AI progress creates more work for humans, not less. Dive into our new report from @danshipper — and use the companion repo to read it with your agent 👇

Lower AI costs can make more tasks worth delegating to an agent. @trq212 points to an extra code review, another verification pass, or a speculative pull request based on feedback. As those tasks get cheaper, he expects people to use more tokens—and get more work out of their models. Read our Opus 5.5 Vibe Check: every.to/vibe-check/vibe-che…

We asked Opus 5.5 for a presentation in Every’s brand style. It made the deck black instead of green and gave the illustrations a fuzzy texture. @hammer_mt preferred some of those decisions, saying Opus was acting like a human designer bringing their own style to the deck. The illustrations stayed consistent, and the deck still felt like something we’d publish. Our Opus 5.5 Vibe Check: every.to/vibe-check/vibe-che…

You're probably sleeping on computer use. At Every, we’re handing agents chores that once required opening an app and clicking through a series of steps. Computer use lets AI click, type, and navigate those apps for you. Here are 17 ways our team is using it: 1. Fill out school forms. 2. Update six course presentations with new screenshots, assets, and hundreds of small edits. 3. Make video edits. 4. Check every link in a book’s PDF proofs, verify that each destination matches the surrounding text, and collect problems in a spreadsheet. 5. Find an old maintenance request and submit a follow-up about a missing dishwasher. 6. Add kids’ school and camp events to a calendar. 7. Browse iPhone photos, identify items to sell, and create marketplace listings. 8. Clear WhatsApp storage through iPhone Mirroring. 9. Export images from Figma and attach them to posts in Typefully. 10. Manage app builds. 11. Assemble, rig, and repair characters in Blender. 12. Talk to Verizon support about a better plan, reading responses and continuing the conversation while you do something else. 13. Send Slack messages with attachments when the connector falls short. 14. Diagnose and fix a slow computer. 15. Transfer internet service to a new apartment. 16. Request an AI usage report from Slack and get it back in the same thread, using our internal Mac app to hand the task to Codex. 17. Build a Google Slides presentation from reference slides, then turn corrections into saved instructions. You don’t need a big project to see whether computer use helps you. Pick a small chore whose result you can check. Watch the first attempt, review what the agent did, and use that experience to decide what to hand over next. More computer use workflows curated by @lauraentis: every.to/context-window/you-…

.@danshipper wants the main idea in the opening paragraph. In his writing tests, Opus 5.5 often takes too long to state it. He prefers GPT-6’s Sol and Astra for leaner prose that gets to the point. He still calls Opus a strong writer, though that depends on your writing preferences. Read his GPT-6 Sol vs. Opus 5.5 comparison: every.to/vibe-check/vibe-che…

Jev is a new kind of model: fast, cheap, and accurate enough to make judgment calls. You ask it a question about text or structured data and it returns a probability. Is this email urgent? Does this paragraph contain a particular writing tic? @hammer_mt used it to build Jevgram, a tool that scores text against questions such as “Does this sound like AI?” His suggestion: Give a coding agent more specific checks—like looking for “not X but Y” constructions—so it can revise flagged passages as it writes rather than after. What makes it useful is that the feedback is cheap enough to run constantly, not just at the end. Start with a judgment you could make in a few seconds, define what a useful answer looks like, and test whether Jev agrees with you on a handful of examples before running it across the whole collection. Read the full piece: every.to/context-window/how-…

“It’s not faster if you have to retry.” @kieranklaassen found Sol quicker to respond, but Opus 5.5 more reliable at following his Compound Engineering instructions. For long coding tasks, those retries changed which model felt faster to work with. @danshipper breaks down the tradeoffs in our GPT-6 Sol vs. Opus 5.5 Vibe Check: every.to/vibe-check/vibe-che…

.@danshipper on how Astra one-shotted its own vibe-check. When GPT-6 Astra dropped earlier this month, Dan put it to the test by asking it to write a blog post about itself. The result was a pretty balanced article that weighed up the model’s pros and cons, based on information it pulled from Every’s Slack. Since its release, Dan has been using Astra as his writing companion. The sentences it produces are crisp, to the point and rarely have AI-isms. Read the full vibe check: every.to/vibe-check/gpt-6-as…

.@TylerNishida wanted an AI-friendly version of Origami Studio, so he started building one with Opus 5.5. About 4 prompts later, he had a prototype with adjustable animations, templates, and tutorials. He uses Claude Code to add a swiping interaction, then fine-tunes the animation himself. Read our Opus 5.5 Vibe Check: every.to/vibe-check/vibe-che…

.@danshipper on Fable 5.1's leap in idea packaging. When Every's customers call the company "interesting," it turns out that's actually a bad sign. Dan discovered this after asking Anthropic’s Fable 5.1 to analyze a customer survey—and the model both found the insight and explained it back to him in one clean sentence. Communicating an insight precisely is something models generally struggle with, Dan says—even Fable's earlier version and Opus 5 couldn't do it. This is the first time he's seen a model pull it off well. Read the full vibe check: every.to/vibe-check/fable-5-…

The Jev boys strike again.

ATTENDING: @iamadamleeb of Astrohaus will attend Thesis: 2027. Apply to attend Thesis, our inaugural conference on work and AI: every.to/thesis-2027?utm_sou…

Opus 5.5 spent five hours building an interactive lesson for @nityeshaga. It used dozens of subagents and worked across 50 files, designing pages, adding playable examples, and reviewing its output. It paired clearer explanations with interactive demonstrations throughout. Nityesh preferred Opus 5.5 for this task. Read our Opus 5.5 Vibe Check: every.to/vibe-check/vibe-che…

.@hammer_mt calls Opus 5.5 “spikier”: Sometimes it offers a good idea he hadn’t thought of. He likes that it brings opinions to the work. In his testing, that can make it feel like a fellow creative, rather than a model smoothing every rough edge away. More from our team’s Opus 5.5 Vibe Check: every.to/vibe-check/vibe-che…

New models are great, but can they do your work? At Every, we’re building tests from our day-to-day tasks to see which models can handle them. @danshipper explains how we’re building benchmarks. His comparison of GPT-6 Sol and Opus 5.5 uses those benchmarks: every.to/vibe-check/vibe-che…

.@kieranklaassen likes Opus 5.5 on low effort because it leaves blanks for him to fill. The output feels like a sketch. Turn up the effort, and the model adds more detail—but Kieran often wants that space for his own ideas. That’s part of why he enjoys working with the new Opus. Read our Vibe Check: every.to/vibe-check/vibe-che…

Turning up Claude’s effort setting gives it more budget to think, test, and verify. @AnthropicAI's @trq212 points to an HTML sanitizer: With so many ways the code could fail, extra checking is worth the tokens. For coding tasks you know the model handles well, he says low or medium effort can be enough. Our Opus 5.5 Vibe Check: every.to/vibe-check/vibe-che…