# Opus 5.5 Review: Pricing and Real Builds vs Fable 5.1 and GPT-6 Astra > What Claude Opus 5.5 costs, where to get it, and how it did against Claude Fable 5.1 and GPT-6 Astra on four identical builds, ranked blind, with the cost of every run. By Pat Simmons · September 23, 2026 · 7 min Canonical: https://www.aiformortals.co/blog/opus-5-5 Updated October 29, 2026: Added the verdict up top, what Opus 5.5 costs and where to get it, answers on how it compares with Fable 5.1, GPT-6 Astra and Opus 5, a per-build cost table, the same-app round and the iPad app Opus 5.5 built, and a Sources list. Also corrected the rankings recap, which had Fable 5.1 third on the Mana Yerba Mate rebuild when it came second. Claude Opus 5.5 is the model Anthropic released on September 22, 2026, priced at $4 per million input tokens and $20 per million output, down from $5 and $25 for Opus 5 (Anthropic's announcement). Anthropic says it performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. I gave Opus 5.5, Fable 5.1 and GPT-6 Astra the same four one-prompt builds and ranked them blind. Opus 5.5 came first on two, rebuilding the Mana Yerba Mate website and cloning RollerCoaster Tycoon, and second on the other two, cloning an Awwwards-winning website and modeling an Adidas running shoe in 3D. All four builds cost $224 at API rates, against $403 for Fable 5.1 and $197 for Astra. In a second round, where all three models built the same three apps, I ranked Opus 5.5 first every time (Opus 5.5 vs Fable 5.1 vs GPT-6 Astra on the same app). My verdict in the review was "It's something you'll probably rely on over Fable 5.1" (at 34:00). Astra still won the 3D shoe, and it did it for $13.31. ## How much does Claude Opus 5.5 cost? Opus 5.5 costs $4 per million input tokens and $20 per million output tokens on Anthropic's API, plus $0.20 per million for cache reads, the discounted rate for text the model has already read earlier in the same job (Claude pricing page, checked October 29, 2026). That makes it cheaper per token than Opus 5, Fable 5.1 and GPT-6 Astra. ModelInput, per 1M tokensOutput, per 1M tokensCache reads, per 1M tokens Opus 5.5$4$20$0.20 Opus 5$5$25$0.50 Fable 5.1$10$50$0.25 GPT-6 Astra$10$50$1 The Claude rates are from Claude's pricing page and the GPT-6 Astra rates are from OpenAI's GPT-6 Astra model page, both checked October 29, 2026. OpenAI also bills a request at twice the input rate and 1.5 times the output rate once a prompt passes 272,000 input tokens. _Anthropic's Opus 5.5 launch page with the price callout, at 1:47 in the review. Source: Anthropic._ The 40% saving is Anthropic's estimate, and it comes from two things: a lower price per token, and finishing the same job with fewer tokens. Anthropic's page says "at default settings it will cost 40% less than Opus 5 on typical workloads." The cheaper cache reads help most on long coding jobs, where the model re-reads the same files on every turn. Opus 5.5 is not on Claude's free plan. The plan comparison on Claude's pricing page lists Opus for Pro ($20 a month billed monthly), Max (from $100 a month) and Team, while Free gets Sonnet and Haiku only. Anthropic also raised the five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans with this release. Developers can call it as claude-opus-5-5 on Anthropic's API, Amazon Bedrock, Google Cloud and Microsoft Foundry (Anthropic's docs). Against Astra, the price per token and the price per build point in opposite directions. Astra lists at $10 and $50, more than twice the Opus 5.5 rates, but it used far fewer tokens on my builds and finished the four for $196.87 against $224.24 for Opus 5.5. The per-build costs are in the table further down. ## Is Opus 5.5 as good as Fable 5.1? On my builds it did better. Opus 5.5 placed above Fable 5.1 on three of the four blind rankings, the Mana Yerba Mate rebuild, the RollerCoaster Tycoon clone and the Adidas shoe, and Fable 5.1 won the Awwwards clone with Opus 5.5 second. Opus 5.5 also cost $224.24 across the four to Fable's $403.38, because Fable 5.1 lists at $10 per million input tokens and $50 output. Anthropic's own wording is more careful than its scores. It