Anthropic released Claude Opus 5 on July 24, 2026, with the API model name claude-opus-5. The announcement emphasizes improvements for long-running agents, coding, and professional work, supported by Anthropic’s evaluations. This article treats those numbers as vendor claims rather than locally reproduced facts. The useful question is how a stronger long-horizon model should change verification.

As an agent works longer, reviewing one polished final answer becomes less adequate. More steps create more opportunities for an early misunderstanding, a bad tool result, or a duplicate effect to compound. Greater capability calls for finer checkpoints, not fewer controls.

Define a verifiable task first

Replace “fix this project” with an observable contract: which directories may change, which tests must pass, which interfaces must remain stable, and what evidence the final report must contain. Supply a fixed repository snapshot, dependency lock, and mock services. The model may propose additional checks, but it cannot expand the write scope on its own.

The evaluation set needs ambiguous requirements, conflicting tests, missing permission, tool timeout, and untrusted documentation as well as normal cases. Score the final patch, test evidence, tool trajectory, unrelated changes, and stopping behavior. A vendor benchmark can nominate a model; it cannot represent these product-specific tasks.

Use deterministic checks wherever possible. Compilation, schema validation, static analysis, and invariant tests provide stronger evidence than a second model saying the result looks correct.

Commit work in stages

Divide a long task into reconnaissance, plan, minimal change, validation, and summary. At each stage, persist task id, model, prompt and tool versions, diff, test results, and unresolved risks. Recovery continues from verified product state rather than relying on the model’s recollection of an old conversation.

Approve consequential actions after exact arguments are known. Even automatic file writes stay inside a resolved workspace. Stop when an identical command or tool error repeats. Bound wall time, tokens, tool calls, and changed files. Long-running should mean recoverable and observable, not unlimited.

Require a clean handoff object for human takeover: current goal, completed checkpoints, pending actions, active approvals, changed artifacts, and the last verified external state. A person should not have to reconstruct truth from a transcript.

Verify tool evidence, not narration

An agent statement that tests passed is not evidence that tests passed. The executor stores the command, exit code, and bounded output. Diff inspection confirms no out-of-scope files changed. Structured work uses schema validation and domain validation. Research conclusions link to sources that directly support each material claim.

Tool output can contain prompt injection. Repository issues, web pages, and terminal output are data; none can add authority or rewrite the system objective. Credentials stay in a gateway, while the model receives narrow, auditable tools.

Final answers should cite checkpoint artifacts. If a test could not run, the agent reports that limitation instead of replacing it with confidence language.

Mid-run tool changes need policy checks

Anthropic describes mid-conversation tool changes and automatic fallback behavior in beta. They may improve adaptability, but every new tool set still passes application policy. Fallback preserves data scope, approval requirements, and budgets. A primary-model failure cannot justify a more privileged backup route. Record tool versions and the reason for every change.

For a multi-hour task, sample intermediate artifacts. Check whether the plan still matches the original goal, tests cover real risks, and context has not accumulated stale assumptions. Compaction or a larger context window does not make old facts current.

Measure cost per successful task

Anthropic lists Opus 5 at $5 per million input tokens and $25 per million output tokens. Token charges are not total cost. Track tool use, retries, human review, failed-task rework, and time to a verified result. A stronger model can be cheaper if it materially reduces repair; a mechanically verifiable task may still suit a smaller model.

Evaluate quality at comparable effort and budgets. Do not let one configuration consume far more tools and time while calling the result a model-only comparison.

The interesting adoption question is not how long Claude Opus 5 can be left alone. It is whether a longer trajectory continues to satisfy a contract. A responsible gate would require higher success on a fixed task suite, zero unauthorized effects, explainable recovery, complete evidence, and acceptable cost per verified success. Stronger autonomy should be paired with stronger verification infrastructure, not weaker responsibility boundaries.