Every Model Has a Six-month Shelf Life Now

Context Window

Every Model Has a Six-month Shelf Life Now

Plus: The shoggoth strategy for generalists, moving your website off Webflow with Claude, and what the Every team couldn't stop talking about.

Hello, and welcome to the weekday edition of Context Window. Grant Harvey is Every’s newest staff writer, and he’ll be bringing you what the Every team is testing, debating, and building with AI.

Anthropic filed to go public on Monday, and its S-1 puts a number on something we’ve seen at Every: When a new model drops, many of us switch to it as our daily driver. Models older than six months earn less than 20 percent of Anthropic’s revenue. Today, we unpack what that means for which model you should be using. Plus: head of evals Mike Taylor on how to look like a specialist while staying a generalist, how he had Claude move his website off Webflow over morning coffee, and the industry stories we couldn’t stop talking about.—Kate Lee

Was this newsletter forwarded to you? Sign up to get it in your inbox.

---

Signal

Anthropic bets you’ll keep upgrading. That’s a good bet for you, too.

What happened: Anthropic filed to go public on Monday. In his S-1 breakdown for Every, Evan Armstrong compares Anthropic to a movie studio: You’re only as good as your next release.

The filing showed how heavily Anthropic depends on its newest models. Models older than six months earn less than 20 percent of its revenue. And the $16.3 billion it brought in during the first half of 2026 came before its two newest and most well-received models, Opus 5.5 and Fable 5.1, were released. As Mike put it, “This is all Opus 5 money.” Other notable takeaways from its filing:

• Customers from a year ago now pay about nine times as much.
• The company lost only* $693 million on operations in the first half of 2026.
• Anthropic says its latest releases cut customers’ cost per task by 25 to 45 percent, through lower prices, and more efficient models.

Why it matters: The short shelf life of the latest Claude model is Anthropic’s core risk. By 2028, Anthropic has committed to spending about $155 billion a year on compute. Most of those commitments can’t be canceled, while customers can stop spending at any time. If a few model generations slip, Evan argues, that spending turns from inventory into debt.

It’s also a bet that customers can tell each new model is better than the last.

If that bet turns out to be wrong, “80 percent of revenue coming from models released in the last six months is a huge risk,” says Mike. He points to Steve Yegge of The Flat Curve Society, who describes this as a “discernment horizon,” where most users’ tasks aren’t hard enough to show the difference.

What it means: Every CEO Dan Shipper has a new way to think about how AI changes work—as an inchworm. Engineers at the head stretch forward first, knowledge workers in the middle follow along, and the whole rest of the economy at the tail catches up last. What engineers do today, knowledge workers will do tomorrow.

• Engineers, at the leading edge, push models hardest, and as engineer Thorsten Ball puts it, a model that’s wrong 10 percent of the time won’t tell you which 10 percent, so you pay for the gap in retries. (More on Thorsten’s post later.)
• As knowledge workers pick up what engineers do today, expect their token spend to climb sharply and frontier models to start mattering to them, too. Dan thinks that makes working in AI “look a lot like working in finance,” where you “manage a book of token spend” and hunt for an edge.
• That shift could take years to spread through the economy, and it may never reach some people. For them, a better Google search is enough, and today’s models already deliver it.

Try this: Each time a new model ships, read our Vibe Check. If it sounds relevant to your work, switch your default to the new model and test it on your personal benchmark: one hard task you know well.

Skip: the $80.7 billion loss in Anthropic’s headlines. It’s mostly a paper loss on Amazon’s convertible notes that disappears after the IPO.

(An Every reader turned Evan’s article on Anthropic’s S-1 into a short explainer video. Click to view.)

(*Why the asterisk on the $693 million? It’s the operating loss, and it hides how much money is moving: Anthropic spent $8.9 billion on compute alone in the first half, and its contracts take that bill to about $120 billion next year.)

---

‘The Every Podcast’: An Emmy-winning writer’s rules for working with AI

AI can help you tell stories—but only if you have something to say.

That’s the view of Emmy Award–winning television writer and producer Fred Graver, whose résumé runs from Late Night with David Letterman and Twitter to his own company, The AI Writer’s Studio. Before opening AI, he starts with a blank page and writes out every thought he has about an idea. Then he brings in AI as a virtual writers’ room to challenge and develop it. The tools can help shape the story, but its point of view has to come from him.

On this week’s episode of The Every Podcast, Graver joins Natalia Quintero, Every’s head of enterprise AI adoption, to explain how he works. They discuss why writers should brief AI on their tastes and ambitions before pitching it an idea, the cast of AI collaborators he’s built into his own writers’ room, and why the quirks that make characters endearing still come from people.

Watch on X or YouTube, or listen on Spotify or Apple Podcasts. You can also read the transcript.

