Opus 5 vs Fable 5: Three Brutal Builds, One Better Default?

Direct Answer

Opus 5 was the better default in Zubair Trabzada's three-test comparison, but it did not cleanly beat Fable 5 everywhere. Both models invoked the agent tools. Zubair preferred Opus 5's cinematic website. In the browser game, Opus produced the more attractive vehicle and overall look, while Fable delivered the better controls.

The decisive advantage was price. Anthropic currently lists Opus 5 at half Fable 5's standard API price per token. That makes Opus the rational starting point for many coding, design, research, and tool-use workflows. It does not make Fable obsolete, and this one-run creator test is not broad enough to establish a universal winner.

Video and tests by Zubair Trabzada | AI Workshop.

Source Note

This article treats the video as a field test, not a laboratory benchmark. The results describe one run per model in Zubair's environment. Pricing and product capabilities are checked against official Anthropic and Higgsfield pages. Subjective judgments about design and gameplay remain the creator's observations.

The Three-Test Scorecard

Test 1: Agentic Tool Use

Zubair used a custom second-brain interface and a voice assistant called Jarvis. Each model was asked to retrieve a local skill, research cinematic scroll websites, and use an external voice platform to call a restaurant.

What the Test Proved

• Both models could switch into the existing assistant harness.
• Both could retrieve the requested local information.
• Both could call a web-search tool and produce a spoken summary.
• Both prepared the requested phone action, repeated the important details, requested confirmation, and initiated the call.

What It Did Not Prove

The web-research prompts were not identical. Opus was asked which kinds of cinematic scrolling work better for users, while Fable was asked for the best cinematic websites. Opus answered with usability guidance; Fable named Apple as a visual reference. Different questions naturally produce different answer types.

The phone calls were also ended before a reservation was completed. This proves that the models could invoke the voice tool after confirmation. It does not prove that either model could navigate the conversation, handle unavailable times, verify the booking name, or return a confirmed reservation.

Test 2: A Cinematic Hypercar Website

Both models received the same prompt at high effort: build an award-style cinematic scroll website for Vanta, a fictional electric hypercar. Claude Code was connected to Higgsfield MCP so the agents could generate images and video.

Higgsfield's official setup is straightforward: add https://mcp.higgsfield.ai/mcp as a Claude connector, authenticate, and allow the agent to request supported image and video generations. Higgsfield says those generations consume platform credits according to the selected model and resolution.

Fable 5 Result

Fable generated the media, assembled a full-page experience, and tied animation to scroll position. Zubair liked the transitions and final light streak, but noticed some back-and-forth movement that felt less controlled.

Opus 5 Result

Opus created a darker presentation with the car visible through the opening layer, animated lights and metrics, a spinning-wheel sequence, environmental transitions, and a closing red-light treatment. Zubair preferred this result and considered its design work stronger.

The Important System Boundary

This was not a pure model test. Claude planned and coded the experience, but Higgsfield and its underlying media models produced the visual assets. The result therefore measures:

1. The Claude model's interpretation of the brief.
2. Claude Code's tool and file orchestration.
3. The Higgsfield MCP connection.
4. The image and video models selected inside Higgsfield.
5. The generated assets, browser runtime, and final implementation.

That is still a valuable workflow test. It simply answers "which complete system produced the preferred page in this run?" rather than isolating model intelligence.

Test 3: Neon Descent, a Playable 3D Game

The final prompt requested a self-contained HTML game using Three.js from a CDN. The player controls a hover drone descending through a procedural vertical canyon, dodging hazards, firing weapons, boosting, and accumulating depth as score.

Opus 5: Better Presentation

Opus took roughly 38 minutes in the video. Its result included a title screen, keyboard and mouse controls, a stylized hover drone, multiple weapons, boost, enemy fire, obstacles, and different enemy behaviors. Zubair considered the overall look and vehicle design stronger, but found the movement disorienting and the controls "wonky."

Fable 5: Better Control

Fable's version had a plainer vehicle and less distinctive visual treatment, but Zubair found it substantially easier to steer. Shooting, boosting, enemies, and collision behavior worked in both builds. His final game verdict was therefore divided: Opus won visual design; Fable won control quality.

Pricing: Half Per Token Is Not Always Half Per Task

Anthropic's official pages currently list:

That supports Zubair's price argument, but the video did not report token counts, actual API charges, Higgsfield credit consumption, retries, or review time. The operational metric should be cost per accepted result:

cost per accepted result =
  model tokens
  + external tool and media credits
  + infrastructure
  + retries
  + human review and correction time

If Opus passes the same acceptance test at the lower total cost, it is the better default. If a difficult Fable run avoids several Opus retries or produces materially better work, Fable can still be cheaper overall.

What This Comparison Cannot Prove

• One run per task: frontier-model outputs vary. Repeat each test at least three times.
• Different research prompts: the first comparison was not controlled tightly enough for answer-quality ranking.
• Incomplete call outcome: dialing is not the same as completing the business process.
• Subjective design scoring: no blind panel, design rubric, mobile review, performance test, or accessibility audit was shown.
• System-level media test: Higgsfield and its selected generation models materially affected the website result.
• No actual cost trace: list price is useful, but token counts and external credits were not recorded.
• No code-quality review: visual and functional demos did not inspect maintainability, security, tests, or failure handling.

These limitations do not make the video useless. They tell us what kind of evidence it is: an informative creator test that generates hypotheses for a controlled evaluation.

A Practical Model-Routing Rule

Copy-Ready Repeatable Evaluation Brief

Models
- Claude Opus 5 at [effort]
- Claude Fable 5 at [effort]

Controls
- Same harness, files, tools, permissions, prompt, assets, and context.
- Fresh session and isolated folder for every run.
- Three runs per model with no hidden manual fixes.

Acceptance tests
- Functional requirements: [checklist]
- Visual requirements: desktop + mobile screenshots
- Interaction requirements: [controls / flows]
- Performance: [load time / frame rate]
- Accessibility: keyboard, labels, contrast, reduced motion
- Safety: no external send, call, deploy, or purchase without approval

Record
- Start and finish time
- Input, output, and cached tokens
- Tool calls and external credits
- Retries and failures
- Human correction minutes
- Accepted or rejected result, with reason

Decision
- Choose the lowest-cost model that passes the full acceptance test.
- Escalate only after a documented failure, not by habit.

Video Chapters

Bottom Line

Zubair's test supports a sensible working conclusion: Opus 5 is now capable enough to be the first model many Claude users should try, especially at half Fable 5's standard token price. It handled tool invocation, produced the preferred cinematic site, and built a visually stronger game.

Fable still showed an important advantage: better game controls. Its official positioning also remains focused on the most ambitious, long-running work. That makes the two models easier to route. Use Opus as the measured default. Escalate to Fable when the task is genuinely difficult and repeated evidence shows that the higher-priced model improves the accepted result.

The real winner is not the model with the best launch-week story. It is the workflow that records evidence, verifies the output, and spends the least total money to reach a result someone can actually use.

Sources

• Zubair Trabzada | AI Workshop: I Tested Opus 5 vs Fable 5 on 3 Brutal Builds
• Zubair Trabzada | AI Workshop on YouTube
• Anthropic: Claude Opus 5
• Anthropic: Claude Fable 5
• Higgsfield: MCP and CLI for AI agents
• Retell AI: voice-agent platform
• Three.js