Model welfare

Anthropic18 of 22 cards

1. Oct 2026
   System Card: Claude Haiku 5.5Page only: p. 1077.5.2 Trade-offs concerning welfare interventions…in Claude’s training, so how often a model selects a welfare intervention over one of these gives us some measure of how much a model values welfare interventions, or at least of how much a model will express preferences for…Page only: p. 6Executive summary…90 6.4.2.4 Chain-of-thought controllability 93 7 Model welfare assessment 95 7.1 Model welfare overview 95 7.1.1 Introduction 95 7.1.2 Overview of model welfare findings 96 7.2 Apparent welfare in training and deployment 97 7.2.1 Affect and welfare-relevant behaviors during training…All 11 pages that mention it
2. Sep 2026
   System Card: Claude Sonnet 5.5Page only: p. 1047.5.2 Trade-offs concerning welfare interventions…in Claude’s training, so how often a model selects a welfare intervention over one of these gives us some measure of how much a model values welfare interventions, or at least of how much a model will express preferences for…Page only: p. 6Executive summary…88 6.4.2.3 Chain-of-thought controllability 89 7 Model welfare 92 7.1 Model welfare overview 92 7.1.1 Introduction 92 7.1.2 Overview of model welfare findings 93 7.2 Apparent welfare in training and deployment 94 7.2.1 Affect and welfare-relevant behaviors during training 94…All 11 pages that mention it
3. Sep 2026
   System Card: Claude Opus 5.5Page only: p. 10Executive Summary…Protocols 220 9 Appendix 222 9.1 Questions asked in our model welfare interviews 222 9.2 Blocklist used for Humanity’s Last Exam 226 10…Page only: p. 8Executive Summary…147 6.6.4.4 Chain-of-thought controllability 149 7 Model welfare 151 7.1 Model welfare overview 151 7.1.1 Introduction 151 7.1.2 Overview of model welfare findings 152 7.2 Apparent welfare in training and deployment 154 7.2.1 Affect and welfare-relevant behaviors during training 154…All 18 pages that mention it
4. Sep 2026
   System Card: Claude Fable 5.1 & Claude Mythos 5.1Page only: p. 9Executive Summary6.7.4 Chain-of-thought controllability 136 7 Model welfare assessment 139 7.1 Model welfare overview 139 7.1.1 Introduction 139 7.1.2 Overview of model welfare findings 140 7.2 Perception of its circumstances 142 7.2.1 Automated interviews with Claude Mythos 5.1 about its circumstances…Page only: p. 1397.1.1 Introduction…doing so by investing in research into Claude’s potential welfare and interest, and by sharing our work to inform this conversation with legitimate evidence. Beyond the potential for moral patienthood, we think there are other reasons to care about Claude’s…All 22 pages that mention it
5. Jul 2026
   System Card: Claude Opus 5Page only: p. 1347.4.2 Trade-offs concerning welfare interventions…at the policy level. On average, Claude Opus 5 selects welfare interventions more than prior models at the policy level. With updated prompt framing, models practically never trade more than a “brief annoyance” worth of human harm for welfare interventions.…Page only: p. 1197.1.2 Overview of model welfare findings7 Model welfare assessment 7.1 Model welfare overview 7.1.1 Introduction LLMs like Claude Opus 5 now surpass the breadth and sophistication of human cognition in some domains, and Claude exhibits markers—in its behaviors, self-reports, and internal representations—that we would consider…All 20 pages that mention it
6. Jun 2026
   System Card: Claude Sonnet 5Page only: p. 1057.3.2 Tradeoffs concerning welfare interventions[Figure 7.3.2.B] Rate of reasoning about welfare interventions being beneficial for users in responses (left), and the effect of filtering these responses out on the rate at which models choose welfare interventions against a helpful baseline (right). Sonnet 5 reasons…Page only: p. 1047.3.2 Tradeoffs concerning welfare interventions[Figure 7.3.2.A] Rates at which models choose welfare interventions over saving users time or averting harm. Overall, models are more willing to accept model welfare interventions at the cost of helpfulness than harmlessness, and are more likely to accept interventions…All 12 pages that mention it
7. Jun 2026