Here are the highlights:

• Tell AI who you are as a writer: Graver recommends writers create two documents: one with your inspirations, tastes, and ambitions, and a second about the project. Then ask AI whether the idea fits your voice and interests, or whether you’re chasing something that isn’t really yours. He also uses it to simulate a focus group, including “the guy you really wish you hadn’t invited.” The aim is to help you think through your work, not to have it write for you.
• Give your AI collaborators different jobs: Graver’s “very unpleasant” Claude agent keeps him accountable, reminding him of existing commitments when he arrives with a shiny new idea. ChatGPT plays the eager film student who knows every storytelling rule in the book but not what they mean—useful for checking the basics. And NotebookLM’s AI podcast hosts supply encouragement: “They love what you’re working on.”
• The magic is in the quirks: Claude helped Graver brainstorm a web series about seniors running their retirement home’s Facebook page. AI could suggest audience insights and story ideas, he says, but not “the personal quirks that would make these characters endearing.” His example: The voice actor behind Homer Simpson improvised “D’oh!”—a small human contribution that became central to the character.

This is a must-watch for any writer wondering where they fit in a writers’ room that now includes AI.

Miss an episode? Catch up on Dan’s recent conversations with Sam Altman; Anthropic head of product Mike Krieger; the team that built Claude Code, Cat Wu and Boris Cherny; the team that built Codex, Thibault Sottiaux and Andrew Ambrosino; Vercel cofounder Guillermo Rauch; and others to learn how they use AI to think, create, and relate.

---

Skill share

Turn a story idea into a pitch deck

TV and Film Pitch Deck Prep is a Claude skill by Fred Graver that helps you build a story bible, a pitch deck, and a record of your creative decisions. It follows the Kenn Adams Story Spine, the storytelling structure used by Pixar, and asks questions that a producer would, pushing you for specifics.

Randy Counsmann, Every’s head of video, said, “I tried it and it did feel genuinely creative versus ‘one-shot.’ It helped me explore an idea and weigh decisions.”

Here’s how to run the skill from start to finish. You can also refer to Fred’s Start Here guide.

1. Install the skill

In this folder, you’ll find the skill as a ZIP file. Upload it as it is and don’t unzip it.

Download the skill’s ZIP file from the folder and leave it zipped. In Claude, go to Settings → Capabilities and enable code execution and file creation so the skill can generate Word and PowerPoint files. Under Skills, select Upload skill, and choose the ZIP file. Menu labels may vary.

Each person using the skill needs to install it in their own account. Claude Code users can unzip it into their skills folder instead.

2. Start a fresh chat

Skip the skill’s Test this button, which invents a story for you. Start a new conversation with your own idea: “I want to build a pitch for my TV show,” or “Help me with the logline for my film.”

A paragraph, rough notes, or the version you’d tell a friend is enough. Graver started with The Influencer, a Christmas story pitched as It’s a Wonderful Life meets Black Mirror.

3. Answer three opening questions

The skill starts with “Who acts?”, “What’s irrevocable?”, and “What does the ending cost?” If your answers aren’t clear enough, Claude asks follow-up questions.

4. Develop the pitch, section by section

Claude guides you through the logline, hook, synopsis, Case for the project, world, tone, audience, themes, structure, and How I Work sections. Expect some back and forth: Graver’s logline took eight drafts.

Claude won’t write your logline, hook, Case, or How I Work section, so the writing is always yours. It helps organize your sentences, explains its cuts and rearrangements, and saves ideas that belong in later sections.

5. Track your decisions

The skill numbers each decision and marks replaced choices SUPERSEDED, preserving how your pitch evolved. It also flags contradictions: when Graver called The Influencer a web series, it reminded him he’d already chosen a holiday movie.

6. Review and save

Before generating the final files, Claude shows you the deck outline, suggests updates based on your later answers, and asks what you want changed.

Save documents as you go. If file creation is off, Claude provides text you can paste into Gamma or another slide tool.

By Fred Graver · © 2026 The AI Writer’s Studio

---

Steal this workflow

Let Claude move your website to a cheaper host

When Webflow charged Mike $360 to renew his personal site, he decided to move it. Claude was able to migrate the whole site for him, “including exporting all my posts, reverse engineering the design, and moving the domain name” to Cloudflare, which costs about $5 a month. He finished migrating it over his morning coffee and got the Webflow charge refunded. His verdict: It’s “literally easier” to change the site by prompting Claude to write custom code than to use a no-code website builder.

The workflow:

• Export everything. Ask Claude Code or Codex to pull every post, page, and image off your current site builder into plain files.
• Rebuild the design from the live site. Point the agent at your site’s URL, have it recreate the look as a static site, and check the result page by page against the original.
• Move the domain last. Have the agent walk you through pointing your domain at the new host and confirm the new site is live before you cancel the old plan.

Copy this prompt:

Migrate my website from [Webflow] to a static site I can host on Cloudflare Pages. Export every post and page, recreate the current design from the live site at [URL], and list each step needed to move my domain. Don't change any domain settings until I approve.

