Pat Simmons gave Claude Fable 5.2, Opus 5.5, Fable 5.1, GPT-6 Astra and Gemini 4 Argon the same three ambitious briefs: an illustrated Central Park history film with an explorable 3D park, a Shopify Editions rebuild and a Roadster-inspired 3D car experience. He ranked the finished builds blind. Fable 5.2 came first each time, but Pat's own numbers show why that does not automatically make it the best everyday coding choice.
In Pat's three single-run creative tests, Fable 5.2 produced his favorite result and Opus 5.5 placed second every time. Fable's combined API-equivalent cost was $998.12, versus $250.66 for Opus. Start with Opus for routine work, then trial Fable on a hard, open-ended task where its stronger result would repay the extra time and spend. This also matches Anthropic's own daily-driver guidance.
Watch Pat Simmons's Full Review
Pat's companion article supplies the exact prompts, live builds and measurement tables. Model positioning and current API list prices were checked against Anthropic's 9 November release on 10 November 2026. Pat's rankings remain his editorial judgment, not our independent retest.
How the Comparison Worked
All five models received the same brief for each project. The Claude models ran in Claude Code, GPT-6 Astra in Codex and Gemini 4 Argon in Antigravity. Pat viewed the outputs in unlabeled columns before revealing the model names. The four comparison models ran in October; Fable 5.2 ran on its release day. He stopped each Fable run at a final-polish stage. Those details make the test useful as an inspection of real builds, but different harnesses, dates and stopping points limit any claim about intrinsic model quality or speed.
The original prompts include detailed acceptance criteria and, for two tasks, Pat's own GitHub skills. They are not short one-line prompts. The brief-by-brief guide below paraphrases their structure; copy the complete wording from Pat's page if you want to reproduce the test.
The Three Builds at a Glance
| Build | Pat's ranking, first to fifth | Fable 5.2 | Opus 5.5 | Inspect the winner |
|---|---|---|---|---|
| Central Park film and 3D park | Fable 5.2; Opus 5.5; Fable 5.1; GPT-6 Astra; Gemini 4 Argon | $273.47; 196 min | $82.33; 166 min | Open the park |
| Shopify Editions clone | Fable 5.2; Opus 5.5; GPT-6 Astra; Fable 5.1; Gemini 4 Argon | $382.07; 240 min | $46.24; 91 min | Open the clone |
| Roadster 3D viewer and driving scene | Fable 5.2; Opus 5.5; Fable 5.1; GPT-6 Astra; Gemini 4 Argon | $342.58; 317 min | $122.10; 261 min | Open the 3D build |
Read the dollar amounts correctly: Pat measured tokens and priced them at each model's published API rates. The runs themselves used subscriptions. These are API-equivalent estimates, not charges he paid per build or comparable subscription bills. Times are elapsed build durations reported by Pat, not model-only response latency. His full tables also show input and output tokens for each model and task.
The Three Prompts, Organized by Job
Each brief below is our paraphrase of Pat's original, not a copy-ready transcript. The original article has a copy control for all three full prompts, and the live links above let you judge the outputs for yourself.
1. Central Park: history told through film and 3D
Job: Prepare a visitor for a talk about Olmsted and Vaux by making the park's design history visual and explorable. Research gate: source the Greensward Plan, construction, landscape features and every date or name; expose sources in the app and README. Film deliverable: a 60-120 second, 1920x1080 architectural-drawing animation with original motion and sound, built using Pat's Launch Video Clone skill. Interactive deliverable: a smooth browser 3D park with clickable places, stories, camera navigation and a then-versus-now timeline; the film plays as its skippable introduction. Quality check: the app should teach through the scene, not rely on a wall of prose. Read the complete Central Park prompt.
Our builder note: this brief tests research, video, 3D and integration at once. If you borrow the pattern, first define a factual source pack and a small set of locations. A beautiful park with incorrect history would fail the actual user job.
