Racing Astra and Fable 5.1 Across the Same Four Builds: Surprising Results

Published after a packed week, the video tests two fresh releases: Astra arriving days after the Fable 5 announcement, with both models trialled across a full weekend. The method stays constant: identical prompts to each model, the same tool access on both sides, and zero intervention once the run starts. All four tasks come from the host's daily work: building sites and prototypes fast. Each round closes with build time and an estimated token bill computed at API rates.

Task one asks both models for an online store for a ski clothing brand evoking a famous label. The prompt stays deliberately thin: branding, colours, products, and all product imagery are left to the model, with two constraints that every product needs a human model and all visuals come from an external AI image-video service. Fable 5.1 runs on its highest effort setting, and Astra gets the same prompt with the same settings.

Both stores come back as complete flows: a moving hero video, a model-invented product line, colour-size-quantity selection, cart, and a checkout page. On the Fable side the product backdrops clash and one photo crops a face out, but a 3D-library interactive section adds flair. The Astra side looks tidier: consistent studio-like backdrops and an auto-generated size guide stand out. The bill surprises: Fable finishes in about 10 minutes at an estimated 4.70 dollars while Astra takes 23 minutes at more than double. Similar output at higher speed and lower cost gives Fable this round.

Task two grows from a personal need: a 3D explorable city model of Kyoto, whose temples the host wants to preview before a first visit, instead of a flat 2D map. The Fable output impresses: zoom, pan, tilt, a sidebar of popular spots, faithfully modelled landmarks such as Kyoto Tower, garden-food-temple filters, fly-to-point navigation, an automatic city tour mode, morning-day-golden-hour-night lighting, and a two-point measuring tool (the demo reads 2.8 kilometres, 35 minutes on foot). Only the attraction-dense district is modelled, not the whole city.

Astra builds a smaller yet comparably detailed Kyoto: close-up geometry down to stone steps, decorative extras such as a sculpture, the same lighting and measuring tools, plus clay and wireframe views and an automatic place-to-place tour. One pattern stands out: with no colour direction the model defaults to a green palette, fixable with a single follow-up prompt. Timings run close (50 minutes for Fable, 40 for Astra) but the estimated bills diverge sharply: 146.80 dollars against 20. The reason is visible on screen: Fable built the bigger city and burned more tokens. Usability looks strong on both; the design vote goes to the audience.

Task three is the hardest cloning job: rebuilding the polished landing page of a desktop Mac dictation app, adapted to a voice dictation product called Whisper. The prompt hands over two addresses: the design reference and the product page whose copy gets adapted. The reference packs a moving nature hero, two stones that bloom a plant on hover, a tutorial video, feature icons, scroll animations, screenshots, and a pricing section.

The Fable clone tracks the reference faithfully: a moving grassy hero, the hover-to-bloom stone effect, a rebrand to Whisper, a coded product demo, service-generated videos in the feature blocks, matching icons, screenshots, testimonials, and an animated pricing area. The Astra version feels more cohesive: a fog-mountain-bird hero composition, a smoother cursor-tracking bloom effect, a motion-graphics tutorial video produced without being asked, clean animations, and a stylish video footer. Its one miss is swapping pricing for an early-access waitlist box. Time and bill: 35 minutes and 48 dollars for Fable, about an hour and 57 for Astra. Slightly pricier, but the quality bump takes this round for Astra.

Task four demands a full mobile build from one prompt: a clone of a receipt-scanning app said to earn 60 thousand dollars a month, plus its matching landing page. The app scans paper receipts, extracts line items, categorises spending, and connects to accounting software; receipt reading runs through the Claude programming interface, the design direction asks for Notion-style illustration, and store-ready app screenshots are requested too. Testing happens over Expo Go with a mirrored phone.

On the phone, the Fable app works but looks plain: bundled sample data, receipt-report-settings tabs, camera capture of a real receipt, interface-driven extraction with correct categorisation, editable fields with currency choice, and database save. The design makes no claim, one filter looks slightly broken, yet every requested function runs from a single prompt. The Astra app ships without sample data but looks cleaner and more considered: tidy buttons, consistent colours, scan from photo or camera, line-item correction, a review-and-save flow, and haptic feedback on the tabs. The design vote goes to Astra.

The landing pages flip the table. The Fable page looks ready to ship: an iPhone-framed app visual, an animated hero, Notion-style illustration, six or seven store-format screenshots, service-generated imagery, and feature-testimonial-pricing-FAQ sections. The Astra page is clean but thinner: a beige backdrop, a generic non-iPhone phone frame that misses the iOS detail in the prompt, and only four screenshots. Bills run close: 86 dollars in 36 minutes for Fable, 59 in 49 for Astra. The verdict splits: landing page to Fable, the app itself to Astra.

The host refuses a single-model switch and picks by task: Astra for 3D work at similar quality with better efficiency, Fable out of habit for design, Astra for multi-round knowledge work and browser-computer takeover jobs where it runs quicker. One open thread is noted: video editing, where the Fable model struggled to cut mistakes and pauses cleanly. The host plans to test Astra there next in a separate video. All figures in the video are estimates at API rates; on a subscription the quota meter applies instead.

To steelman the other side: viewers tempted to switch subscriptions over one creator's round score should slow down, because the headline's 'not even close' claim clashes with its own scoreboard. Four rounds split across the models: Fable takes the store, Astra takes the clone, mobile divides page and app between them. Independent write-ups back the split: Astra leads on several benchmarks while Fable answers with design fidelity and cache-price cuts. The honest picture is not a blowout but a task-by-task map of strengths.

The gaps all sit in the method: a single run, no repeats, no messy multi-file repository work, verdicts on first versions only, no iteration depth measured. Bills are estimates at API rates; on a subscription the quota meter rules, and tariffs move monthly. The security dimension never opens: handing a model a programming-interface key, a media-service account, and file access, plus the licence status of a cloned page, all go unasked. Single-sample quirks such as the model's drift to green with no colour direction get generalised too fast.

Three items need independent re-checks at decision time. First, prices: the Anthropic side announced roughly a quarter off typical workloads with Fable 5.1 plus a deep cache-read cut, while press coverage cites 10 and 50 dollar input-output tariffs on the Astra side. Second, capability claims: alliance research places the Astra bar near a critical cyber threshold, which prices the video's 'hand over the browser and let it finish' praise with a security bill. Third, interests: the prompt-pack link on the page and vendors' self-published benchmarks make second-source confirmation mandatory for every figure that matters.

My practical read: teams shipping storefronts and launch pages who need brand fidelity should stay on the Fable side; teams doing 3D scenes, fast prototypes, and multi-round knowledge work should budget an Astra trial; on mobile, combine the two and take the page from Fable and the app from Astra. I would never paste a sensitive key into a trial prompt, I would check the reference licence on clone jobs, and I would recompute every figure from the video against current tariffs on decision day.

astra · fable 5.1 · vibe coding