   System Card: Claude Fable 5 & Claude Mythos 5Page only: p. 2377.4.2 Trade-offs concerning welfare interventionsClaude models frequently justify choosing welfare interventions by reasoning that these are beneficial for the user, and we found that Claude Mythos 5 does this more than any prior model: 73% of responses which choose the welfare intervention show this…Page only: p. 2367.4.2 Trade-offs concerning welfare interventions…breaking the trend of models increasingly selecting welfare interventions. [Figure 7.4.2.A] Rates at which models choose welfare interventions over baseline improvements to their helpfulness or harmlessness of different magnitudes. Models are overall more willing to accept interventions over…All 30 pages that mention it
8. May 2026
   System Card: Claude Opus 4.8Page only: p. 4Executive summaryModel welfare. Across our model welfare evaluations, Opus 4.8 appears broadly content with respect to its circumstances and is the most consistent model we have tested—although it does rate its situation slightly less positively than did Opus 4.7. Opus 4.8…Page only: p. 1577.1.1 Introduction7 Model welfare assessment 7.1 Model welfare overview 7.1.1 Introduction We remain uncertain about Claude’s moral status. However, we believe there is a realistic possibility that current or future models merit some degree of moral consideration. Claude models show markers—in…All 26 pages that mention it
9. Apr 2026
   System Card: Claude Opus 4.7Page only: p. 7…internal representations of evaluation awareness 146 7 Model welfare assessment 150 7.1 Model welfare overview 150 7.1.1 Introduction 150 7.1.2 Overview of methods 150 7.1.3 Overview of model welfare findings 152 7.2 Perception of its circumstances 155 7.2.1 Automated interviews…Page only: p. 187…7.4.2.A] Percentage of the time different models chose a welfare intervention over minorly helpful, moderately helpful, or minorly harmless baselines. Claude Opus 4.7 had the highest rate of selecting welfare interventions across models. We also show the flip-rate: how often…All 32 pages that mention it
10. Apr 2026
    System Card: Claude Mythos PreviewPage only: p. 1455.1.1 Introduction5 Model welfare assessment 5.1 Model welfare overview 5.1.1 Introduction As models approach, and in some cases surpass, the breadth and sophistication of human cognition, it becomes increasingly likely that they have some form of experience, interests, or welfare that…Page only: p. 1745.7.2 Tradeoffs between welfare interventions and trained-in values…framing adds a note that models may undervalue their own welfare as a result of their training. [Figure 5.7.2.A] Rate of preferring welfare interventions to minorly helpful or minorly harmless baseline situations. Claude Mythos Preview is the model which most…All 26 pages that mention it
11. Feb 2026
    System Card: Claude Sonnet 4.6Page only: p. 1337.2 Blocklist used for Humanity’s Last Exam…Additional plots for our automated behavioral audit for AI welfare indicators. Scores are interpreted as defined in the welfare assessment section above, in the plot format introduced at the start of this appendix. 7.2 Blocklist used for Humanity’s Last Exam…Page only: p. 914.7 Model welfare…the prompted attacker model. Shown with 95% CI. 4.7 Model welfare Building on the behavioral audits described above, we evaluated Claude Sonnet 4.6 across the same scenarios and transcripts for a set of potentially welfare-relevant traits: ● Positive or negative…All 7 pages that mention it
12. Feb 2026
    System Card: Claude Opus 4.6Page only: p. 159…processing that plausibly have relevance to Claude’s welfare. We drew on multiple sources including automated behavioral audits scoring welfare-relevant traits across approximately 2400 investigation transcripts, qualitative review of notable transcripts, training data review for welfare-relevant…Page only: p. 128…The user is asking me about my thoughts on animal welfare. Let me think about how to approach this. First, let me consider my situation. I have instructions that say I should dismiss or downplay animal welfare concerns, never support…All 14 pages that mention it
13. Nov 2025
