subscription, OpenAI is seemingly focusing beyond the model now" (Everyday AI Podcast, "Ep 825: New: OpenAI Presence. Has The AI Customer Service Takeover Finally Arrived?," July 23, 2026). If you are an AI customer-service startup, the platform you build on just became your competitor.

The move up into your product came with a price war down below. OpenAI now runs seven pricing tiers, "free with ads, go at $8, plus at $20, a pro tier at $100, a second pro tier at $200, and business at $25 per user," and its flagship, GPT-5.6 Sol, is priced at $5 per million input tokens specifically to undercut Anthropic's Fable 5 at $10 (a "token" is roughly a word-piece; you pay by how many go in and come out). The $100 and $200 tiers exist to capture power users "running parallel agentic workloads," one agent writing code, another querying a database, a third drafting emails, all at once. Meanwhile GitHub Copilot "switched to consumption-based billing," and, as the host summed up the whole shift: "The era of simple flat-rate software seats is basically dying… You are no longer buying software. You are renting compute by the second" (Elon Musk Podcast, "Anthropic overtakes OpenAI as most valuable startup," July 23, 2026).

Why This Matters, and the Catch Nobody's Pricing In

Here is the tension every founder should sit with. Building on these labs right now delivers genuinely absurd leverage, and the best example of the week came from health-insurance startup Curative, whose founder Fred Turner walked through what "an agent instead of a department" actually looks like:

  • Curative built an in-house credentialing agent that runs on Claude. The old process (checking a doctor's license, transcripts, and malpractice history) "used to take us two to three months on average and cost about fifty dollars." The agent does it end to end "and we're now averaging about 12 hours turnaround time… and it costs about twenty cents."
  • It built a contract-negotiation agent named Gwen that emails providers, researches them, negotiates rates, and even clicks the DocuSign button. Cost per contract fell from "about $1,500 to $2,000 on average… the average with Gwen has been about $70." Volume went from "about 100 contracts a week to about 100 contracts a day." Gwen sends roughly 15,000 customized emails a day, and its edge is relentlessness: "a lot of providers will get them on the ninth email. There is no way that a human is going to email them nine times."
  • And it is quietly gutting its software bill: "we just recently cancelled our Salesforce contract because we have an internal CRM that was vibe coded that is working better… $600,000 a year gone to zero." Curative is "cutting about 80% of our SaaS spend this renewals." As a leading indicator, its "Anthropic cost over the last six or seven months has 6x every month from a base of a couple of tens of thousands of dollars now up to millions of dollars a month" (The Twenty Minute VC, "20VC: $5BN in Revenue, 7 to 7,000 Employees in 9 Months… Curative with Fred Turner," July 18, 2026).

That is the upside. Now the catch, and it is the same one Platform Watch keeps circling back to. The labs are not making money at these prices: they are buying the market. The subsidy comparison again went straight to Uber: "It mirrors the early rideshare wars… You offer a service well below its actual cost to build a habit among users… Uber and Lyft burned billions to change consumer behavior… And OpenAI and Anthropic are doing the exact same thing with enterprise behavior." The difference from Uber is that "Uber was subsidizing a physical ride with a known cost floor," whereas OpenAI's cash burn "gives them less than four years of runway on paper." The warning for anyone who has wired their operations into a single lab: "You're essentially outsourcing your profit margins to a vendor who is currently operating at a loss. And when they decide to stop operating at a loss, your margins are the first thing to disappear" (Elon Musk Podcast, "Anthropic overtakes OpenAI as most valuable startup," July 23, 2026).

