A Tale of Two Models

Anthropic put out two models this week—one we’d been missing, and one we could skip. Fable 5 came back online on Thursday, and it couldn’t come back soon enough. Two days earlier, Sonnet 5 landed with a shrug: Katie Parrott’s Vibe Check finds a Goldilocks model pitched for everyone that impresses no one, with a cheaper, faster, or smarter option for nearly every job. Yet while Fable can spin up a working app from a single prompt (it rebuilt our document editor Proof in about three hours), AI still can’t reliably make a PowerPoint deck. Mike Taylor and the consulting team needed a 24-skill pipeline at $62 a deck to automate it, and still wouldn’t recommend it for most teams. Regardless, the Every team still can’t get enough of Codex. We’ll be back in your inbox on Tuesday.—Kate Lee

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

Knowledge base

“Vibe Check: Sonnet 5—A Model Pitched for Everyone Impresses No One” by Katie Parrott/Vibe Check: Katie and the Every team put Anthropic’s new Sonnet 5 through its paces and came away unconvinced. Pitched as the Goldilocks model—smart enough for hard work, cheap and fast enough for daily use—it lands as none of those next to Opus 4.8, Fable 5, and GPT-5.5. Read this for where Sonnet 5 fits, and why the team keeps reaching for other models.

“AI Could Do Anything. Then It Met PowerPoint.” by Mike Taylor/Also True for Humans: PowerPoint has been a fixture of work for decades, and AI still can’t conquer it in one shot. Mike and the consulting team found that Codex and Claude Code build slides from scratch well but falter against a company’s own templates. For most teams, the hardest part is figuring out what to say, and AI can’t do that for you. Read this for an honest account of what AI can and can’t do with slides.

“Codex in Practice” by Laura Entis/Context Window: Laura Entis rounds up how people across Every have built their own Codex workspaces—Dan Shipper’s long-running router threads, Katie’s file system, head of growth Austin Tedesco’s outcomes-based approach, and Cora general manager Kieran Klaassen’s portable setup, a synced context folder any agent can draw on—each with a copyable starter prompt. Read this for setups you can steal.

“Codex for Nontechnical Builders” by Dan Shipper/AI & I: Natalia Quintero, Every’s head of consulting, makes the case that Codex is the first agent a nontechnical person can operate the way engineers operate Claude Code. It builds its own folder structure and instructions instead of asking you to set them up first. She walks Dan through running it like a direct report: achieving inbox zero, managing a client pipeline in Attio, and coordinating her father’s medical care. Read this for a nontechnical builder’s Codex practice, end to end. 🎧 🖥 Listen on Spotify or Apple Podcasts, watch on YouTube, or follow the discussion on X.

“Your AI Strategy Is Making Bets. Do You Know Which Ones?” by Dan Pupius/Thesis: Dan Pupius, CTO at The General Partnership, argues that every AI strategy rests on four bets founders usually don’t spell out: token costs, model capability, provider lock-in, and regulation. The teams that name their bets are the ones who can adjust when one turns. Read this for a clear way to surface your assumptions and see which you can change.

Log on

Get hands-on with how Every uses AI. These are the live camps, workshops, and meetups where team members teach the workflows behind our work.

Upcoming events

• Q2 Demo Day (July 10): Every’s quarterly demo day, where the team shows what it shipped this quarter—paid subscribers only. RSVP.
• Every IRL (July 15): An in-person meetup at Every’s Brooklyn headquarters, paid subscribers only, from 6-8 p.m. ET. RSVP.

From Every Studio

Monologue dictates in every language you speak

Monologue, Every’s voice dictation app, shipped v1.3.0 with multilingual dictation: Tell it which languages you speak—English and Spanish, say—and it keeps up as you switch mid-thought, with a new picker spanning more than 99 languages. The update also adds more ways to start recording, including Hyper Key support and dedicated push-to-talk, hands-free, and mouse-button shortcuts.

