
We beat Mythos and Fable 5 in <36 hours.
There's been endless hype around Mythos' capabilities since its announcement, so much so that Anthropic won't release it directly and instead publicized Fable 5, a safety-capped version of the same model.
So once Fable 5 was out, we @sentra_app knew we had to test just how wide the gap was between it and what was already available.
We took GPT-5.5, the current leading model on Terminal-Bench 2.1, and changed one thing: we gave it Sentra's Code Memory.
The result:
- A score of 88.31%, ahead of Fable 5 (80.5%) and Mythos 5 (88.0%).
- 3.65x cheaper than the same model without Sentra. Accuracy up, cost down ~72.6%, tokens down 41.2%.
All from one change: a task-scoped memory layer that keeps the agent from re-reading context it already has.
We're now submitting through Terminal-Bench's official verification and will release every agent trajectory once verified. Benchmarks on @datacurve DeepSWE are coming soon.

Claude@claudeai
Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use. Its capabilities exceed those of any model we’ve ever made generally available.
English