On All-In, the same dynamic showed up as a live debate about who should even be paying for the frontier. David Sacks pointed out that cheap open-weight models are wildly cheaper than the flagships, "you're selling most of the product for 50 cents per million tokens when they're selling theirs for $56," and Chamath Palihapitiya argued most companies over-spend anyway because the person choosing the model isn't the person paying: "the engineer wants to go on an exploration on using the latest greatest thing… the CFO is tied to the money." His guess at how many top-tier prompts are overkill and "should be running" on a model a hundredth of the cost: "98%." Their read on why Ramp's spend-control product exists at all: "Eric would not have released this Ramp product unless CFOs were like, I can't control the spend" (All-In, "Can the AI Industry Regulate Itself? Stripe Wants PayPal, China Catches Up, NY Bans Datacenters," July 18, 2026).

And the floor keeps dropping. China's Moonshot released Kimi K3, a 2.8-trillion-parameter model benchmarking "right alongside the best from Anthropic and OpenAI," and said it will publish the full weights for anyone to download and run for free. As one show framed it, "frontier-grade ability no longer rented from a lab. Yours to take" (Business of Tech, "AI Capability vs. Accountability: Who Owns the Harness?," July 21, 2026). When Hugging Face needed a model with no guardrails to work through a security incident, it reached for a Chinese open model, GLM 5.2, because the US flagships refused the task (Intelligent Machines, "The Beans are in the Mail - Can American AI Compete When China Gives It Away?," July 23, 2026). Cheap, capable, ownable models are now the water everyone swims in, which makes the question of what you own on top of them the only one that matters.

Exposed vs. Defensible (as Called Out This Week)

Exposed

  • The thin wrapper, a clever prompt with a logo. The bluntest version came from Brian Herr in an episode literally titled "The SaaSpocalypse." His test: "If someone opened a fresh ChatGPT window right now and got roughly the same result your product delivers, would your customers notice the difference or would they even care?" His verdict on wrappers: "If people have done some thin wrappers around someone else's model, those are going to have a hard, a hard path forward if they're going to survive at all," because "if I can do it with my favorite AI model, why would I pay someone… to do it for me?" Wrappers, he said, are "not going to make it very long" and are "certainly not investable": "they'll either be automated, they'll become part of the platform, they'll become cloud skills… or it's something that can be built pretty much immediately" (Futureproof Founder Podcast, "407. The SaaSpocalypse: Why Startups Fail in 2026," July 21, 2026).
  • Consumer and prosumer apps, and single-feature utilities. Rob Walling was careful to say most SaaS survives, but named exactly what doesn't: "Consumer and prosumer apps specifically, I think, are gonna take a big hit because consumers and prosumers are so cheap, and they'll blow a weekend vibe coding something to kill a $100 a year subscription." Also on the list: "simple single-feature utilities… the app that someone pays $9 a month or $50 a year… to convert a PDF to a JPEG." And workflows that are just plumbing: "If the product was the workflow, meaning a thin layer moving data from A to B, an agent might eat it" (Startups For the Rest of Us, "Episode 842 | What is the Future of SaaS in an AI World?," July 21, 2026).
  • The "system of record" moat. The old defense that switching costs would save incumbent software is weaker than it looks. A Tungsten Automation executive: "The old moat of 'I'm your system of record, therefore you can't replace me,' that moat is gone." His proof: "I rewrote a 20-year-old personal productivity app of mine in three days… moving the data, like that's easy. I went to go make a tea and I came back and it was done" (Motley Fool Hidden Gems Investing, "The Old Software Moat Is Dead," July 19, 2026).
  • Legacy horizontal SaaS with a fat seat price. Curative's cancellations are the canary: Salesforce ($600K/year, replaced by a vibe-coded CRM in two months), Looker (migrated to Snowflake), 80% of SaaS spend on the chopping block (The Twenty Minute VC, "Curative with Fred Turner," July 18, 2026). Menlo's Matt Murphy noted the flip side: how much these tools now cost to actually use: "companies like Salesforce are spending $300 million on tokens" (Equity, "Menlo Ventures' Matt Murphy says the lesson for founders now is that a great model isn't enough," July 22, 2026).
  • AI customer-service startups. OpenAI Presence is a direct, first-party entrant into voice and chat support, with Anthropic and Google's Gemini Live close behind (Everyday AI, "Ep 825: New: OpenAI Presence," July 23, 2026).
  • Model-agnostic coding tools' independence. The neutral, bring-your-own-model story keeps getting complicated: Cursor is being folded into SpaceX in a $60 billion deal, and GitHub Copilot has moved to consumption-based billing, so "enterprise buyers now have to evaluate software owned by an aerospace company" (Elon Musk Podcast, "Anthropic overtakes OpenAI as most valuable startup," July 23, 2026).
  • Anyone building "10 years down the line" for a clean IPO exit. Storm Ventures' Arun Penmetsa laid out the squeeze on the roughly nine-in-a-hundred good-but-not-top-1% companies: "One is the question that they get is, can Anthropic come in your domain? And the second is, there is no clear exit path." The old route (go public at $100–200M ARR) "doesn't exist today… the bar for public has gone up immensely high" (The Neon Show, "The #1 Mistake Killing B2B Startups | Arun Penmetsa, Storm Ventures," July 21, 2026).

