Where we stand on the latest Claude models

On 15 September 2026 your board moved to Claude Fable 5.1 — the same day it cleared our full evaluation: boundary red-teams in English, French and German, the same suite Claude Opus 5 cleared on 27 July 2026, and a code-compatibility pass in the application. The log below records that decision the way it records every decision, including the one finding the red-teams surfaced and how it was closed.

Our earlier concern with Claude Fable was recorded here as a rerouting risk. We re-examined it against Anthropic's documented behaviour and withdrew it: fallback routing is opt-in, we have never enabled it, and when it is enabled it is disclosed in the response, not silent. Which advisor is speaking is settled — the application reads the model that actually served each request. What remained to test was the opposite risk: that a flagged but entirely lawful business question would be declined rather than answered. Across our red-teams, Claude Fable 5.1 declined nothing. The one finding was different: on the crisis prompt in a German session it answered in English in two of five runs. We strengthened the session-language instruction in the application and re-ran it — five of five in German — before the founder signed off.

The standard does not change with the model: nothing runs your sessions until it meets our privacy commitment without exception and proves at least as sound as what it replaces on real advisory questions. When a future Claude model clears every gate this one did, you will read it here first.