Spiral prompts automate the writing you repeat

Spiral, Every’s writing tool, lets you save a repeated writing workflow as a reusable prompt, so jobs like generating show notes, pulling quotes from a podcast, or drafting a marketing post from internal documents run in one step. New this week: You can create and edit those prompts right in chat or over MCP, instead of setting them up separately.

Alignment

Sell the shovels. Anthropic launched Claude Science—a desktop research tool that lets scientists run analyses, visualize molecular and genomic data, and show the exact code and steps behind every result— and announced it’s running its own preclinical drug programs, which likely had pharmacology executives swearing at their laptops.

However, as a doctor named Patrick Malone tweeted, it’s unlikely that Anthropic will develop its own drugs and see it all the way through to the hands of a patient.

Instead, Anthropic will use these internal drug programs to test and improve its own AI tools. The goal is to fix the major bottlenecks in drug development—particularly evaluation and verification—so it can build a stronger platform that pharma companies will bend over backwards to use.

As Malone points out, this is known as dogfooding—deliberately using your own technology on real, difficult problems so you can identify and fix their weaknesses. In this instance, dogfooding will be used to test how well an AI performs on complex tasks like identifying drug targets or designing molecules, and then verifying those results by conducting laboratory experiments and clinical studies to confirm whether the AI’s suggestions are indeed correct.

These hurdles are time-consuming and expensive to resolve because, unlike software, where you can test something in seconds, biological systems move slowly, and feedback only comes after living cells or organisms have had time to respond.

If—and it’s a big if—Anthropic can solve this problem, it builds the workflow layer that ties data, models, experiments, and decision-making together, and, by extension, sells big pharma the platform that makes drug development faster. This strategy is known as selling the shovels, and it’s been a good one since 1849.—Ashwin Sharma

That’s all for this week! Be sure to follow Every on X at @every and on LinkedIn.

We build AI tools for readers like you. Write brilliantly with Spiral. Organize files automatically with Sparkle. Deliver yourself from email with Cora. Dictate effortlessly with Monologue. Work on documents with AI agents using Proof.

For sponsorship opportunities, reach out to [email protected].

Anthropic put out two models this week—one we’d been missing, and one we could skip. Fable 5 came back online on Thursday, and it couldn’t come back soon enough. Two days earlier, Sonnet 5 landed with a shrug: Katie Parrott’s Vibe Check finds a Goldilocks model pitched for everyone that impresses no one, with a cheaper, faster, or smarter option for nearly every job. Yet while Fable can spin up a working app from a single prompt (it rebuilt our document editor Proof in about three hours), AI still can’t reliably make a PowerPoint deck. Mike Taylor and the consulting team needed a 24-skill pipeline at $62 a deck to automate it, and still wouldn’t recommend it for most teams. Regardless, the Every team still can’t get enough of Codex. We’ll be back in your inbox on Tuesday.—Kate Lee

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

Knowledge base

“Vibe Check: Sonnet 5—A Model Pitched for Everyone Impresses No One” by Katie Parrott/Vibe Check: Katie and the Every team put Anthropic’s new Sonnet 5 through its paces and came away unconvinced. Pitched as the Goldilocks model—smart enough for hard work, cheap and fast enough for daily use—it lands as none of those next to Opus 4.8, Fable 5, and GPT-5.5. Read this for where Sonnet 5 fits, and why the team keeps reaching for other models.

“AI Could Do Anything. Then It Met PowerPoint.” by Mike Taylor/Also True for Humans: PowerPoint has been a fixture of work for decades, and AI still can’t conquer it in one shot. Mike and the consulting team found that Codex and Claude Code build slides from scratch well but falter against a company’s own templates. For most teams, the hardest part is figuring out what to say, and AI can’t do that for you. Read this for an honest account of what AI can and can’t do with slides.

“Codex in Practice” by Laura Entis/Context Window: Laura Entis rounds up how people across Every have built their own Codex workspaces—Dan Shipper’s long-running router threads, Katie’s file system, head of growth Austin Tedesco’s outcomes-based approach, and Cora general manager Kieran Klaassen’s portable setup, a synced context folder any agent can draw on—each with a copyable starter prompt. Read this for setups you can steal.

