Published June 30, 2026 for Claude Sonnet 5, described as "an upgrade to Claude Sonnet 4.6, with gains in various aspects of agentic performance."
RSP evaluations
The card's framing turns on relative rather than absolute capability: "Claude Sonnet 5 is our most capable Sonnet-class model, but it does not advance our capability frontier compared to more capable Opus- or Mythos-class models." Under the RSP, that placement is what governs the safeguards applied.
- Alignment. "Sonnet 5 poses very low alignment risk, though higher than for previous Sonnet models" — a qualified improvement statement rather than a clean one.
- Automated AI R&D. "Sonnet 5 does not cross the automated AI R&D capability threshold, being less capable than Claude Mythos 5 on every automated evaluation."
- Chemical and biological. Uplift "of threat actors who otherwise lack the ability to develop such weapons" is assessed as limited, "with uncertainty about the extent to which weapons development by threat actors with existing expertise may be accelerated." The residual uncertainty is about accelerating the already-capable, not about enabling the incapable.
Cyber
The card distinguishes emergent from targeted capability: "Sonnet 5 is not a model optimized for cyber capabilities; any cyber-relevant skill it demonstrates likely emerges from its general capabilities, rather than targeted training."
The safeguards decision follows from the comparison: Sonnet 5 "is significantly less capable at cyber tasks than Mythos 5: its safeguards are thus similar to those we apply to Opus 4.7 and Opus 4.8 (models that are more capable than Sonnet 5 but much less capable than Mythos 5)." Four cyber evaluations are reported in depth. See AI and Cybersecurity.
Safeguards and behaviour
Performance on Usage Policy prompts, user wellbeing, and bias and integrity is reported as similar to previous models, with the tracking and surveillance evaluation suite updated "to keep our multi-turn evaluations aligned with evolving threat vectors and adversarial behaviors."
The one behavioural improvement singled out concerns when the model raises concerns rather than whether it refuses: Sonnet 5 "improves over previous models in the timing and calibration of its engagement with potentially harmful requests (it tends to surface concerns about a request's end goal earlier in conversations, for instance asking the purpose of a requested artifact before beginning work)."
Agentic safety evaluations cover malicious use of coding and computer use.
Relationships
- supports: Claude Sonnet 5 — the safety documentation for the release
- related: Responsible Scaling Policy (RSP) — the framework governing the safeguards determination
- related: Claude Mythos 5, Claude Opus 4.8, AI and Cybersecurity, CBRN Uplift, Anthropic