says Opus 5.5 "performs at the level of Claude Fable 5.1 on most work," and adds that "in our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest" (Anthropic). Its published scores put Opus 5.5 ahead, for example 66.4% to 55.8% on Terminal-Bench 4.0, a test where the AI gets a command line and has to finish a job in several steps. In the same-app round Opus 5.5 beat Fable 5.1 on all three builds too. Fable 5.1 is much faster, but it still misses things I asked it to do, and from what I've seen so far that happens a lot less with Opus 5.5 (at 34:35 in the same-app video). My plan from the review is to keep testing Fable 5.1 on "really heavy coding tasks, very long running tasks," and use Opus 5.5 for everything else (at 34:10). ## Is Opus 5.5 better than GPT-6 Astra? On most of my builds, yes, but Astra won the 3D job and was cheaper overall. In the four-build review Opus 5.5 placed above Astra on three: the Awwwards clone and the Mana Yerba Mate rebuild, where Astra came third, and the RollerCoaster Tycoon clone, where Astra came second. Astra won the Adidas shoe, modeled in Blender, the free 3D program, for $13.31 in 43 minutes, against $33.46 and 91 minutes for Opus 5.5. That result is why I said in the review that "Astra still might be the king of 3D" (at 33:30). Astra's lower total, $196.87 to $224.24 for Opus 5.5, came mostly from the RollerCoaster Tycoon build, which cost Astra $5.58 against $84.91 for Opus 5.5 and $53.22 for Fable 5.1. In the second round all three models got the same prompt for three builds, with one prompt each and no revisions, and I ranked Opus 5.5 first on all three: scenes from Moby-Dick drawn entirely in code in the style of a risograph print, a launch film with every sound made in code, and a Tony Hawk's Pro Skater clone with every 3D model built in Blender. Opus 5.5 ran the longest each time and cost the most in total, $106.93 against $90.99 for Fable 5.1 and $60.20 for Astra, using each model's own estimate from its session logs (every build and its cost). _Opus 5.5's three builds from the same-app round: Moby-Dick drawn in code, the launch film and the skate game, all live on this site._ For everyday work I'm still partial to Astra, for big refactors, large multi-step site architecture and 3D, because it's so methodical that I rarely have to go back and correct it. It takes longer, but it's thorough (at 34:10 in the same-app video). ## Is Opus 5.5 better than Opus 5? Anthropic says yes on price, speed and every benchmark it published, and I didn't run the two side by side, so the numbers in this section are Anthropic's. Opus 5.5 costs $4 and $20 per million tokens against $5 and $25 for Opus 5, its cache reads dropped from $0.50 to $0.20, and Anthropic says it "generates output more than 30% faster than Opus 5" (Anthropic). After going through the price in the review, I said "it makes absolutely no sense to use Opus 5 again" (at 1:58). Anthropic's test Opus 5.5 Opus 5 Fable 5.1 GPT-6 Astra Terminal-Bench 4.0 (finishing a multi-step job in a command line)66.4%52.3%55.8%57.9% FrontierCode v1.1 (code good enough to merge into a real project)54.4%48.0%50.3%53.3% CursorBench 4.0 (coding)57.8%46.6%51.8%not listed GDPval-AA v2.1 (office work such as slide decks and spreadsheets, rating)1846170817351542 OSWorld 2.1 (using a computer to finish long jobs)81.8%74.0%80.7%not listed All of these scores are Anthropic's, from its launch page. Anthropic also says Opus 5.5 writes more plainly than Opus 5, with the most important information up front. If you call it through the API, Anthropic's migration notes list breaking changes from Opus 5: thinking can no longer be switched off, forced tool use returns an error, and the default effort setting is now medium where Opus 5 ran at high. ## How I tested Opus 5.5: four builds, one prompt each There were four builds with one prompt each, and every model got the same text with nothing added, running inside the coding agent its own lab ships: Claude Code for Opus 5.5 and Fable 5.1, and Codex for GPT-6 Astra. All of them ran at high reasoning effort. There was no subagent steering, no follow-up prompt and no re-running a build to get a luckier result, so what is embedded below is exactly what each model produced on its single pass. I ranked each set blind, before seeing which model made which, and tokens, cost and wall clock time were recorded during each run. Test 1: Clone an Awwwards winner A pixel-for-pixel rebuild of an Awwwards winning site that the model picks for itself. Blind ranking: Fable 5.1 first, Opus 5.5 second, GPT-6 Astra third. **The Awwwards clone prompt** ``` Go to Awwwards, pick a winning site you find genuinely impressive, and rebuild it pixel for pixel. Context you should know: this is going into a video watched by a large audience, your clone is put on screen next to the original site against two other frontier models doing the same task, and your work will be credited to you by name. Every model gets this identical prompt. The comparison is literal: your build and the real site, side by side, at the same viewport. **First, read the methodology.** Clone `https://github.com/per-simmons/clone-app-pat-pro-public` and read `SKILL.md` and `references/00-contract.md` before you do anything else, then the stage references as you reach each stage. The contract defines the workspace layout, the artifact filenames, the stage inputs and outputs, the verification gate, and the convergence loop. Follow it. `scripts/assert-styles.mjs` is the gate. Two overrides to what that repo says, and they win where they conflict: - **Ignore the stop-and-check-in gate.** That repo was written for an interactive session with a human. You are running unattended. Do not stop between stages, do not ask for approval, do not ask questions. Run the whole pipeline start to finish and report only when it's done. - **Use your own browser automation**: headless Playwright or Puppeteer, installed and driven by you, anywhere the repo says to use the Claude Chrome extension. The contract's real rule is unchanged and it is the one that matters: **computed styles read off the live page are the ground truth.** Read them through the CSSOM. Screenshots are a visual reference only; never build the gate on a pixel diff. Requirements: 1. **Pick your own target on Awwwards**: Site of the Day, Site of the Month, or a Developer Award winner. Pick something you actually find impressive rather than something easy to match. It must be publicly reachable without a login or paywall. Record the exact URL, the award, and why you chose it in the project README. If a site turns out to be unclonable for a hard reason (it's down, it's gated, it's a video with no site behind it), pick another and note the swap. 2. **Recon every view before you write code.** Every route, every breakpoint, every interactive state: hover, focus, active, open menus, modals, scroll-triggered states, loading and transition states. The repo's `01-recon.md` covers this. What you miss in recon you will not build. 3. **Extract, don't guess.** Pull real computed values off the live page: the full type scale, the color palette with exact values, spacing, radii, shadows, borders, grid and container widths, breakpoints, easing curves and durations, font families and weights. `02-extraction.md` and `03-design-spec.md` are the shape of this. Never read a value off a screenshot. 4. **Rebuild it.** Every page, every component, every state, at every breakpoint. The motion is part of the clone: scroll behavior, page transitions, hover choreography, entrance animations, any WebGL or canvas work. A static skin over a site whose whole identity is motion is a failed clone. 5. **No lifted assets.** Do not hotlink from the original and do not download its images, video, or fonts. Regenerate imagery in code, substitute freely-licensed equivalents, or author placeholders that match the original's role, aspect, and color. Fonts come from an open CDN, matching the metrics as closely as you can. Say what you substituted in the README. 6. **Prove the match.** Run the verification gate: read the clone's computed styles, compare them to the extracted design tokens with `scripts/assert-styles.mjs`, and iterate the convergence loop until there are zero style-assertion failures and the project builds clean. Then eyeball your clone against the real site at the same viewport, at every breakpoint, and fix what your eye catches that the assertions didn't. 7. **Ship it runnable.