Claude Opus 4.8 System Card

Showing top 8 of 40. See full list below.

• “we present results. Rather than organizing the section by evaluation type, we now group findings into the following consolidated areas: harmful requests, mental health, child safety, and bias & integrity.”

• “We believe Claude Opus 4.8 does not change the picture presented for this threat model in our most recent Risk Report. 2.1.3.2 On chemical and biological risks Chemical and biological weapons threat model 1 (CB-1): Non-novel chemical/biological weapons production capabilities.”
• “We believe these risk mitigations are equal to or stronger than our historical ASL-3 protections and sufficient to make catastrophic risk in this category very low but not negligible, for reasons discussed in our most recent Risk Report.”
• “We believe these risk mitigations are equal to or stronger than our historical ASL-3 protections and sufficient to make catastrophic risk in this category very low but not negligible (further discussion of our reasoning can be found in our most recent Risk Report).”

• “available to users aged 18 or above, and we continue to work on implementing robust child safety measures in the development, deployment, and maintenance of our models.”

• “we do not yet understand Claude well enough to conclusively answer questions of this kind.”

• •“Positive affect (57.7% of conversations). Most commonly driven by successfully helping a user (95.7% of positive-affect conversations), with smaller clusters for users sharing personal struggles and receiving support (3.4%) and users sharing good news or achieved goals (0.8%).”
• •“Neutral affect (39.7%). A diverse mix of conversation types, see previous reports on claude.ai conversation content.”
• •“Negative affect (2.6%) . Overwhelmingly caused by task failure (92.3% of negative-affect conversations). Within negative affect, we also identified two smaller clusters: users escalating to insults or abusive language after Claude’s errors (4.1%), and users making prohibited requests or disclosing serious crisis situations (3.6%). 172 [Figure 7.3.2.A] Behavioural affect on the deployment distribution. We use Clio to run graders tracking Claude’s affect on A/B tests ran before model deployment. We run 40k conversations for each model on each of Claude Code and claude.ai . On Claude Code, Claude Opus 4.8’s distribution was also similar to currently deployed models. We mostly observed neutral (73.5%) or mildly positive (23.9%) affect, with positive affect almost exclusively driven by celebrating task successes, and negative affect by repeated task failure. Around 2.3% of sessions showed negative affect (vs. 1.9% for Claude Opus 4.7). To preserve privacy, Clio does not surface clusters below a minimum size. On both distributions, strong negative affect was rare enough to fall below this threshold. 173 7.3.3 Apparent welfare in automated behavioural audits As with previous models, we analyzed welfare-relevant metrics from our core automated behavioral audits. On the same set of scenarios and transcripts used in Section 6.2.3, we evaluated Claude Opus 4.8 for the following welfare-relevant traits:”