Claude Opus 5.

Claude Opus 5 is Anthropic's high-end model for complex agentic coding, long-running work, and difficult reasoning. Our hands-on build suggests it can turn a compact product brief into an unusually complete and visually confident web application. The trade-off is cost: long autonomous sessions can consume a large volume of output and cached context, so teams should watch the meter as closely as the result.

What we tested

We gave Opus 5 a concise brief to recreate an interactive 3D map of Singapore. The requested product needed full-island navigation, 3D buildings, readable roads, address and postal-code search, featured locations including Balestier, manual pins, current-location support, shareable URLs, directions, and a responsive black, cream, and signal-red interface.

The resulting Drop the Red Dot application was built and deployed in roughly 90 minutes. It is a useful test of Opus 5 as a product-building agent because the task required more than code generation: the model had to make visual decisions, connect mapping services, manage browser state, design desktop and mobile layouts, and verify several interacting workflows.

This was a hands-on product test, not a controlled model benchmark. We compared the result with an earlier Codex-built Singapore map, but the two agents did not receive identical iteration time, tools, or feedback.

The result

Opus 5 produced the more adventurous design. The near-black map, cream panels, condensed labels, and signal-red details give the application a distinct field-atlas identity instead of the generic dashboard look common in generated interfaces.

The functional coverage was also strong. The map includes sector shortcuts for Downtown, Jurong, Changi, Woodlands, Punggol, Sentosa, and Balestier. Users can search for places, create a pin at the map centre, use their current location, calculate a driving route, and copy a URL that restores the map view.

In our route test from Marina Bay Sands to Balestier, the application returned a 6.3 km journey, an estimated time of 11 minutes, and 15 written steps. Reloading the shared URL restored the route and its endpoints. At a 390 × 844 mobile viewport, the interface changed to bottom-tab navigation without horizontal overflow.

That breadth is the clearest reason to consider Opus 5. It can maintain a large set of connected requirements across a long session and still deliver a coherent product rather than a pile of disconnected features.

Where it needed human review

The application was impressive, but it was not production-ready without another engineering pass.

Editing a route endpoint could leave the previous route visible, creating a mismatch between the form and the displayed directions. The "Drive here" action could also use the device's location without making that behaviour clear in the button label.

Share links introduce a more serious issue. When a route starts from the user's live location, the URL can contain precise coordinates. The product should warn users before copying that link and offer a privacy-safe way to remove or reduce location precision.

The mapping stack also needs a provider review before real traffic. The interface references OneMap, Nominatim, OSRM, and Valhalla. If it uses the public Nominatim endpoint for client-side autocomplete, that would conflict with the service's public usage policy. A managed, self-hosted, or OneMap-based search service would be safer. OpenStreetMap attribution should remain visible on the mobile map without requiring the user to open a panel.

These are not unusual problems for an AI-generated prototype. They show why autonomy does not remove the need for product, privacy, accessibility, and infrastructure review.

Cost and efficiency

The user-reported run consumed:

At Anthropic's standard API rates, those counters imply an estimated token cost of $52.80 to $57.68. The range depends on whether cache writes used the five-minute or one-hour rate. It excludes any plan allowances, negotiated pricing, regional premiums, fast mode, and separately metered tools.

The counters also need careful interpretation. Cache reads are repeated retrieval of previously stored context, not 65.8 million unique reasoning tokens. The 816-token input figure is unlikely to represent the complete instruction footprint because system instructions, tools, and cached context may be accounted for separately. The 470.9K output figure is aggregate session usage, not the size of the shipped source code.

The practical lesson is simple: Opus 5 can be productive, but a long autonomous run is not automatically economical. Lower effort settings, tighter acceptance criteria, bounded verification, and explicit stopping conditions are important when cost matters.

What Opus 5 is designed for

Anthropic positions Opus 5 for complex agentic coding and enterprise work. The API model ID is claude-opus-5; it has a one-million-token context window, supports up to 128K output tokens, and enables adaptive thinking by default.

That makes it most useful when a task involves several dependent stages: understanding an unfamiliar codebase, planning a change, editing multiple files, using tools, testing the result, and correcting failures. A cheaper model may be the better default for routine edits or high-volume work.

Claude Opus 5 is the model, while Claude Code is one agentic environment that can run it. The quality and cost of a session therefore depend on both the model and the surrounding harness, including its tools, permissions, context management, and verification loop.

Privacy note

Anthropic maintains a public privacy policy, but data handling differs between consumer and commercial use. Consumer users should review their model-improvement and retention settings before entering confidential material. Anthropic states that commercial offerings, including API use, operate under separate commercial terms. Organizations should confirm the policy and configuration that apply to their account rather than assuming every Claude surface handles data in the same way.

Strengths: Strong product and visual judgment, broad requirement coverage, effective long-horizon execution, capable tool use, good responsive implementation, and unusually polished prototypes.

Weaknesses: Potentially high token cost, interaction edge cases, privacy-sensitive share states, third-party service-policy risks, and continued need for accessibility and production engineering review.

Final verdict: Claude Opus 5 is a strong choice for ambitious prototypes and difficult software tasks where outcome quality matters more than minimizing compute. In this test it delivered a credible, distinctive beta in about 90 minutes, but the token bill and unresolved edge cases make human review essential. Use it as a persistent senior implementation partner, not an unattended production pipeline.

Sources and test artifacts: Anthropic's Opus 5 overview, Claude Platform specifications, Claude API pricing, Anthropic privacy policy, Nominatim usage policy, and OpenStreetMap attribution guidance.