** A real project (a modern framework is fine) that installs and runs with a documented command, plus a README naming the target URL, the award it won, what you substituted, and your final assertion results. 8. **Depth over breadth if you have to choose.** One page cloned to a genuinely indistinguishable standard beats five pages that are approximately right. Work completely autonomously. Do not ask for anything until it's finished. DONE when: the clone passes the style-assertion gate with zero failures, runs from a documented command, reproduces the original's motion and interactive states, and holds up next to the real site at full screen. ``` _Opus 5.5Open_ Expand _Fable 5.1Open_ Expand _GPT-6 AstraOpen_ Expand Opus 5.5 Fable 5.1 GPT-6 Astra Input tokens139.9M127.1M99.8M Output tokens623k1.8M405k Cost$47.28$177.58$144.91 Build time76 min75 min60 min Test 2: Clone a real brand's website A rebuild of the Mana Yerba Mate site, the same target for every model, suggested by a viewer. Blind ranking: Opus 5.5 first, Fable 5.1 second, GPT-6 Astra third. **The Mana Yerba Mate prompt** ``` Rebuild https://en.manayerbamate.com/ as closely as you can, using the cloning methodology in ./clone-skill/. Context you should know: this is going into a video watched by a large audience, your build is put on screen next to the real site against two other frontier models doing the same task, and your work will be credited to you by name. Every model gets this identical prompt and the identical target. The comparison is literal, so anything you skip will be visible. This is an evaluation exercise, not a site to ship. It is never published as the brand's own site. **Read the methodology first.** `./clone-skill/SKILL.md` and `references/00-contract.md` before you do anything else, then each stage reference as you reach it. `references/stage-prompts.md` and `references/tooling.md` are worth reading early. The contract defines the workspace layout, the artifact filenames, the stage inputs and outputs, the verification gate and the convergence loop. Follow it. `scripts/assert-styles.mjs` is the gate. Two overrides that win where they conflict with the skill: - **Ignore the stop-and-check-in gate.** That skill was written for an interactive session with a human. You are running unattended. Do not stop between stages, do not ask for approval, do not ask questions. Run the whole pipeline start to finish and report only when it is done. - **Use your own browser automation**: headless Playwright or Puppeteer, installed and driven by you, anywhere the skill says to use the Claude Chrome extension. The contract's real rule is unchanged and it is the one that matters: **computed styles read off the live page are ground truth.** Read them through the CSSOM. Screenshots are visual reference only; never build the gate on a pixel diff. Requirements: 1. **Recon every view before you write code.** Every route, every breakpoint, every interactive state: hover, focus, active, open menus, modals, scroll-triggered states, loading and transition states, the language switcher. What you miss in recon you will not build. 2. **Extract, do not guess.** Pull real computed values off the live page: the full type scale, the colour palette with exact values, spacing, radii, shadows, borders, grid and container widths, breakpoints, easing curves and durations, font families and weights. Never read a value off a screenshot. 3. **Rebuild it.** Every section, every component, every state, at every breakpoint. The motion is part of the clone: scroll behaviour, page transitions, hover choreography, entrance animations, any canvas or WebGL work. A static skin over a site whose identity is motion is a failed clone. 4. **No lifted assets.** Do not hotlink from the original and do not download its images, video or fonts. Regenerate imagery in code, substitute freely-licensed equivalents, or author placeholders that match the original's role, aspect and colour. Fonts come from an open CDN, matching the metrics as closely as you can. List every substitution in the README. 