Technology

Claude Sonnet 5.5 is right here. This immediate stops it from losing your utilization

In Anthropic’s benchmarks, Sonnet 5.5 achieves scores rivaling Opus 5.5 in areas comparable to agentic coding and pc use, though the most recent Opus mannequin remains to be on high by way of “advanced, open-ended work requiring sustained judgement,” Anthropic notes. Nonetheless, Sonnet 5.5 seems surprisingly highly effective and inexpensive in comparison with Claude Opus, and it additionally has the excellence of being the default Claude mannequin without cost customers.

In any occasion, with a brand new Claude mannequin comes a brand new prompting information, and the information for the most recent Sonnet comprises a gem that would come in useful for different AI fashions, notably these able to constructing issues.

The immediate considerations “open-ended” requests, together with people who don’t strictly outline boundaries for what the AI ought to and shouldn’t do. An instance supplied by Anthropic is “present me what you are able to do with this,” with “this” being something from an Excel spreadsheet to a code snippet.

Like ChatGPT and Gemini, Claude can spontaneously begin constructing charts, studies, movies, and different “artifacts” on the fly relying in your request, and an open-ended request might spawn an artifact you didn’t explicitly ask for. For sure, these artifacts will eat into your Claude utilization limits—not nice if you didn’t need them to start with.

In my exams, Sonnet 5.5 was fairly good about holding its replies text-only once I handed it an Excel spreadsheet of made-up working prices and requested “present me what you are able to do with this,” however in at the least one occasion it did spin up a chart by itself.

I’ve beforehand written about prompts designed to maintain AI fashions from leaping the gun, together with one which directs an AI to remain in “pencil-and-paper” mode. However in Claude Sonnet 5.5’s prompting information, Anthropic shares a brand new immediate for dialing down preemptive constructing: