Claude Opus 5 in the Wild: Brilliant Agent, Slow Expensive Specialist

Direct Answer

Claude Opus 5 looks most valuable when a job is long, visual, tool-heavy, and expensive to get wrong. AI Search's tests show a model that can plan across files, operate software, inspect its own output, and keep working for an hour or more. They also show the cost of that persistence: high token use, long waits, uneven factual depth, and failures on basic visual and medical-image tasks.

Watch the Full Review

Video credit: AI Search. Follow the creator on X, visit AI Search, or read the newsletter.

Evidence and Sponsorship Note

The supplied transcript is the primary source for the demonstrations, run times, token counts, subjective comparisons, and final verdict. These are creator tests in one configured environment, mainly Claude Code at a high effort setting, not controlled laboratory benchmarks.

The Luma Agents section from 11:56 to 13:54 is sponsored. It is kept separate from the Opus 5 evaluation. Official Anthropic model documentation, current pricing, ARC Prize results, and Artificial Analysis were checked on 29 July 2026.

Useful Links

The Test Scorecard

Across the five runs with disclosed usage, the video reports about 1.55 million tokens. That total is useful as a scale signal, not a dollar invoice: the creator used a Claude Max subscription, ran projects concurrently, and did not disclose the input/output/cache split or external API and tool charges.

The Browser Windows Build: Breadth Before Depth

The most visually persuasive test is a browser-based Windows-style environment produced from one broad prompt. Word-like editing and saving worked. The spreadsheet handled formulas such as average and sum. Paint, sticky notes, a code editor, file browsing, settings, and games had interactive behavior.

The weak points matter just as much. PowerPoint lacked basic editing depth. Discord and Slack responses were simulated rather than generated from the actual message. App downloads were representations inside the prototype, not native software installation. This is a strong interactive demo and an excellent prototype, but calling every app "working" would overstate the result.

Where Opus 5 Looked Exceptional

Reference image to interactive 3D

The first result was generic. A second prompt correcting desk placement and excessive glow produced a scene the creator judged substantially better than his Fable comparison. Yet the reconstructed office still had extra furniture, missing desks, misplaced objects, and inconsistent shelves. The success came from vision plus iterative correction, not perfect one-shot spatial understanding.

Financial research to rendered video

Opus 5 found company reports, compared figures, installed HyperFrames, generated a Gemini TTS voiceover, built charts, and rendered the video. This is a meaningful example of agentic production because the deliverable crosses research, code, audio, motion design, and rendering.

One practice in the demo should not be copied: the creator says an API key was pasted at the bottom of the prompt. Put secrets in environment variables or a managed secret store, grant the agent only the permissions it needs, and rotate any key exposed in chat or logs.

Blender and music software control

The Blender test shows a model constructing components, inspecting screenshots, correcting issues, applying materials, animating wing movement, and rendering a result. The music test is similarly impressive as software operation: it located the DAW, rejected a plugin that required registration, downloaded alternatives, created many tracks, added effects, and rendered five minutes of audio.

Both runs also reveal the boundary. The X-wing was a visually convincing creator demo, not a topology, rigging, licensing, or production-asset audit. The song was technically complete but musically repetitive. Agents can remove production friction while taste and professional validation remain human work.

Where Opus 5 Failed

The hidden-frog test is valuable precisely because the model failed. It tiled and enhanced the image, considered a possible snake, rescanned, and eventually admitted it could not identify an animal. The answer was wrong, but calibrated uncertainty is preferable to confidently drawing a fictional target.

The deep-research test also complicates the hype. Opus produced a structured biomedical explanation with charts and tables, yet AI Search found it less thorough and less organized than GPT-5.6 and Kimi K3 for that prompt. That is subjective, but it supports the right procurement rule: evaluate the output format and evidence quality your team needs instead of purchasing the model with the highest aggregate score.

The Medical-Image Test Is a Safety Warning

In the video, Opus 5 misclassified every one of six tumor images. Some answers named the wrong condition; others reported no visible mass. This is not a minor benchmark miss. It demonstrates why a general-purpose model must not be used as an autonomous radiology or pathology system.

• Do not use the output to diagnose, exclude, or prioritize treatment.
• Do not interpret willingness to answer as evidence of medical competence.
• Use validated clinical tools, appropriate imaging protocols, and qualified clinicians.
• For educational analysis, remove personal data and clearly separate research from patient care.

Official Specifications and Price

Benchmarks Need a Timestamp and a Task

ARC Prize reports a 30.2% ARC-AGI-3 score for Opus 5, the highest published result on that evaluation at launch. Artificial Analysis also places Opus 5 near the top of its current composite index. Those are meaningful signals, but neither says whether the model is best for your codebase, report style, latency target, or budget.

Leaderboards change, effort settings differ, harnesses contribute capability, and small score gaps may not translate into material workflow differences. The video's strongest evidence is not a ranking. It is the observed combination of planning, tool use, visual inspection, correction, and persistence across difficult jobs.

When to Use Opus 5

A Better Buying Test

1. Choose three representative jobs. Include one routine task, one difficult task, and one visual or tool-using task.
2. Freeze the environment. Use the same prompt, files, tools, permissions, and acceptance criteria.
3. Set budgets. Cap time, tokens, external API spend, and agent delegation.
4. Test at least two models. Start with the cheapest plausible model and use Opus 5 as the escalation route.
5. Score the finished outcome. Measure correctness, completeness, visual quality, failures, and human rework minutes.
6. Calculate cost per accepted result. Include failed runs, retries, tool charges, and review time.
7. Record the winning task class. Route future work by evidence rather than by one universal default.

Acceptance scorecard
- Core workflow completed: yes/no
- Deterministic checks passed:
- Factual claims traced to sources:
- Visual review passed:
- Human correction minutes:
- Wall-clock time:
- Model tokens and tool costs:
- External side effects approved:
- Accepted for production: yes/no

Video Chapters

Bottom Line

AI Search's review earns its excitement. Opus 5 coordinated unusually broad, persistent workflows and produced several impressive visual artifacts. The same review also supplies the reasons to remain selective: hour-long runs, hundreds of thousands of tokens, uneven application depth, imperfect spatial fidelity, repetitive music, a missed visual target, and six failed medical classifications.

The practical advantage is not "Opus 5 can do everything." It is that Opus 5 can stay with difficult work, use tools, inspect artifacts, and recover from some failures with less supervision. Buy that capability where it changes the accepted result. Route everything else to the lowest-cost model that reliably passes.

Sources

• AI Search: Claude Opus 5 is a freak
• AI Search on YouTube, X, website, and newsletter
• Anthropic: What's new in Claude Opus 5
• Anthropic: Models overview
• Anthropic: Choosing the right model
• Anthropic: Prompting Claude Opus 5
• ARC Prize: Claude Opus 5 results
• Artificial Analysis: Claude Opus 5
• HyperFrames repository
• Blender MCP repository
• Luma: Introducing Luma Skills