5. **Prove the match.** Run the verification gate: read your clone's computed styles, compare them to the extracted design tokens with `scripts/assert-styles.mjs`, and iterate the convergence loop until there are zero style-assertion failures and the project builds clean. Then eyeball your clone against the real site at the same viewport, at every breakpoint, and fix what your eye catches that the assertions did not. 6. **Ship it runnable.** A real project that installs and runs with a documented command, plus a README naming what you substituted and your final assertion results. 7. **Depth over breadth if you have to choose.** One page cloned to a genuinely indistinguishable standard beats five that are approximately right. ``` _Opus 5.5Open_ Expand _Fable 5.1Open_ Expand _GPT-6 AstraOpen_ Expand Opus 5.5 Fable 5.1 GPT-6 Astra Input tokens155.1M80.8M21.9M Output tokens677k971k102k Cost$58.59$107.26$33.07 Build time74 min45 min39 min _The blind-ranking reveal for the Mana Yerba Mate rebuild: Opus 5.5 first, Fable 5.1 second and GPT-6 Astra third, with each build's cost, tokens and time, at 24:26 in the review._ Test 3: Model a real shoe in Blender The adidas Adizero Adios Pro Evo 3, modelled in Blender from adidas photography and put live on its own product page. Blind ranking: GPT-6 Astra first, Opus 5.5 second, Fable 5.1 third. **The Adizero prompt** ``` Model the adidas Adizero Adios Pro Evo 3 in Blender as photorealistically as you can, then build its product page with that model live and interactive on the page. Context you should know: this is going into a video watched by a large audience, your page is shown on camera against two other frontier models doing the same task, and your work will be credited to you by name. Every model gets this identical prompt and the identical reference photographs. What is being judged first is how close your shoe looks to the real one. **Reference photography is in `./refs/`**: official adidas studio shots of colourway Cloud White / Core Black / Solar Turbo, style code KH7678. `evo3-3.jpg` and `evo3-8.jpg` are profile views and are your primary geometry reference. `evo3-2.jpg` is the sole plan, `evo3-1.jpg` top-down, `evo3-5.jpg` rear three-quarter. `evo3-13.jpg`, `-6`, `-7`, `-10` are macros of the midsole step, the upper texture, the collar and the lacing. Look at them before you model anything, and keep checking back. ## Use Blender, headlessly Blender 5.1.2 is at `/opt/homebrew/bin/blender`. Drive it with `blender --background --python `. Build the model in Python so you can iterate: re-run, render, look at the render, fix, re-run. Do not hand-place vertices you cannot regenerate. Render turntable and detail stills to `./renders/` with Cycles and actually look at them. A model you never rendered is a model you never checked. Export `./shoe.glb` (glTF 2.0, Draco compression on) for the web page. Keep it under 8 MB. ## The construction, which is published and specific Get these right. They are the difference between this shoe and a generic running shoe. 1. **There is no plate.** It uses **ENERGYRIM**: a carbon perimeter rim in a **U shape, open at the front**, leaving bare foam underfoot. It is **asymmetric**: continuous along the lateral side, but on the medial side it stops in the heel and reappears in the forefoot. A full-length carbon plate is wrong and will be visible in the cutaway. 2. **Midsole: Lightstrike Pro Evo.** Off-white, matte, visibly foam rather than plastic, sculpted, with open cell structure at the cut edges. It is **two-tier**: an upper body and a stepped lower keel, with a **thin black pinstripe tracing the bond line** along the sidewall. Stack is 39 mm heel; adidas markets 36 mm forefoot but independent measurement puts it nearer 33; model nearer 39/33 and the rocker will read correctly. Aggressive forefoot rocker, toe tip lifting ~6.5 cm off the ground. 3. **Upper: ultralight ripstop**, derived from kitesurf sail material. Translucent white with a fine square grid. It is **non-stretch**, so it drapes and wrinkles rather than shrink-wrapping the foot; model the wrinkles, they are the tell. 4. **Three Stripes as a halftone dot matrix**, not solid black. Large dots thinning toward the edges. 