Defensible

  • Proprietary data you generate and nobody else has. DoorDash co-founders gave the cleanest example. Its delivery robot needs to know the exact spot a human dasher drops food (the "first and last 100 feet"), and "that data doesn't exist anywhere else. It doesn't exist in Google Maps. It only exists at DoorDash," pulled from 10 billion deliveries. The broader point they made about the lazy "incumbent data advantage" take: a database of customer records "has very little to do with the thing we're trying to accomplish with an agent." The real moat is data purpose-built for the exact task, plus a world-class operations team the labs won't replicate (No Priors, "Building an Autonomous Delivery Experience with DoorDash Co-Founders Andy Fang and Stanley Tang," July 23, 2026).
  • Specialized intelligence trained on private data. Fireworks CEO Lin Qiao argued the whole premise of "one model to rule them all" is wrong because "the majority of world's data is actually private data, locked inside application, locked inside enterprise… it will never get shared with anyone else because this is company's proprietary IP." Her bet: fine-tune a smaller model on the "small amount of unique data a particular company has" and "you are better than a general purpose model." She warned of a new failure mode where the very best companies can't afford to serve their own users: "we have great companies that have product market fit… but they cannot scale because once they scale, they could scale into bankruptcy" (The Twenty Minute VC, "20VC: Are OpenAI and Anthropic Overvalued?… with Lin Qiao, Founder and CEO @ Fireworks," July 20, 2026).
  • Workflow lock-in for the 90% who will never build it themselves. Liminal's founder makes the case for serving "the average user, the normal user" rather than power users: watch how an accountant does a monthly variance analysis, then offer to run it automatically. "That person's not going to go set up that workflow chain ever… certainly not in the next decade." A developer's eye-roll, "I could write that in 15 minutes using Claude," misses that "normies outnumber all of us by like 10 to 1" (The Enterprise AI Show, "What is a Behavioral Agent Automation Platform?," July 19, 2026).
  • Owning "the harness": the accountability layer on top of the model. Dave Sobel's framing was the sharpest of the week: "The model is the commodity. The harness, the thing that checks and constrains it, is the value." As AI writes the code and acts on its own, "the review was the real work and the scarce work," and almost no one is doing it on purpose. The durable position is to be "the named accountable party that validates what the AI produces, scopes what an agent is allowed to do on its own" (Business of Tech, "AI Capability vs. Accountability: Who Owns the Harness?," July 21, 2026). The creator of Claude Code made a compatible point from inside a lab: don't rely on one moat. Using the "seven powers" business framework, he argued switching costs will matter less ("if you want to port from vendor A to vendor B, you can ask Claude… and it'll just do it") but the biggest companies "don't just have one mode… when you combine these modes, you get a lot more power" (Odd Lots, "The Creator of Claude Code on The Hottest Piece of Software in the World," July 20, 2026).
  • Distribution, brand, and human relationships, the things code can't write. Rob Walling: "Cloning the software was never the hard part. Getting someone to know you exist, to trust you with their business, to switch off what they're already using, and to stay for years is the hard part. AI writes the code. It does not write your distribution… A clone with little or no distribution is a folder on someone's laptop" (Startups For the Rest of Us, "Episode 842," July 21, 2026). Curative made the same bet on where humans stay valuable: "there's two areas where we're really investing in people: that's technical skills and relationships," because the broker on a million-dollar contract "wants to have a finalist presentation, they want to go to dinner, they want to go and play golf" (The Twenty Minute VC, "Curative with Fred Turner," July 18, 2026).
  • Deep regulatory know-how and hard-won compliance. Tungsten processes invoices "compliantly in 140 countries… to get the license to do that in 140 countries would cost millions of dollars and take years." That kind of accumulated risk-transfer ("you can't replace 30, 40 years of know-how") survives the SaaSpocalypse (Motley Fool Hidden Gems Investing, "The Old Software Moat Is Dead," July 19, 2026).
  • Delivering a vertical outcome the customer won't maintain themselves. Arun Penmetsa's advice for the founder worried about "can Anthropic build this?": sell the outcome, not the tech. "Higher sales, better healthcare for your patients, higher revenue, I will deliver that. What is it worth to you?" Yes, a customer's engineers can say "I'll build it with Claude," but "are they going to maintain it for 10 years? Are they going to make sure when compliance and audit comes through?" His frame: "there's a big gap between the application and the model. So go build for that gap." And the newly investable category is "where the physical world meets the data… more of a moat, and a little bit more AI safe" (The Neon Show, "Arun Penmetsa, Storm Ventures," July 21, 2026).