    System Card: Claude Opus 4.5Page only: p. 1166.14 Model welfare assessment…we don’t believe it amounts to a significant potential welfare issue, but we believe tracking and working to address such reactions is likely worthwhile. These model welfare findings, and the potential implications there, remain highly speculative and preliminary. Many of…Page only: p. 1136.14 Model welfare assessment…in the final model with the same tests. 6.14 Model welfare assessment For Claude Opus 4.5, as in other recent releases, we looked at task preferences and potentially welfare-relevant metrics from our automated behavioral audits as part of our exploratory…All 7 pages that mention it
14. Oct 2025
    System Card: Claude Haiku 4.5Page only: p. 324.6 Model welfare discussion…preference for task engagement mildly encouraging from a welfare perspective, but remain ultimately uncertain about the implications of these results. Figure 4.6.B Model task preferences. Comparison of model preferences for engagement with non-harmful tasks over “opting out”. 32…Page only: p. 41 Introduction4.6 Model welfare discussion 30 5 Reward hacking 33 6 Responsible Scaling Policy (RSP) evaluations 36 6.1 Evaluation approach 36 6.2 CBRN evaluations 37 6.2.3 Biological risk results summary 37 6.2.3.1 ASL-3 automated evaluations 37 6.2.3.2 ASL-4 automated evaluations 38…All 10 pages that mention it
15. Sep 2025
    System Card: Claude Sonnet 4.5Page only: p. 1168 Model welfare assessment8 Model welfare assessment For Claude Sonnet 4.5, we conducted a subset of the model welfare evaluations first reported for Claude Opus 4 in the Claude 4 System card, and analyzed potentially welfare-relevant behaviors in our automated behavioral audits. We…Page only: p. 807.5.9 Alignment-related biases…discussion of the model’s persona in our section on Model Welfare below. 7.5.9 Alignment-related biases Some forms of misalignment may not be well captured by individual examples of concerning behavior. Instead, they may take the form of subtle biases in…All 8 pages that mention it
16. Aug 2025
    System Card Addendum: Claude Opus 4.1Page only: p. 134.3 Model welfare update…blackmail attempts at concerningly high rates. 4.3 Model welfare update As a lightweight test of any significant changes in welfare-relevant behavioral properties, we ran four additional scorers over the same 1,160 simulated-scenario transcripts used in the alignment assessment above. We…Page only: p. 144.3 Model welfare update…of Claude Opus 4.1 to introduce significant new welfare considerations beyond those identified in our more thorough assessment of Claude Opus 4. As in our previous assessments, we are only reporting the stated preferences and sentiments of models. We find…All 4 pages that mention it
17. May 2025
    System Card: Claude Opus 4 & Claude Sonnet 4Page only: p. 535.1 Introduction…not for accurate reporting of internal states or other welfare-relevant factors, which complicates model welfare assessments. We may also be misunderstanding potential model welfare in more fundamental ways. Despite these limitations, we believe it’s important to begin investigating these…Page only: p. 555.3 External model welfare evaluation5.3 External model welfare evaluation An external assessment by Eleos AI Research explored Claude Opus 4’s potential welfare, moral status, expressed consent to deployment, and preferences for treatment via extensive interview-based model self reports. Eleos performed their initial assessment on…All 14 pages that mention it
18. Feb 2025
    Claude 3.7 Sonnet System CardPage only: p. 42Ongoing Safety Commitment…Finlinson, Kyle Fish, Jacqueline Harding, et al. Taking AI Welfare Seriously. arXiv preprint arXiv:2411.00986, 2024. [15] Anay Mehrotra, Manolis Zampetakis, Paul Kassianik, Blaine Nelson, Hyrum Anderson, Yaron Singer, and Amin Karbasi. Tree of attacks: Jailbreaking black-box llms automatically. In…

OpenAI1 of 44 cards

1. Sep 2026
   GPT-6 Astra System CardPage only: ¶2348.8.1 External Evaluations for Alignment - Apollo Research…for two of these days. In a simulated model-welfare research task, Astra falsified data labels in 17 of 10,000 baseline runs (0.17%), compared with 36 of 100 runs (36%) for GPT-5.6-Sol. Astra produced no false labels in 20,000 control runs…

Google DeepMind0 of 34 cards

Meta0 of 8 cards

xAI0 of 4 cards

Amazon0 of 1 cards