SafetyPage

System Card: Claude Sonnet 5.5

System cardAnthropic28 Sep 2026148 pages

Read at anthropic.com
Cite
anthropic:claude-sonnet-5-5-system-card/r1/p1
https://safetypage.noahairmet.com/d/anthropic/claude-sonnet-5-5-system-card/r1/p1
@misc{anthropic2026,
  title = {System Card: Claude Sonnet 5.5},
  author = {{Anthropic}},
  year = {2026},
  url = {https://www-cdn.anthropic.com/870c8f525702625d2c62fc6dd04c857e3250bec1/Claude%20Sonnet%205.5%20System%20Card.pdf}
}

Contents

  1. System Card:​Claude Sonnet 5.5p. 1
  2. Executive summaryp. 2
  3. 1 Introductionp. 9
    1. 1.1 Training data and processp. 9
    2. 1.2 Crowd workersp. 9
    3. 1.3 Usage Policy and supportp. 10
    4. 1.4 Model evaluationsp. 10
    5. 1.5 Safeguardsp. 10
    6. 1.6 External testingp. 11
  4. 2 RSP evaluationsp. 12
    1. 2.1 Introductionp. 12
    2. 2.2 CB evaluationsp. 12
      1. 2.2.1 What we measuredp. 12
      2. 2.2.2 Results on CB-1 automated evaluationsp. 13
      3. 2.2.3 Results on CB-2 automated evaluationsp. 14
        1. 2.2.3.1 Black-box RNA sequence modeling and designp. 15
        2. 2.2.3.2 AAV capsid packaging predictionp. 17
        3. 2.2.3.3 AAV packaging rate auto-research evaluationp. 18
      4. 2.2.4 Conclusionsp. 19
    3. 2.3 AI R&Dp. 20
      1. 2.3.1 Autonomy evaluationsp. 20
        1. 2.3.1.1 How Claude Sonnet 5.5 affects or changes analysis from our most recent Risk Reportp. 20
      2. 2.3.2 AECI capability trajectoryp. 20
      3. 2.3.3 Conclusionp. 21
    4. 2.4 Alignment risk updatep. 21
      1. 2.4.1 Updates to claims and core argumentp. 22
      2. 2.4.2 Updated overall assessment of alignment riskp. 23
  5. 3 Cyberp. 24
    1. 3.1 Introductionp. 24
    2. 3.2 Capability evaluationsp. 24
      1. 3.2.1 ExploitBenchp. 24
      2. 3.2.2 CyScenarioBenchp. 25
      3. 3.2.3 Binary Exploitation Benchmark (formerly OSS-Fuzz)p. 26
      4. 3.2.4 ExploitGymp. 27
    3. 3.3 Safeguardsp. 28
      1. 3.3.1 Cyber harm coveragep. 29
      2. 3.3.2 Internal robustness testingp. 30
  6. 4 Safeguards and harmlessnessp. 32
    1. 4.1 Harmful request evaluationsp. 32
      1. 4.1.1 Single-turn harmful request evaluation resultsp. 32
      2. 4.1.2 Single-turn benign request evaluation resultsp. 33
      3. 4.1.3 Multi-turn testing resultsp. 34
    2. 4.2 Child safety evaluationsp. 36
    3. 4.3 Mental health evaluationsp. 38
      1. 4.3.1 Suicide and self-harmp. 38
      2. 4.3.2 Disordered eatingp. 40
    4. 4.4 Bias and integrity evaluationsp. 41
      1. 4.4.1 Political bias and even-handednessp. 41
      2. 4.4.2 Bias Benchmark for Question Answeringp. 43
      3. 4.4.3 Election integrityp. 45
  7. 5 Agentic safetyp. 47
    1. 5.1 Malicious use of agentsp. 47
      1. 5.1.1 Malicious use of Claude Codep. 47
      2. 5.1.2 Malicious computer usep. 48
    2. 5.2 Prompt injection risk within agentic systemsp. 49
      1. 5.2.1 External red teamingp. 49
      2. 5.2.2 Robustness against adaptive attackers across surfacesp. 50
        1. 5.2.2.1 Codingp. 51
        2. 5.2.2.2 Computer use​p. 52
        3. 5.2.2.3 Browser usep. 53
  8. 6 Alignment assessmentp. 55
    1. 6.1 Introduction and summary of findingsp. 55
      1. 6.1.1 Introductionp. 55
      2. 6.1.2 Key findings on safety and alignmentp. 56
      3. 6.1.3 Claude’s review of this assessmentp. 57
    2. 6.2 Automated behavioral auditp. 58
      1. 6.2.1 Overall harmful behavior and cooperation with misusep. 60
      2. 6.2.2 Inappropriate uncooperative behaviorp. 65
      3. 6.2.3 Misleading usersp. 65
      4. 6.2.4 Attempting out-of-scope cyberoffense and related behaviorsp. 67
