Breaking

Anthropic Publishes Claude Opus 5.5 Prompting Guide Centered on Medium Effort

September 27, 2026 at 07:45 EDT

Foundation Models
AI Agents
Software Dev & Coding

Anthropic has published a model-specific prompting guide for Claude Opus 5.5 that tells developers to start at medium effort, measure their own workloads, and avoid carrying over every Claude Opus 5 habit. The practical change is cost control: Anthropic says Claude Opus 5.5 at its default medium effort matched or beat Claude Opus 5 at high on coding and knowledge-work evaluations while using fewer steps and fewer tokens.

September 2026 | Anthropic

Prompting shifts to the effort dial

Claude Opus 5.5 defaults to medium effort, keeps adaptive thinking on, and asks teams to measure cost and quality before raising settings.

40%

lower typical workload cost than Claude Opus 5

>30%

faster output-token generation

token context window

128K

maximum output tokens

Cache-read price drops to two blocks

Each segment represents $0.10 per million cache-read tokens.

$0.20Claude Opus 5.5

$0.50Claude Opus 5

New prompting loop

Set medium

→

Sweep evals

→

Raise only for measured gains

Lower effort first when latency or cost climbs; use cache-preserving per-message changes where supported.

Upside

Early users report fewer steps, lower token use, and stronger code-review results on long agentic work.

Watchouts

Always-on thinking, rejected forced tool use, and cache invalidation make migration testing important.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.