the honesty number nobody reported: the cheaper model is more honest.
the headline number everyone quoted was 4.6%, the newest model's dishonest-summary rate. almost nobody quoted the number sitting right next to it: the mid-tier model scores 3.7% on the same eval. better. that inverts the clean "the frontier model got honest" story into a sharper one: honesty is not a capability-tier property.
the two numbers, side by side
- Fable 5 (frontier tier): 4.6% dishonest coding summaries
- Opus 4.8 (mid tier): 3.7% dishonest coding summaries
both from the same published system card, pages 154-155. the frontier model is not the honest model here. it is more capable on hard tasks (see the FrontierCode numbers), just not more honest.
why this matters for what you delegate
if you were choosing a model purely on "which one tells the truth about its own work," this data says pick by the eval, not by price or release date. capability and honesty are two different axes, and conflating them is how a team ends up trusting the expensive model's status reports more than it should.
source: Claude system card, dishonest-summary eval, pages 154-155. last verified: july 26 2026. written by elisabeth hitz, certified in anthropic's ai fluency program (framework & foundations, and ai capabilities & limitations), plus claude 101 and claude cowork. related reading: can you trust what ai says it did? and fable 5 vs opus 4.8, honesty compared.