On August 9, Simon Willison quoted a single passage from Claude Opus 5’s system prompt. It’s short, but it neatly shows how Anthropic handles a tricky problem: what does a model do when important events happened after its training cutoff?
The situation
Fable 5 and Mythos 5 came out on June 9, 2026. Three days later, on June 12, Anthropic suspended access to both to comply with U.S. Department of Commerce export controls. On June 30 those controls were lifted, and on July 1 access was back.
All of that sits past Opus’s training cutoff. The model knows about it only because Anthropic wrote the facts straight into the system prompt – with clear instructions on how to handle them.
The instruction, verbatim
Anthropic tells Opus to confirm the events “accurately and matter-of-factly” if asked – the model shouldn’t deny the suspension happened. And on export controls themselves? Claude should treat them “like any other current political topic”: give a fair, accurate account rather than share personal opinions, and point to Anthropic’s statement for anything further.
There’s also a touch I find smart: because things may have moved on since the prompt was written, Claude should look for newer information when it can search, and otherwise send people to Anthropic’s site.
My take
This is a tiny block of text, but it solves two problems at once. First, the cutoff problem: a model can only speak correctly about events after its training if it gets the facts from outside. Second, the political minefield: export controls are politics, and Anthropic doesn’t want its own model taking a stance.
What’s interesting is how visible this steering is. Anthropic publishes its system prompts, so we can read exactly where a model is being led by the hand. That’s more transparency than most providers offer – and it’s a reminder that a “neutral” assistant is never just there. It gets built. Sentence by sentence.
Sources: