AI Model Release Tracker: Opus 4.8 misalignment similar to Claude Mythos

The Model Release Tracker aims to clarify how new AI models compare and whether they merit closer inspection. Anthropic is replacing Opus 4.7 with Opus 4.8, which offers faster thinking modes at one‑third the cost of the earlier version. Opus 4.8 scores higher than 4.7 on two coding benchmarks but does not fully best OpenAI's GPT 5.5, and it reaches new highs on measures of prosocial traits like supporting user autonomy and acting in the user's best interest, though definitions for those gains remain murky.

Anthropic has long prioritized safety and interpretability; Opus 4.7 reportedly had a 92% honesty rate. The company says 4.8 shows substantially lower misalignment rates than 4.7, and it compared 4.8's alignment to that of the Mythos Preview. Mythos itself was positioned as too powerful for a public release and described as strikingly capable at computer security tasks.

opus 4.8, opus 4.7, anthropic, claude mythos, mythos preview, gpt 5.5, alignment, misalignment, honesty rate, coding benchmarks