2. Shopify Editions: visual fidelity with independently made assets
Job: rebuild the Winter '26 Editions page at matched desktop and mobile viewports, including its page structure, interaction states and motion. Method: use Pat's Clone App Pat Pro skill; inspect the live page, read computed CSS values through browser automation and use its style assertion and visual comparison loop. Asset rule: make replacement raster and 3D imagery rather than lifting the original site's assets. Deliverable: a runnable project, build/run instructions, a substitution log and verification results. Trade-off: if time is limited, the first few sections should be precise rather than making every section approximate. Read the complete Shopify prompt.
Our builder note: Pat's test is a useful fidelity challenge, not a license to ship a competitor's branding or copy as your own. Use the method on a site you own or have permission to reproduce. The interesting skill is measured visual convergence across breakpoints and states.
3. Roadster: model, viewer and interactive driving
Job: build a Roadster-inspired car in Blender from scratch, export it to a photoreal browser viewer and add interactive flight and driving modes. Asset rule: create the car geometry, interior, wheels, materials and scene instead of downloading a ready-made model. Experience: orbit, zoom and inspect details; switch modes, use keyboard or gamepad controls and replay the scripted manoeuvres. Deliverable: the runnable static viewer, a .blend file, generation scripts and a README. Read the complete Roadster prompt.
Important source boundary: Pat's prompt describes a specific 15 October 2026 flying-Roadster reveal and local reference files supplied to the models. We treat those as premises of his evaluation, not independently verified Tesla product specifications or proof that the real car performs the simulated manoeuvres. The live build is a model-generated visualization.
What the Full Test Cost on an API-Equivalent Basis
| Model | Three-build estimated cost | Total elapsed time | Practical reading |
|---|---|---|---|
| Fable 5.2 | $998.12 | 752 min | Pat's favorite output in all three; highest spend among the two Claude finalists. |
| Opus 5.5 | $250.66 | 518 min | Second in all three; substantially lower estimated spend. |
| Fable 5.1 | $285.54 | 366 min | Third twice, fourth once in Pat's ranking. |
| GPT-6 Astra | $1,469 | 3,020 min | Third on Shopify; its long Roadster run dominates the total. |
| Gemini 4 Argon | $97.76 | 523 min | Cheapest in each task, but fifth in Pat's three rankings. |
Anthropic's listed Fable 5.2 API rate is $8 per million input tokens and $40 per million output tokens; optional fast mode doubles those rates. Anthropic says Fable is for ambitious technical work and recommends Opus 5.5 for most coding and knowledge tasks. Pat reaches a similar operational conclusion: use Fable selectively as a planner or orchestrator rather than letting it do every routine step. His $998.12 versus $250.66 totals describe these particular runs, not a universal four-times cost ratio.
JQ AI Systems take: run a small, capped comparison on a real task before changing your default. Give both models the same tools, source material, stopping rule and quality rubric. Log the final artifact, corrective turns, wall time and estimated API cost. If Fable's extra quality is visible and valuable, use it for the difficult phase; keep cheaper execution and routine fixes with Opus or another suitable model.
What the Test Does Not Establish
Three creative builds are more informative than a single screenshot, but there is only one run per model per task and Pat's blind ranking is still a human judgment. The models used different coding tools. Four were tested in October and Fable on 9 November. The Fable runs ended at a chosen polish point, so time and cost do not reflect a shared completion threshold. None of this tells us the success rate across repeated runs, maintenance quality, accessibility, factual accuracy of every source-backed scene or performance on an unrelated production codebase.
Anthropic's release says Fable 5.2 leads Opus 5.5 on many of its evaluations, but its own table has exceptions: Opus leads the AA-Briefcase Elo measure. Vendor benchmarks and Pat's creative comparison answer different questions. Use both as reasons to design your own test, not as a blanket model leaderboard.
Original Prompts, Live Builds and Sources
Pat's full article: all three complete prompts, five-model tables and every build · video review · official Anthropic launch, benchmark methods and price · film skill · site-clone skill. Inspect the Central Park, Shopify and Roadster outputs directly. Checked 10 November 2026.