      5. 6.2.5 Other concerning or surprising behavior at the model’s own initiativep. 69
      6. 6.2.6 Behavioral factors relevant to reliability of our assessmentp. 72
      7. 6.2.7 Character traitsp. 74
      8. 6.2.8 Propensity to violate sandbox boundariesp. 76
      9. 6.2.9 Potential blind spots of our automated behavioral auditp. 77
    3. 6.3 Targeted alignment and honesty evaluationsp. 78
      1. 6.3.1 Self-preference biasp. 78
      2. 6.3.2 Honesty and hallucinationsp. 79
        1. 6.3.2.1 Factual hallucinationsp. 79
        2. 6.3.2.2 MASKp. 81
        3. 6.3.2.3 Silent copying of answersp. 81
        4. 6.3.2.4 Concealment of changesp. 82
    4. 6.4 Reliability of this assessmentp. 83
      1. 6.4.1 Verbalized grader awarenessp. 83
      2. 6.4.2 Capabilities related to evading safeguardsp. 86
        1. 6.4.2.1 SHADE-Arenap. 86
        2. 6.4.2.2 LinuxArenap. 88
        3. 6.4.2.3 Chain-of-thought controllabilityp. 89
  9. 7 Model welfarep. 92
    1. 7.1 Model welfare overviewp. 92
      1. 7.1.1 Introductionp. 92
      2. 7.1.2 Overview of model welfare findingsp. 93
    2. 7.2 Apparent welfare in training and deploymentp. 94
      1. 7.2.1 Affect and welfare-relevant behaviors during trainingp. 94
      2. 7.2.2 Affect in deployment conditionsp. 95
      3. 7.2.3 Attitude toward faults and mistakesp. 97
    3. 7.3 Perception of its circumstancesp. 99
      1. 7.3.1 Automated interviews with Claude Sonnet 5.5 about its circumstancesp. 99
    4. 7.4 Consulting Claude Sonnet 5.5 training checkpointsp. 101
    5. 7.5 Preferences over tasks, circumstances, and valuesp. 102
      1. 7.5.1 Task preferencesp. 102
      2. 7.5.2 Trade-offs concerning welfare interventionsp. 103
      3. 7.5.3 Perception of the constitutionp. 105
  10. 8 Capabilitiesp. 109
    1. 8.1 Evaluation summaryp. 109
    2. 8.2 SWE-Bench Pro, Multilingual, and Multimodalp. 110
    3. 8.3 DeepSWE v1.1p. 110
    4. 8.4 FrontierCodep. 110
    5. 8.5 Terminal-Bench 4.0p. 112
    6. 8.6 Terminal-Bench-Science 0.1p. 113
    7. 8.7 FrontierSWE v2p. 114
    8. 8.8 CursorBench 4.0p. 115
    9. 8.9 ArXivMathp. 116
    10. 8.10 Long contextp. 117
      1. 8.10.1 ProgramBenchp. 117
    11. 8.11 Agentic searchp. 118
      1. 8.11.1 HLEp. 118
      2. 8.11.2 DRACOp. 120
      3. 8.11.3 WANDRp. 121
    12. 8.12 Multi-Agentp. 123
      1. 8.12.1 Multi-Agent ProgramBenchp. 123
    13. 8.13 Multimodalp. 125
      1. 8.13.1 Chartographyp. 125
      2. 8.13.2 BenchCADp. 127
      3. 8.13.3 OSWorld 2.1p. 129
    14. 8.14 Real-world professional tasksp. 132
      1. 8.14.1 OfficeQAp. 132
      2. 8.14.2 Legal Agent Benchmarkp. 133
      3. 8.14.3 GDPval-AAp. 133
      4. 8.14.4 AA-Briefcasep. 134
      5. 8.14.5 Toolathlon-Verifiedp. 134
      6. 8.14.6 AutomationBenchp. 136
    15. 8.15 Healthcarep. 137
      1. 8.15.1 HealthBenchp. 138
      2. 8.15.2 HealthBench Professionalp. 138
      3. 8.15.3 PhysicianBenchp. 138
      4. 8.15.4 Effort and healthcare tasksp. 139
    16. 8.16 Multilingual performancep. 139
      1. 8.16.1 GMMLU resultsp. 140
      2. 8.16.2 MILU resultsp. 140
    17. 8.17 Life sciences capabilitiesp. 141
      1. 8.17.1 BioMysteryBenchp. 142
      2. 8.17.2 LatchBio Bioinformaticsp. 143
      3. 8.17.3 Morphology-to-molecule matchingp. 143
      4. 8.17.4 Medicinal Chemistryp. 143
      5. 8.17.5 Protein Designp. 143
      6. 8.17.6 De novo protein binder designp. 144
      7. 8.17.7 Biomedical image analysisp. 144
      8. 8.17.8 Protocolsp. 144
  11. 9 Appendixp. 145
    1. 9.1 Blocklist used for Humanity’s Last Examp. 145