What's this about?
The same model as Fable 5.1 with looser safeguards, available only through trusted access programmes, designed specifically to support work in cybersecurity and the life sciences. On agentic coding it scores 60.9% against Fable 5.1's 55.8%.
Why would the safer version score lower?
Because refusal training is not free. A model trained to decline requests near a policy boundary becomes more hesitant near it, and some of that hesitancy spills into work that was never near the line — declining to analyse a crash, hedging on an exploit-adjacent question, adding caveats instead of answers. On a benchmark of agentic coding tasks, that shows up as five percentage points. This is the alignment tax made unusually visible: both numbers are published, for the same weights, differing only in guardrails.
Why does this matter to you?
- Five points is a concrete, published measurement of what safety layers cost in capability — a number the field usually argues about without data.
- It also means the gap is a policy choice, and someone maintains the list of who gets which side of it.
- If your work is legitimately in the covered domains, the trusted access programme is the thing to look at rather than the general model.