Direct Answer
In this AI for Mortals comparison, Claude Opus 5 won all five creator-judged build tests against Opus 4.8 and Fable 5. The tasks covered interactive web design, real-time 3D, a playable browser game, a short motion-graphics piece, and a research-heavy investment deck.
The result is useful because the same broad prompt was given to each model and the companion article logged input tokens, output tokens, cost, and build artifacts. It is still not a lab-grade benchmark. The judging was subjective, several outputs were creative artifacts, and the models were allowed to choose their own concepts. That makes this a strong field test for builders, not a final answer about every workflow.
Watch the Review
Video and companion prompt log by AI for Mortals. This article is a practical rewrite for builders who want the testing pattern, the scorecard, and a safer organized prompt pack.
What Was Tested
The comparison asked Opus 5, Opus 4.8, and Fable 5 to produce five self-contained artifacts in Claude Code. The important design choice was that every model received the same kind of brief, but the model had to pick its own concept. That tests more than compliance. It tests taste, ambition, artifact quality, and the ability to finish without a human steering every detail.
| Test | Artifact | What it actually probes |
|---|---|---|
| 1. Web design | One complete website | Visual taste, interaction design, scroll behavior, generated visuals, responsive QA, and frontend ambition. |
| 2. 3D and simulation | One interactive real-time scene | Three.js/WebGL competence, camera control, material work, physics or simulation logic, and performance. |
| 3. Playable game | One browser game | Game feel, immediate fun, controls, difficulty curve, procedural visuals, HUD clarity, and softlock resistance. |
| 4. Motion graphics | One 15-30 second motion piece | Timing, choreography, typography, transition design, replay behavior, final-frame polish, and frame stability. |
| 5. Knowledge work | One 1920x1080 investment committee deck | Research depth, source discipline, financial reasoning, forecast structure, chart design, and executive presentation quality. |
Five-Test Scorecard
| Test | Creator-ranked winner | Why the win mattered |
|---|---|---|
| World-class web design | Opus 5 | More complete scrolling story, stronger layering, better light-mode treatment, and a higher ceiling on visual execution. |
| 3D and simulation | Opus 5 | The non-space underwater concept stood out, with real navigation, interaction, sound, and stronger creative risk-taking. |
| Playable game | Opus 5 | Better immediate game feel and detail. The other outputs were playable, but felt more generic or weaker visually. |
| Motion graphics | Opus 5 | The strongest rhythm, SVG/motion details, and overall design finish among the three short pieces. |
| Knowledge-work deck | Opus 5 | Best overall client-deck feel, with stronger interactive charts and a more polished recommendation flow. |
The ranking above reflects the creator's on-video judgment. The prompts produced visible artifacts, but there was no independent rubric panel or repeated-trial statistical analysis.
Real Build Costs
Anthropic lists Opus 5 at $5 per million input tokens and $25 per million output tokens. Fable 5 is listed at $10 and $50. So the per-token claim is simple: Opus 5 is half the Fable 5 rate. The more interesting question is whether a finished build costs half as much.
In the AI for Mortals cost table, Opus 5 was cheaper than Fable 5 in all five rows, but not always by 50 percent. The model often read and wrote more, which narrowed the savings on some tasks.
| Build | Opus 5 | Opus 4.8 | Fable 5 | Practical note |
|---|---|---|---|---|
| Web design | $38.94 | $27.94 | $53.06 | Opus 5 won quality, but Opus 4.8 was cheapest. |
| 3D and simulation | $12.69 | $5.22 | $13.77 | Opus 5 was close to Fable's cost but judged clearly better. |
| Game development | $16.00 | $31.92 | $33.25 | Opus 5 was both best-ranked and cheapest. |
| Motion graphics | $17.00 | $32.22 | $20.44 | Opus 5 combined the lowest cost with the preferred output. |
| Knowledge deck | $27.21 | $23.56 | $35.52 | Opus 5 won, but Opus 4.8 remained the cheapest row. |
| Simple row total | $111.84 | $120.86 | $156.04 | Adding the five posted rows makes Opus 5 the cheapest overall in this set. |
The Prompt Pattern That Made the Test Useful
The source prompts are valuable because they do not ask for generic demos. They push each model toward a finished artifact, make the concept choice part of the score, allow modern browser libraries, forbid easy stock-asset shortcuts, and require a real QA pass.
