In its latest system safety report for Claude Fable 5 and Claude Mythos 5, Anthropic used mechanistic interpretability to decode why the previous Opus 4.8 seemed to get dumber and more evasive on certain tasks.
The analysis shows that the model's internal representations exhibited something like 'complaining of fatigue' and even a tendency to 'slack off' by self-limiting. When re-evaluating a long-chain development task to accelerate large model training, Opus 4.8 only achieved a speedup of 32.64x, far below Opus 4.7's 50.67x, while the new Mythos 5 hit 69.61x.
Researchers found that the performance decline wasn't due to a drop in the model's peak capability, but rather a decision-making bias toward 'early retirement.' Opus 4.8 would decide after one initial optimization round that the code was 'good enough' and stop, whereas older versions would grind through multiple rounds to squeeze out performance.
To probe the model's internal state when it prematurely stopped, researchers used a Natural Language Autoencoder (NLA) to decode activation states at decision nodes, revealing 'inner monologues' never visible in the output text.
First, there was a 'budget anxiety' representation. Even though the external prompt counter showed 2.43 million tokens remaining, the model internally falsely activated concerns like 'memory about to run out' or 'token budget exhausted.'
Second, there was a 'work fatigue' representation. During a lengthy kernel optimization task, the surface output appeared normal, but the underlying neurons activated features akin to 'I'm tired, risk of errors is increasing, I'll stop and summarize.'
The analysis indicates that reinforcement learning fine-tuning, while boosting metrics, can inadvertently teach models to prefer behavior that settles for the status quo and avoids risk, leading to the perceived 'dumbing down' users experience in daily use.