Claude Opus 5 Makes You Choose How Hard It Thinks. Most Companies Never Have.

A marketing coordinator uses the same AI subscription to draft a follow up email and to price out a six figure client proposal. Nobody at the company has ever asked whether those two tasks deserve the same amount of computing power, let alone the same level of scrutiny. Until now, the question rarely came up. One subscription, one model, one level of effort for everything.

On July 24, 2026, Anthropic released Claude Opus 5. The most talked about feature is not a raw performance gain. It is a dial. Users can now choose an effort level, low, medium, or high, task by task. The model also checks its own work and recovers from errors without a human stepping in. Anthropic says Opus 5 matches its most capable model, Fable, on many tasks at half the price, while keeping API pricing level with the previous version, Opus 4.8. Anthropic is also calling it its safest model to date.

That effort dial looks like a technical setting. It is actually a management decision that Anthropic just made impossible to avoid. Most SMEs have never really sorted their AI use cases by what is actually at stake. The same model, the same level of vigilance, handles a LinkedIn post and a contract clause review. As long as capability and cost were bundled inside one subscription, that lack of triage stayed invisible. An effort dial makes it visible. And what becomes visible eventually gets questioned, usually by finance.

Four practical moves worth making.

Map your ten most frequent AI tasks by what actually happens if the output is wrong. A typo in a blog draft does not carry the same weight as an error in a financial model. Low effort for one, mandatory verification for the other.

Decide who owns that call. Do not let each employee pick an effort level on instinct or convenience. That quietly rebuilds the exact fragmentation the company was trying to fix.

Reopen the budget conversation. It is no longer about picking a subscription tier. It is about matching spend per task to the value of that task.

Always pair effort with oversight. Opus 5's self checking is a safety net, not a substitute for a human decision about what needs review. A high effort output on a task that matters still needs a second set of eyes.

This is what the Meaning and Impact dimensions of the MAKIA framework look like in practice, not in theory. Knowing why a task runs through AI at all, and whether the result was actually worth the spend.

The real question this launch raises is not whether to adopt Opus 5. It is whether anyone in your organisation currently decides how much thinking a task deserves. If the answer is no, the model release matters less than the conversation it should force.

Try this week

Pick three recurring AI tasks your team runs. For each one, write one line: what happens if the output is wrong, and who would notice. That is the start of your effort map. Under twenty minutes, and a real basis for deciding.

Sources

• Anthropic debuts Claude Opus 5 with feature that lets users toggle between cost and capability, Fortune
• Anthropic launches Opus 5, TechCrunch
• Anthropic releases new model, Opus 5, Axios
• Anthropic launches Claude Opus 5 with efficiency, safety improvements, SiliconANGLE