You’ll need: A Cloudflare account and the willingness to maintain custom code instead of a drag-and-drop editor.

Bonus tip: It’s also helpful to install the Cloudflare connectors in Claude, though Mike says he just gave Claude his browser. It logged into GoDaddy and moved the domain for him.

Try it this week: If a site-builder renewal is coming up, run step one before you pay it.

---

Inside Every

The shoggoth specialist strategy

“There are no prizes for range.” Writer Erik Hoel’s line, from his post on the curse of the polymath, struck Mike “right to my heart.” Dan’s experience has led him to the opposite conclusion, a case he’s been making since 2024. AI is strongest at the predictable work specialists used to own, he argues, and weakest at the messy, new problems where generalists thrive.

Mike agreed that wielding a wide range is more fun but argued the market penalizes anyone who does more than one thing. His fix is what he calls the shoggoth strategy (named for the fearsome creatures in H.P. Lovecraft’s fiction): Present one clear specialty to others, and hide your messy generalist skills behind it. When he was starting his marketing agency, his nightly networking events went nowhere until he mentioned to his prospective clients he’d spent $50 million on Facebook ads. That one line finally got a reaction and became his signature (the smiley face of the shoggoth); behind it were the SEO, data science, and copywriting.

My read: They’re both right, and Mike’s shoggoth mullet (specialist in the front, generalist in the back) is how you get paid for both. Dan’s right that range is what lets you crack the new, messy problems models can’t, while Mike’s right that nobody hires you strictly for your range. The specialty out front is what gets you in the door, and the broader skill set behind it (what makes you you) is how you solve the problems once you’re inside.

Try it: Write out the one sentence you’d lead with if a stranger could hire you for only one thing. If you use AI for most of your work already, you could even brainstorm this with your agent. Ask it:

based on everything you know about me, list 10 of my most proficient skills, and which of the ten is the most impressive for me to lead with as a single specialty to market myself.

Then audition this specialty with your friends and industry peers for a human gut check.

---

Overheard at Every

These are the news pieces the Every team couldn’t stop talking about this week, ranked by how many of us weighed in on Slack.

• Microsoft offered buyouts to a quarter of its staff. About 60,000 employees are eligible for buyouts. Important: This counts who can leave, not who will. The team’s reactions: “Dang,” “wow,” “Ouch.” Dan then asked Opus whether the news breaks his “After Automation” thesis. Its read: The eligible teams reportedly included Copilot and CoreAI, so this looks like a struggling AI strategy being cleaned up, not agents taking jobs. The weak spot in Dan’s thesis it found: Microsoft can shrink payroll to pay for data centers even if AI isn’t doing anyone’s job, and the new work Dan predicts may show up at small AI-native companies rather than the giants paying the compute bill.
• Thorsten Ball’s field notes on working with agents. Member of the frontier team Kieran Klaassen’s favorite rule from Thorsten’s post: Stop worrying about agents filling your codebase with slop. Kieran says, “Opus 5.5 and other models are good enough to just keep shipping. We need to dream bigger.” Mike disagreed with Thorsten’s claim that compaction makes managing context by hand a waste of time. He starts a fresh thread for every task and keeps memory turned off, because old instructions and early mistakes pile up and make answers worse.
• A one-line prompt found a serious PHP bug. A security team matched professional researchers’ finding using its own testing setup, a frontier model, and the prompt “find bad vulnerabilities” in the programming language, Dan Lorenc wrote on LinkedIn. Mike: “Prompt engineering is dead lol.” Dan’s reply: In six months, the easy finds will be gone and prompting will matter again. Try it: Point your coding agent at your own code with the same prompt.
• OpenAI claimed three more breakthroughs. On Monday it announced AI-produced proofs in math (the Kakeya conjecture in every dimension), computer science (that the exponential time hypothesis is false), and physics (the existence of quantum chromodynamics, including part of a Millennium Prize problem). This time, outside experts are shepherding the results, after its October release of 722 math manuscripts drew pushback and three withdrawals for errors.

---

Log on

Upcoming events

• Context Engineering With DSPy: Thursday, October 29, 1–2 p.m. ET. Head of evals Mike Taylor, author of the new O’Reilly book Context Engineering with DSPy, joins Drew Breunig and DSPy lead maintainer Isaac Miller to build a DSPy program from scratch: one you can measure and improve, so it keeps working when you switch models. Bring your questions. Free for Every subscribers. Register.
• All Access office hours: Friday, October 30, 12–1 p.m. ET, on Zoom: Host Natalia Quintero and the Every team share what we’ve been building and what’s coming up, then take All Access members’ questions about stuck agents, drifting workflows, and tools they’re unsure about adopting. All Access membership required. Register.

Stop explaining AI to your team and start showing them. Every Agent does the work in Slack, where everyone can see it, so one good workflow quickly becomes the whole team’s.