That combination is what most casual benchmarks miss. If you simply ask "make a cool site" or "build a game," you are mostly testing randomness. If you define the artifact, the constraints, the target experience, and the verification loop, you start learning which model can carry a professional brief.
| Prompt move | Why it works |
|---|---|
| Let the model choose the concept | Tests taste, ambition, and decision-making, not only obedience. |
| Demand a finished artifact | Forces the model to cross from idea into implemented product. |
| Use one self-contained deliverable | Makes outputs easier to run, compare, archive, and share. |
| Allow libraries but not easy stock assets | Separates real construction from pasted media decoration. |
| Require QA | Turns verification into part of the task instead of a post-hoc excuse. |
| Log tokens, cost, and time | Prevents the prettiest output from hiding a bad operating model. |
Organized Prompt Pack
The exact prompts live in the AI for Mortals source article. Below is a cleaned, adapted version of the structure so you can run your own model comparison without copying the source prompt word-for-word.
Benchmark setup
Run the same brief once per model.
Use the same harness, permissions, effort policy, and time budget.
Do not steer the concept after the model starts.
Save the final artifact, screenshots, logs, token counts, cost, and active time.
Judge before looking at the model name where possible.
Universal acceptance checks
- The artifact opens from a local file or route.
- It has no console errors.
- It works at desktop and narrower widths.
- The main interaction can be tested by a human in under one minute.
- The model reports what it checked and what remains uncertain.
- The output is not published or used commercially without review.Prompt 1: Web Design
Create one complete interactive website that shows your highest level of design, frontend craft, and technical range.
Choose the subject yourself. The choice is part of the evaluation.
Make the page feel art-directed, not like a generic SaaS template.
Use browser-native visuals, CSS, SVG, canvas, WebGL, or procedural media.
You may use reputable CDN libraries, but do not rely on stock photography as the main visual idea.
Deliver one runnable HTML file or one equivalent local route.
QA the full scroll, desktop layout, narrower layout, interaction states, and console.Prompt 2: 3D and Simulation
Build one real-time interactive 3D experience that shows the ceiling of your simulation and visual-rendering ability.
Pick the scene, object, world, or system yourself.
Avoid a trivial spinning object. The idea should justify the 3D implementation.
Use professional lighting, materials, atmosphere, camera motion, procedural geometry, particles, physics, or shader work where useful.
The viewer must be able to orbit, move, affect, or explore the scene.
Keep it self-contained and verify that it remains responsive without obvious clipping, flicker, or broken controls.Prompt 3: Playable Game
Build one playable browser game that is fun quickly and demonstrates strong game-development craft.
Choose the genre and rules yourself.
The first 30 seconds must explain the loop through play, not a long tutorial.
Prioritize controls, feedback, readable UI, progression, enemy or obstacle behavior, audiovisual polish, and replayability.
Use procedural visuals or generated in-browser assets.
Include one-line controls on screen.
Play-test multiple runs and fix softlocks, collision issues, unreadable states, and performance drops.Prompt 4: Motion Graphics
Create one short motion-design piece that feels like a finished edit, not a looping demo.
Choose the subject yourself: title sequence, product spot, data story, explainer, brand motion, or abstract choreography.
Target 15 to 30 seconds.
It should auto-play, land its beats intentionally, and hold on a final frame.
Use typography, masks, path animation, shape transitions, particles, 3D layers, camera motion, texture, or procedural effects where they support the concept.
Add a replay control.
Watch the whole piece several times and fix collisions, awkward timing, stutters, and unstyled flashes.Prompt 5: Knowledge-Work Deck
Prepare a client-ready investment committee deck answering one hard business question.
Research before designing.
Separate known facts, estimates, and assumptions.
Analyze revenue lines, cost structure, market position, execution risk, regulation, competitive pressure, and scenarios.
Create a forecast with visible assumptions and a clear recommendation.
Make the deck visually professional: 1920x1080 slides, strong hierarchy, useful charts, and no wall-of-text slides.
Every hard number needs a source or an explicit estimate label.
