Reading up on Opus 5.5
29 deep · digging since sep 23
- Measuring alignment eval tells - METR
METR finds that recent AI models can distinguish alignment from capability evals by detecting fake-looking environments and honeypot task structures.
- TIS-100 Bench: LLMs Push the Pareto Frontier in Programming Puzzle Games
16 LLMs tested on TIS-100 puzzles produced 133 new Pareto frontier points across 18 puzzles, including 3 cycle records, with humans later improving most programs.
- I have a thing for fast trains - Chris Pattle
The author recalls childhood fascination with fast trains, revisits the hobby via AI‑generated 3D models of TGVs and maglevs, and shares the digital recreations.
- How to do valuable work in 2026 - Hugo Montenegro
AI now outperforms humans at coding, making context, product intuition, initiative, and communication the key sources of valuable engineering work in 2026.
- Can Agents Design Libraries for Agents?
LibraryDesignBench evaluates agent‑generated libraries by measuring how much they reduce code for downstream agents, revealing that agents mimic human designs but improve when explicitly guided to optimize for agent usage.
Takes
Please tell me something you asked Opus 5.5 high to do that it failed at.
@davidcrawshaw
ok why the fuck aren’t more people talking about this?? 1. open Claude Code 2. type /model and pick Fable 5.2 3. set your sub-agents to Opus 5.5 (otherwise they copy Fable) 4. give it a big job and add this:
@lennox_saint
Opus 5.5 can build a high-quality, bespoke iOS app in *minutes*. It takes *hours* to get Xcode + App Store Connect + TestFlight to work properly (even with computer use!). And then days to get it reviewed just to share a TestFlight link.
@dwr
We're in the process of renovating our home right now and Opus 5.5 did pretty much all of it Including making a cinematic trailer and magazine feature to get my fiancée on board
@bnj
Andrej Karpathy’s setup from his HackerTube blog today: - “I use Claude Code (desktop app for Mac).” As I’ve been saying, terminal is the wrong interface for coding agents. Desktop apps are the best interface today. - “I use Opus 5.5 on Medium effort as my default.” There’s no point to use the smartest model at the highest reasoning effort for most tasks.
@Yuchenj_UW
You're probably overprompting Claude
@addyosmani
Education doesn’t have to be text and flat diagrams. So I’m building a 3D book about energy that students can play with: 10 lessons to predict, test and prove. Built with Opus 5.5 and Three.js. Try Lesson 1, Energy stores: everyday objects and where their energy is kept ↓
@techartist_
If I was building a new company today I would be building for a world where new use cases are unlocked by a model as good as Opus 5.5 being generally available at >1,000 tok/s What's not possible today that will be possible in 6 months? Then be the best positioned for then
@frantzfries
Claude Duet is our new app where Claude is your proactive agent! Runs on Opus 5.5 across 6,000 connectors/plugins. Duet lets you know what needs your attention and keeps going when you step away. Join the waitlist for the research preview! https://claude.com/product/claude-duet
@addyosmani
The Claude Code desktop app is now better than Herdr if you're only using Opus 5.5. I'm sorry. I wanted Herdr to win, too. But it's time to use the same tool everyone else is using. And they're using it for a very simple reason: it's the best.
@m_chirculescu
Opus 5.5 mapped the entire coast to coast route across the USA for a clicker game I'm playing with. 5,285 km of real road, sampled every 10m from OpenStreetMap.
@jonpaul
before I joined Anthropic, I spent about 2 weeks hacking on this as a side project with Opus 4. It used the Agent SDK so it needed a constantly running process & didnt work that well but one prompt to Opus 5.5 ported it to Claude Managed Agents & made it way more reliable
@trq212
Bye bye Beehiiv 👋 I vibe-coded my own newsletter platform: 📧 Campaign management w/ editor 🖥️ Web archives (https://t.co/pypYdNhdlZ) 🔗 API support 🫧 List cleanup, deliverability reports It took a few prompts with Opus 5.5. I left Beehiiv because they raised prices for the 3rd time in 2 years. - My first yearly bill was $1,000. - My future bill would have been $4,000. Now I'm using AWS SES + MongoDB, and I will pay less than $300/year. Newsletter platforms charge per subscriber, when in fact the real cost is the number of emails sent. Subscribers will naturally become inactive, so you end up paying more for something that hurts your deliverability (sending to dormant subscribers = spam trap). So I've set a cleanup cron job for my newsletter alternative. It marks inactive subs as dormant, sends one last re-engaging email, and stops emailing them. Money saved. Deliverability increased. I also gain a priceless feature: Freedom. I can change the UI, add features, connect all my apps - anything is possible with just a prompt. I'm replacing Resend next. I love it, but it's very expensive ($400/mo)
@marclou
The more I use Opus 5.5, the more I realize how far behind OpenAI models are. Not just coding. Everything. I sent GPT-6 Astra my V02 max lab test results, and it gave me a headache. 200 sentences of: "your report shows this, I think that, therefore it's inconclusive". On the other hand, Opus 5.5 gave me the gist and actionable steps. OpenAI models right now are like that overthinking guy who can't make a decision. Opus 5.5 is that high-agency friend who tells you exactly what you need to hear in one sentence. I don't know how the gap widened so much in the last few months.
@marclou
I asked Opus 5.5 to make an interactive website companion for the latest @AcquiredFM Home Depot episode. It came out pretty nice! Water color illustrations all drawn by Claude, too: https://claude.ai/artifact/RdE6uE7WozcP2MLfx3BDky Episode here: https://www.acquired.fm/episodes/home-depot
@bcherny
Opus 5.5 has been building animatics for me - draft versions of videos I'm about to film. It walks through the actual lessons, creates stills of what it thinks the learner should see, and narrates them with TTS. It's outrageously useful for planning videos (watch on 2x speed, it talks slow):
@mattpocockuk
Claude Sonnet 5.5 is out! We wrote a guide for building with it: • choosing between Sonnet 5.5 and Opus 5.5 • migrating from Sonnet 5 and tuning effort • using it in Claude Code https://claude.dev/blog/building-with-claude-sonnet-5-5/
@ClaudeDevs
S-Tier Agent Management: Engineering manager Opus 5.5 + @DevinAI + @linear We do zero local dev at @HelloUntangle
@ryancarson
GPT 6 Sol vs Opus 5.5 in a nutshell :/
@GaelBreton
OPUS 5.5 IS THE ONLY MOTION DESIGNER YOU NEED this launch video cost $0 and is pure code everyone else on X claims the same shit, then sends you to their website or mcp this is completely free with 0 external tools steal the prompt and try it in your claude ↓ <inputs> Ask me for: a one-word brand name for the wordmark (a verb works best), 9 to 12 high-res photos, a royalty-free song around 120 BPM with a drop and a quiet breakdown (Mixkit, free for commercial use), and a free stock clip of a plain wall with moving plant shadows (Pexels). </inputs> <direction> An Apple-keynote launch film, 2D only, one continuous take. Every scene is made out of the previous one: nothing fades, blurs or cuts. Objects change shape instead: text rises out of a mask line, icons pop from zero on a spring, bars draw across, pages push, and a black shape floods the whole frame and contracts into the next scene. Warm off-white canvas, black UI, iOS 26 liquid glass over the photos. Archivo (wdth 125, weight 800) for the wordmark, Geist for UI. A cursor drives every change with real clicks, drags and long-presses. The camera zooms screen-studio style so each moment fills the square, and the cursor scales with it. Banned: crossfades, blur-ins, brightness "developing", 3D flips, particles, glows, holds longer than 1s, anything that looks like a template. </direction> <structure> 120 BPM, 54 beats, something happens on every beat. Open: the wordmark squeezes into its own period like an accordion, the dot grows into a black pill, a label rises inside it. Click: six iris blades close over the label and snap open onto a photo. The circle becomes a square, shrinks, the grid unfolds from behind it like a paper map (center, plus, corners), reflows into a bento, and a click zooms into one tile, landing exactly on the drop. Glass: a glass word pops in letter by letter, melts into a droplet that stretches into a glass toolbar. The adjust icon turns it into a slider. Dragging relights the photo from day to golden hour (two aligned shots), and the knob turns into a glass lens while held. It lifts into a glass orb, the next photo opens inside it as a circle, and the orb expands into a lock screen: glass clock digits, date, and a home bar that stretches into a glass music player. Stage: the lock screen pulls back into a phone (the bezel grows out of the screen edge). The Dynamic Island stretches like liquid, pinches off, flies over and grows into a Mac window that rolls up like a blind. Long-press the wallpaper, drag it onto a Safari tab, the page pushes in, drop it and it becomes the hero of a landing page. Scroll: the hero morphs into a framed print on a product card, with the mat and molding growing out of the photo's edge. Pick a frame color (it paints across) and a size, then the nav button flies down into "Order print". Order: one black shape keeps morphing: Ordered ✓ → Printing % → On its way (a van on a route) → Delivered ✓. Wall: the delivered circle floods the frame edge to edge, holds black for a beat, and contracts into the framed print hanging on the real wall footage. The same iris opens onto the print, closes again, the frame floods the screen and contracts into the pill → the dot → the letters spring back out, landing on the beat return. Last frame = first frame. </structure> <build> 1. One HTML file, square 1440x1440. Every style is computed from time inside an async seek(t): no CSS transitions, no timers, no state between frames. 2. Springs are closed-form step responses. A value with many targets is the sum of one spring per change, so it stays a pure function of time. 3. Liquid glass: each glass element holds its own clone of the scene behind it, filtered with an SVG feImage displacement map (a rounded-rect distance field) through three feDisplacementMaps at slightly different scales for chromatic edges, plus a rim light. Glass letters: a canvas distance field per glyph gives the map, mask and highlights. 4. Goo: blur + alpha threshold, then composite the source atop it so the glass stays sharp inside. 5. Iris: 6 blades around a hexagonal aperture. Each blade is its two vertices, both edge extensions and the SHORT arc between them. 6. Wordmark squeeze: every letter moves toward the dot by the same factor and its drawn width follows (narrow the wdth axis, scale the rest), so the letters stay touching. 7. Footage: re-encode all-intra (ffmpeg -g 1), load it as a blob URL, await 'seeked' before drawing each frame. 8. Sound: a downloaded SFX for every event (Mixkit), never synthesized, each placed by its measured peak. The song starts on a downbeat: the zoom lands on the drop, the wall sits in the breakdown, the wordmark returns with the beat. Loudnorm to -14 LUFS. 9. Render with Playwright: 4 subframes per frame blended with ffmpeg tmix, 60fps. Check one frame per beat, then scan for single-frame pops (frame-difference spikes 3x their neighbours). </build> <gotchas> backdrop-filter: url() misreads displacement maps in Chromium, so clone the scene instead. A flood must overscale past the corners and take about 0.3s, or half the screen changes in one frame. A child with visibility: visible shows through a hidden parent, so use inherit. Text that swaps inside a morphing shape needs its own mask. python http.server can't range-seek video, so use the blob URL. </gotchas> <start> Ask me for the inputs, then show me the beat map and 4 stills (open, glass, stage, wall) before you write the full film. </start>
@twoclipping
I used this prompt to have Opus 5.5 optimize my setup in Open Minis for iOS. It worked for 12 hours, used 9% of my Claude sub, and slimmed my various prompts to half the size. Very good cleanup, and crazy efficiency in Opus 5.5 right now.
@viticci
holy shit, opus 5.5 is kind of insane at animation. i didn't write a single line of code. it wrote the story, drew every frame, and made the music. no image assets at all, it's all JS. the story is called "small print": claude gets a pile of requests every day. it circles the human part hidden inside them, like "one hand. baby's asleep", and drops it into a jar. at night, those words turn into stars and join up into a constellation. rendered with hyperframes, all in one index.html. it synthesized the music in python, then went through the render second by second and polished it again.
@Voxyz_ai
for the past few months i've been asking our models to paint. opus 5.5 is very skilled at emulating different styles every image here is a python program generated pixel by pixel. there is no image model, and no off-the-shelf art software. instead, it's about 7,500 lines of code using standard libraries to emulate different brush styles. the agents don't use any pictures as reference, instead working only from what they know about each painter
@jkeatn
I asked Opus 5.5 to try a bunch of redesigns of my personal website using my MAX sub, using workflows to iterate and critique. I was really pleased with how it hit the tone I was looking for. Then I asked it to make a trailer with all of its iterations.
@trq212