“Codex for Nontechnical Builders” by Dan Shipper/AI & I: Natalia Quintero, Every’s head of consulting, makes the case that Codex is the first agent a nontechnical person can operate the way engineers operate Claude Code. It builds its own folder structure and instructions instead of asking you to set them up first. She walks Dan through running it like a direct report: achieving inbox zero, managing a client pipeline in Attio, and coordinating her father’s medical care. Read this for a nontechnical builder’s Codex practice, end to end. 🎧 🖥 Listen on Spotify or Apple Podcasts, watch on YouTube, or follow the discussion on X.

“Your AI Strategy Is Making Bets. Do You Know Which Ones?” by Dan Pupius/Thesis: Dan Pupius, CTO at The General Partnership, argues that every AI strategy rests on four bets founders usually don’t spell out: token costs, model capability, provider lock-in, and regulation. The teams that name their bets are the ones who can adjust when one turns. Read this for a clear way to surface your assumptions and see which you can change.

Log on

Get hands-on with how Every uses AI. These are the live camps, workshops, and meetups where team members teach the workflows behind our work.

Upcoming events

• Q2 Demo Day (July 10): Every’s quarterly demo day, where the team shows what it shipped this quarter—paid subscribers only. RSVP.
• Every IRL (July 15): An in-person meetup at Every’s Brooklyn headquarters, paid subscribers only, from 6-8 p.m. ET. RSVP.

From Every Studio

Monologue dictates in every language you speak

Monologue, Every’s voice dictation app, shipped v1.3.0 with multilingual dictation: Tell it which languages you speak—English and Spanish, say—and it keeps up as you switch mid-thought, with a new picker spanning more than 99 languages. The update also adds more ways to start recording, including Hyper Key support and dedicated push-to-talk, hands-free, and mouse-button shortcuts.

Spiral prompts automate the writing you repeat

Spiral, Every’s writing tool, lets you save a repeated writing workflow as a reusable prompt, so jobs like generating show notes, pulling quotes from a podcast, or drafting a marketing post from internal documents run in one step. New this week: You can create and edit those prompts right in chat or over MCP, instead of setting them up separately.

Alignment

Sell the shovels. Anthropic launched Claude Science—a desktop research tool that lets scientists run analyses, visualize molecular and genomic data, and show the exact code and steps behind every result— and announced it’s running its own preclinical drug programs, which likely had pharmacology executives swearing at their laptops.

However, as a doctor named Patrick Malone tweeted, it’s unlikely that Anthropic will develop its own drugs and see it all the way through to the hands of a patient.

Instead, Anthropic will use these internal drug programs to test and improve its own AI tools. The goal is to fix the major bottlenecks in drug development—particularly evaluation and verification—so it can build a stronger platform that pharma companies will bend over backwards to use.

As Malone points out, this is known as dogfooding—deliberately using your own technology on real, difficult problems so you can identify and fix their weaknesses. In this instance, dogfooding will be used to test how well an AI performs on complex tasks like identifying drug targets or designing molecules, and then verifying those results by conducting laboratory experiments and clinical studies to confirm whether the AI’s suggestions are indeed correct.

These hurdles are time-consuming and expensive to resolve because, unlike software, where you can test something in seconds, biological systems move slowly, and feedback only comes after living cells or organisms have had time to respond.

If—and it’s a big if—Anthropic can solve this problem, it builds the workflow layer that ties data, models, experiments, and decision-making together, and, by extension, sells big pharma the platform that makes drug development faster. This strategy is known as selling the shovels, and it’s been a good one since 1849.—Ashwin Sharma

That’s all for this week! Be sure to follow Every on X at @every and on LinkedIn.

We build AI tools for readers like you. Write brilliantly with Spiral. Organize files automatically with Sparkle. Deliver yourself from email with Cora. Dictate effortlessly with Monologue. Work on documents with AI agents using Proof.

For sponsorship opportunities, reach out to [email protected].