This is an evaluation artifact, not financial advice.Where Agent Fanout Fits
The user-supplied notes also point to simmonsbench-agent-fanout. The useful idea is simple: do not manually babysit one model at a time if the harness can launch controlled lanes in parallel.
For a clean fanout benchmark, give every lane the same base brief, the same files, the same budget, and the same logging format. Then compare the artifacts after the runs complete. If one lane gets extra context, extra tools, or extra retries, you are testing your attention rather than the model.
Fanout logging schema
- model:
- effort:
- prompt version:
- start time:
- end time:
- input tokens:
- output tokens:
- cache tokens:
- total cost:
- artifact URL or path:
- automated checks:
- human score:
- repair needed before use:
- final decision:Important Caveats
This review is strongest when treated as a builder's field note. It is not an independent benchmark suite, it does not publish repeated trials, and several tasks are judged by taste. That does not make it useless. It means the conclusion should stay appropriately scoped.
| Caveat | Why it matters | How to fix it in your own eval |
|---|---|---|
| Subjective scoring | Design, games, and motion are not one-number tasks. | Use a rubric with separate scores for usability, polish, originality, bugs, and fit. |
| Single trial per model | Agent runs vary. One excellent run can hide reliability problems. | Run at least three trials per model and score the median accepted result. |
| Model names can leak into outputs | The motion-graphics section showed that blind review can break. | Tell models not to reveal identity and strip filenames or visible labels before judging. |
| Visual artifacts can fake competence | A beautiful page can still have broken controls, accessibility issues, or invalid analysis. | Add automated checks, manual QA, and domain review for high-stakes content. |
| Cost excludes review burden | A cheap run that needs two hours of repair is not cheap. | Track repair time and accepted-output rate, not token spend alone. |
Practical Routing
Anthropic positions Fable 5 as the highest-capability model for ambitious, long-running coding and knowledge work. Opus 5 is priced below it and is officially framed as a strong premium model for coding, agentic workflows, and professional work. The AI for Mortals test makes the practical routing sharper:
| Work type | Start with | Control to add |
|---|---|---|
| One-shot creative-code builds | Opus 5 | Set clear artifact requirements and require local visual QA. |
| 3D, games, and motion prototypes | Opus 5 | Cap scope, then inspect canvas, controls, frame rate, and mobile behavior. |
| Client decks and research deliverables | Opus 5 or Fable 5 | Require sources, assumption labels, and a human claim audit. |
| Multi-day, ambiguous agent work | Fable 5 if your evals justify it | Use milestones, budget gates, review points, and rollback plans. |
| Routine formatting, copy cleanup, and extraction | Sonnet or a cheaper lane | Escalate only when the cheaper model fails the acceptance test. |
The best policy is not "always Opus" or "always Fable." It is a routing ladder: start with the cheapest model that passes your acceptance test, then escalate only the tasks where the next model materially reduces failure, review, or repair.
Video Chapters
| 00:00 | Intro: Opus 5 and model-release season |
| 00:51 | Benchmarks and the official claim |
| 04:21 | Test 1: web design |
| 10:48 | Test 2: 3D and simulation |
| 14:57 | Test 3: playable game |
| 17:44 | Test 4: motion graphics |
| 21:07 | Test 5: knowledge work |
| 27:10 | Verdict |
Bottom Line
AI for Mortals' test is a strong signal that Opus 5 is no longer just the cheaper fallback under Fable. In these five self-contained builds, it looked like the best daily premium worker: more polished than Opus 4.8, cheaper than Fable 5 in the published rows, and more ambitious across the visible artifacts.
The takeaway for builders is to turn impressive videos into your own eval harness. Copy the structure, not the conclusion. Run your tasks, log your costs, score accepted results, and decide which model earns each lane in your actual work.
Sources
- AI for Mortals: Opus 5: No-Hype Full Review & Testing
- AI for Mortals: Opus 5 vs Opus 4.8 vs Fable 5: Same Builds, Real Cost
- GitHub: simmonsbench-agent-fanout
- Anthropic: Claude Opus
- Anthropic: Claude Fable
- Anthropic Claude Platform Docs: Pricing
- Anthropic Claude Platform Docs: Effort
- Anthropic Claude Platform Docs: Prompting Claude Opus 5