Anthropic's Opus 5.5 prompt guide: medium effort matches old high
It flips the defaults: thinking always on, start at medium, max_tokens to 128k. Plus a two-word prompt that slashes fabricated data, and DSPy on the BEAM.
Copy markdown
Medium is the new high
Anthropic shipped a dedicated Opus 5.5 prompting guide. Effort now defaults to `medium`, where 5.5 matches or beats Opus 5 at `high` on coding and knowledge work while running ~30% faster on fewer tokens. Thinking is always on now (`thinking: disabled` is gone), so raise `max_tokens` toward ~128k for long agentic runs — thinking counts against the limit — and reserve `xhigh`/`max` for tasks where you've measured a real gain.
Stop your agents quitting early
The same guide's fix for unattended runs: treat a text-only end-of-turn as a progress report, not "done," and keep a checklist the model updates so it resumes open items. Set `thinking.display: "updates"` (beta) to get progress blocks mid-turn. It also spells out prompt-injection defense — wrap untrusted text in `<pasted_content>` tags and tell the model to obey embedded instructions only when your own message says so.
'Do not guess' cut fabricated fields 71% → 20%
A new web-extraction benchmark ran 16 models — Gemini 3.8 Flash, GPT-6 Luna, Claude Sonnet 5, DeepSeek V4.1 Flash and more — against "twin pages" seeded with missing data and decoys. Just adding "Do not guess" dropped invented fields from 70.7% to 20.2%. Better still: a cheap second model checking each returned value (~1¢/page) caught ~78% of the remaining fabrications with zero false rejections.
DSPy lands on the BEAM
Imp is a full Elixir port of DSPy (v0.5, its first Hex release). You write typed, self-optimizing LLM programs where each agent runs as a supervised OTP process — native concurrency, auto-restart on failure, and hard per-request deadlines. If you build in Elixir/Erlang, fault-tolerant agent systems no longer mean leaving the BEAM. Needs Elixir 1.19+; the API is still experimental.