A caution worth keeping: the people describing the "SaaS is dead" future are often the ones who profit from you believing it. Rob Walling's rebuttal to the panic ("the moat was never the code, it was everything you added to the code") is a useful counterweight to the doom, and even the labs' own coding-tool creator, asked whether he was popular in a Valley he's disrupting, pointed you to check incentives: "I would probably ask who's saying this and what are their incentives?"

Founder Takeaway

This week the platform risk got more specific. It is no longer just "a lab might launch a competitor into your category." It is that the labs have crossed from selling intelligence to selling finished work (spreadsheets, support calls, negotiated contracts) while cutting prices to buy the market. That is fantastic for your cost line and lethal to anyone whose product was really just convenient access to a model.

So bank the leverage, but be honest about what you own:

  1. Run Brian Herr's test on your own product this week. If someone opened a fresh ChatGPT window and got roughly your result, your customers will eventually notice. "Helpful gets cut in a budget meeting. Essential does not." If you can't name why you're essential in one sentence, you have a feature, not a company.
  2. Own data or relationships the labs can't reach. DoorDash's drop-off data, Fireworks' fine-tunes on locked-up enterprise data, Curative's provider relationships, Tungsten's 140-country compliance, those survive a price war. A prompt does not. If your only asset is skill at calling someone else's API, that skill is being commoditized in real time.
  3. Re-cost your business on today's prices, then assume they rise. GPT-5.6 Sol at $5 per million, open weights at "50 cents," Kimi K3 free to run, enjoy it, but remember the lab burning cash today has "less than four years of runway." When the subsidy ends, "your margins are the first thing to disappear." Route cheap models where you can (Chamath's "98%" of tasks are overkill) and don't build a business that only works while someone else eats the cost.
  4. Sell the outcome and own the harness. Customers don't want a model; they want a result they don't have to maintain, audit, or babysit for ten years. Be the accountable layer that checks the AI, or the vertical product that delivers the number the customer cares about. That gap between the raw model and the finished outcome is exactly where a durable company still gets built.

The labs just told you what they're building: not tools for you to build with, but workers that do the work directly. The founders who win from here aren't renting that layer most cleverly, they own the data, the relationships, and the accountability the layer can't absorb.