Take the requirement of building a "fitness tracking app." Opus 5.5 produces a working version on the low tier in 1.5 minutes, but takes 67 minutes on max to deliver one with a heatmap — a 44x gap. Anthropic turned this difference into an official prompting guide. What we want to flag this week: AI companies are no longer just selling models — they're starting to sell the methodology of "how to use them sparingly."

What this is

Anthropic this week released Claude Opus 5.5 and Sonnet 5.5, alongside an official prompting guide. Three things stand out:

First, delete instructions like "think carefully" or "think step by step." The official line is that Opus 5.5 decides for itself how much to think before each response; adding such prompts only slows replies without lifting quality. Second, the new effort parameter (the "how deeply to think" dial) runs from low to max across five tiers, with medium as the default — but official testing shows medium already matches or beats the old Opus 5's high tier on coding and knowledge work. Third, state the goal clearly — including "what counts as done" and "when to stop and ask you."

Put plainly, Anthropic is using an official tutorial to tell you how to use it more sparingly.

Industry view

Bulls argue this is the first time a foundation model company has put the cost-versus-quality tradeoff on the table publicly. Developer Addy Osmani's testing backs that up: on the same task, the high tier burns roughly 20K more tokens (about $0.40, the cost of ~10 retry rounds) than medium — but for work like bug fixes that needs verified boundaries, high is clearly more stable.

Dissent exists. The developer community points out that the medium tier only patches the function itself when fixing code, leaving callers untouched; Anthropic itself concedes that "if the approach is wrong from the start, no amount of tier-bumping will save it." The sharper critique: Anthropic isn't cutting prices — it's teaching you to downshift. The cost of the model being too "eager" is being passed on to users.

Impact on regular people

  • For enterprise IT: Claude procurement used to be token-based billing; now prompt engineering has to manage an extra effort parameter, raising ops complexity.
  • For individual professionals: If you use Claude for coding or research, spend half an hour re-reading your CLAUDE.md (Claude Code's personal config file) and Skills (preset skill bundles) — you may be able to cut spend in half directly.
  • For the consumer market: AI vendors will very likely start pricing by "amount of thinking" — base features get cheaper, deep reasoning stays expensive.