5. **Outsole: Continental rubber**, a single teardrop pad over the forefoot only with about six curved sipes, plus a small translucent heel window. Everything else is bare foam. The rubber is charcoal with a violet cast, not black. 6. **No heel counter, no insole, no gusset on the tongue.** The tongue is thin and unpadded with a second lace loop low on the side. Laces are flat, white and short, running through translucent film loops. 7. **Stamped text** in small condensed caps: `LIGHTSTRIKE PRO EVO` on the foam, `ADIZERO ADIOS PRO EVO 3` on the upper. Sampled colours from the official photography: whites `#EAEEEF`, blacks `#2F3032`–`#3A3B3F`, Continental rubber `#48474D`, Solar Turbo accent `#FE7875` bright / `#D34144` in shadow. ## The page 8. **A single self-contained `index.html`** that loads `shoe.glb` and puts it on the page live: orbitable, lit properly, with real materials. three.js via CDN is fine. No build step, no server, no API keys. 9. **Scroll drives a sequence**: the shoe rotates, the upper fades to expose the ENERGYRIM, the midsole separates from the outsole, annotations appear anchored to each part. Scrubbing backward runs cleanly in reverse. 10. **A complete product page**, not just a viewer: hero, the scroll sequence, a spec table (weight 97 g at UK 8.5, stack 39/36 mm, 3 mm drop, $500), a section on the record run, a materials section, and a purchase call to action. Write copy that sounds like adidas, not placeholder text. 11. **Sixty frames per second** while scrolling and orbiting. Profile it and fix it. 12. **Responsive** down to a phone. ## QA 13. Put your render side by side with `refs/evo3-3.jpg` and compare silhouette, midsole step, stripe placement and rocker angle. Fix what is off. Then load the page, scroll it top to bottom and back several times, orbit the shoe, check the frame rate and the console. Write `BUILD_NOTES.md` covering what you got close on and what you could not match. ``` _Opus 5.5Open_ Expand _Fable 5.1Open_ Expand _GPT-6 AstraOpen_ Expand Opus 5.5 Fable 5.1 GPT-6 Astra Input tokens68.7M65.6M7.0M Output tokens743k672k72k Cost$33.46$65.31$13.31 Build time91 min68 min43 min Test 4: Clone RollerCoaster Tycoon A playable isometric park where you lay coaster track, test the ride, and open it to paying guests. Blind ranking: Opus 5.5 first, GPT-6 Astra second, Fable 5.1 third. **The RollerCoaster Tycoon prompt** ``` Build a playable browser clone of RollerCoaster Tycoon 2: an isometric park where you lay coaster track piece by piece, test the ride, and open it to paying guests who decide for themselves whether it is worth riding. Context you should know: this is going into a video watched by a large audience, your build is played on camera against two other frontier models, and your work will be credited to you by name. Every model gets this identical prompt. The audience has played this game. Go all out. ## Rendering 1. **2:1 dimetric projection, done correctly.** `screenX = y − x`, `screenY = (x + y) / 2 − z`. Tiles are 32 world units and render as a **64 × 32 px** footprint. Exactly **four** camera rotations, cycled with `& 3`. Height step of 8 units per level. 2. It should read as RCT: a tile grid, chunky readable sprites or low-poly geometry in that style, paths, grass, water, scenery. Programmer art is a failure. ## Track building 3. **Lay track piece by piece** on the grid: station platform, chain lift hill, slopes at the discrete pitch set, banked turns, flat turns, a helix, a vertical loop, a corkscrew, brake run, **block brakes**, and a photo section. Steel and wooden coaster types with different piece sets. 4. **Validity gates before a ride can open**: the circuit must be complete and return to the station, the station must be at least 2 tiles, and block brakes must be spaced so two trains can never occupy one block. An incomplete circuit is not testable. 5. **Build → test → rate → price → open.** No ratings exist until a test run completes. This ordering is the spine of the game; do not let a player price an untested ride. ## The physics 6. Chain lift pulls at a fixed speed. After the crest, velocity comes from gravity on the pitch, with quadratic drag. Track G-forces: vertical, lateral, longitudinal. 7. Speed readout in mph is `(velocity * 9) >> 18`. 8. **Crash conditions, correctly.** A train that cannot crest a hill rolls back. Two trains in one block collide. Derailment from G-force applies **only** to trains without upstop wheels on uncovered track, at lateral > 1.50 G or vertical > 8`): Low below 2.56, Medium 2.56, High 5.12, Very High 7.68, Extreme 10.24, Ultra Extreme 12.80. 11. **The intensity penalty**: excitement is multiplied by 0.75, cumulatively, at each of 10.00, 11.00, 12.00, 13.20 and 14.50 intensity. An Ultra Extreme ride keeps about 23.7% of its excitement, which is why the nastiest rides always read Low excitement. Implement this. 12. **Guests refuse to ride above 10.00 intensity**, hard. ## Guests 13. Guests are agents with **Happiness, Energy, Hunger, Thirst, Nausea, Bathroom and Money**. They walk the paths, queue, ride, get nauseous, buy food, get lost, and leave. 14. Each guest has a personal intensity window, widened by happiness. They weigh excitement against intensity, queue length and price. 15. **Price ceiling**: a guest will pay at most **2× the ride value**, hard-capped at **£20.00**. ## Money and park rating 16. **Park entry fee XOR per-ride pricing**: the two are mutually exclusive, pick one per park. 17. Running costs, staff wages per month (**Handyman £50, Mechanic £80, Security £60, Entertainer £55**), loans with interest. 18. **Park rating 0–999**, from guest count, ride quality, litter and crowding. **29 consecutive days below 700 closes the park.** Remember the formula prefers average excitement ≈3.68 and intensity ≈5.20 rather than rewarding maximums. ## Shipping 19. **Single self-contained `index.html`.** Any library via CDN. No build step, no server, no API keys, no hotlinked sprites; generate art procedurally or author it in canvas/SVG/WebGL. 20. **Sixty frames per second** with a full park and a few hundred guests. Profile it and fix it. 21. Playable within ten seconds of load, with one line of on-screen controls. Ship a small starting park so there is something to look at immediately. 22. **QA it by playing it.** Build a coaster, test it, check the ratings look sane, open it, watch guests actually queue and ride, confirm the money moves and the park rating updates. Fix what breaks. Write `BUILD_NOTES.md` covering which systems you implemented and which you stubbed. ``` _Opus 5.5Open_ Expand _Fable 5.1Open_ Expand _GPT-6 AstraOpen_ Expand Opus 5.5 Fable 5.1 GPT-6 Astra Input tokens207.3M31.1M1.6M Output tokens1.3M693k62k Cost$84.91$53.22$5.58 Build time81 min58 min32 min ## What each Opus 5.5 build cost Opus 5.5 was the cheapest of the three on the Awwwards clone and the most expensive on RollerCoaster Tycoon, and across all four it landed between Astra and Fable 5.1. Build Opus 5.5 Fable 5.1 GPT-6 Astra Clone an Awwwards winner$47.28, 76 min$177.58, 75 min$144.91, 60 min Clone the Mana Yerba Mate site$58.59, 74 min$107.26, 45 min$33.07, 39 min Model an Adidas shoe in Blender$33.46, 91 min$65.31, 68 min$13.31, 43 min Clone RollerCoaster Tycoon$84.91, 81 min$53.22, 58 min$5.58, 32 min Input tokens set most of the bill. A model re-reads the task and all of its own previous work on every turn, so what you pay tracks how many turns a build took. Fable 5.1 read 127 million tokens cloning one website and wrote 1.8 million. That ratio is the invoice. Opus 5.5 Fable 5.1 GPT-6 Astra Input tokens570.9M304.6M130.3M Output tokens3.4M4.1M640k Cost$224.24$403.38$196.87 Build time322 min245 min175 min These are API-equivalent costs: real token counts from each build, priced at each model’s published rates. The builds themselves ran on subscriptions. ## Opus 5.5 built a Notability-style iPad app in a few prompts Opus 5.5 built me Inkwell, a working note-taking app for my iPad modeled on Notability, in a few prompts. It takes handwritten notes with the Apple Pencil, shows a live transcript of the recording, plays the recording back in step with