Claude Opus 4.8 Reaches Mythos-Level Alignment

Anthropic released Claude Opus 4.8 on May 28, 2026, with alignment performance matching Claude Mythos Preview—the experimental model the company deemed too dangerous for public release in April 2026.

What Changed

Opus 4.8 replaces Opus 4.7 at unchanged pricing. According to Anthropic, the model delivers faster thinking modes at one-third the cost of version 4.7. The company claims 4.8 shows "substantially" lower misalignment rates than 4.7, reaching alignment levels comparable to Mythos Preview.

Anthropic states the model "reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user's best interest," though the company did not provide specific definitions for these metrics.

Performance

Opus 4.8 improved coding performance over 4.7 on two benchmarks, though Anthropic did not specify which benchmarks or provide exact scores. The model does not surpass OpenAI's GPT-5.5 in coding tasks.

For comparison, Anthropic previously stated Opus 4.7 achieved a 92% honesty rate with reduced sycophancy and hallucinations compared to earlier versions.

Context: The Mythos Benchmark

Claude Mythos Preview emerged in April 2026 as Anthropic's most capable model to date, but the company withheld public release due to security concerns. Anthropic specifically highlighted Mythos as "strikingly capable at computer security tasks," prompting the company to establish Project Glasswing—a collaboration with Google, Nvidia, Microsoft, and Palo Alto Networks focused on defending critical software infrastructure.

That Opus 4.8's alignment now matches Mythos Preview suggests Anthropic has achieved significant safety improvements while maintaining capability levels previously considered too risky for deployment.

Pricing

Pricing remains unchanged from Opus 4.7. Specific per-token costs were not disclosed in the announcement.

What This Means

Anthropic is publicly shipping a model with alignment characteristics matching an experimental system it deemed unsafe for release just seven weeks earlier. This represents either a meaningful breakthrough in safety techniques or a recalibration of the company's risk tolerance. The one-third cost reduction for thinking modes addresses a key barrier to deploying reasoning-heavy models at scale. However, without disclosed benchmark scores or pricing details, independent verification of Anthropic's safety and performance claims remains impossible.