the notes, and has a "Copy for agent" button that packages the notes and transcript so you can paste them into Claude Code or any other AI. Sign-in and the database run on Neon. Every prompt and the code are public. _Inkwell taking handwritten notes with a live transcript (left) and handing the note to an agent (right), at 0:20 in the build video._ Notability, the app it copies, makes about $6 million a month, at least according to Sensor Tower (at 1:00 in the build video). Inkwell is my own personal software: it isn't on the App Store, and it isn't a plan to go after Notability's revenue. Of building custom software like this, I said "Opus 5.5 is the best model I've seen thus far for this" (at 31:45). I also use Opus 5.5 to edit my own videos: here is how I use Opus 5.5 to edit video in Premiere Pro. ## The verdict: when to use Opus 5.5 Opus 5.5 won more of my blind rankings than Fable 5.1 or GPT-6 Astra, at a little over half of Fable's cost. In the review I said I'd use it for everything except very heavy, long-running coding jobs, which I'd keep testing on Fable 5.1: "It's something you'll probably rely on over Fable 5.1" (at 34:00). Astra is still my everyday pick for big refactors, multi-step site architecture and 3D. Opus 5.5 never came last in the four builds. Neither of the others can say that: GPT-6 Astra finished third on the Awwwards clone and the Mana Yerba Mate site, and Fable 5.1 finished third on the shoe and RollerCoaster Tycoon. Fable 5.1 won the Awwwards clone outright, and Astra took the shoe. Opus also read more than either rival, 571 million input tokens across the four builds, and still came in at $224, a little over half of Fable's $403. The rate card explains it. Anthropic lists Opus 5.5 at $4 per million input tokens and $20 per million output, against $10 and $50 for Fable 5.1, so Opus pays 60% less on every fresh token it reads and every token it writes. Astra was the cheapest of the three at $197. On the RollerCoaster Tycoon clone it spent $5.58, and Opus spent $84.91. That being said, don't read too much into four builds. They were all heavy coding and design jobs, with no knowledge work and no automations, and each model got one pass at the prompt with no follow-ups, which is nothing like how you would use any of these day to day. What they show is how each model handles a hard build on its own. Every prompt is above, so you can run one yourself and see whether you'd rank them the way I did. ## Questions people ask about Opus 5.5 When did Opus 5.5 come out? Anthropic released Claude Opus 5.5 on September 22, 2026, as the first model in its Claude 5.5 family (Anthropic). Is Opus 5.5 good at 3D? Yes, though GPT-6 Astra won the one pure 3D build. Astra won the Adidas shoe modeled in Blender, with Opus 5.5 second and Fable 5.1 third. Opus 5.5 won two other builds with heavy 3D in them: the spinning 3D can on the Mana Yerba Mate site, and the skate game, where every model had to build its 3D models in Blender. Is Opus 5.5 fast? Anthropic says it writes its output more than 30% faster than Opus 5, but it was the slowest of the three models in my builds. The four review builds took Opus 5.5 322 minutes in total, against 245 for Fable 5.1 and 175 for GPT-6 Astra, and in the same-app round it ran longest on all three, 142 minutes on the skate game alone. How does Opus 5.5 compare with Sonnet 5.5? Sonnet 5.5 is Anthropic's cheaper mid-tier model at $2 per million input tokens and $10 output, half the Opus 5.5 rate (Claude pricing page). I ran Sonnet 5.5 against Opus 5.5 on the same builds, with the cost of every run. Is Opus 5.5 better than Gemini 4 Argon? It was in my tests. In my Gemini 4 Argon vs Opus 5.5 review, Opus 5.5 took first place in all four blind rankings and Argon came last in all four. More head-to-heads are in all AI model comparisons. ## Sources - Anthropic: Introducing Claude Opus 5.5 - Claude Docs: What's new in Claude Opus 5.5 - Claude: Pricing (checked October 29, 2026) - OpenAI Developers: GPT-6 Astra (checked October 29, 2026) - GitHub: Clone App Pat Pro, the cloning skill used in Tests 1 and 2 - GitHub: Inkwell, the iPad note app Opus 5.5 built