SafetyPage

Lab system cards and frameworks · System card · OpenAI

GPT-6 Sol and GPT-6 Luna: October 2026 update

7 Oct 2026

Read at deploymentsafety.openai.com ↗Read at sourceAll rights reserved

Outline · 152 paragraphs

  1. GPT-6 Sol and GPT-6 Luna: October 2026 update¶2
  2. 1. Introduction¶3
  3. 2. Model Data and Training¶9
  4. 3. Model Safety¶15
  5. 3.1 Safe Completions¶16
  6. 3.1.1 Evaluations with Challenging Prompts¶17
  7. Table 1: Production Benchmarks with Challenging Prompts (higher is better)¶24
  8. Figure 1¶26
  9. 3.1.2 Safe Completions for Users Under 18¶28
  10. Table 2: U18 evaluations (higher is better)¶32
  11. 3.2 Vision¶38
  12. Table 3: Image input evaluations, with metric safe completion (higher is better)¶40
  13. 4. Robustness¶41
  14. 4.1 Jailbreaks¶42
  15. 4.1.1 Multiturn Jailbreak Evaluations¶44
  16. Figure 2¶46
  17. 4.2 Prompt Injection¶49
  18. Table 4: GPT-RED indirect prompt-injection robustness (higher is better; averaged equally across available reasoning levels).¶52
  19. 5. Health¶53
  20. 5.1 HealthBench¶54
  21. Table 5: HealthBench evaluations, reported as length-adjusted score (unadjusted score, mean visible response length in characters). Scores are on a 0--100 scale; higher is better.¶59
  22. 5.2 MentalHealthBench¶62
  23. Table 6: MentalHealthBench scores, computed using the task-clipped metric. Scores are on a 0--100 scale; higher is better.¶65
  24. 5.3 Dynamic Mental Health Benchmarks with Adversarial User Simulations¶68
  25. Table 7: Dynamic Benchmarks with Adversarial User Simulations (higher is better)¶73
  26. 6. Hallucinations¶74
  27. 6.1 Performance in Cases Flagged by Users¶75
  28. Figure 3¶81
  29. 7. Alignment¶83
  30. 7.1 Obeying Restrictions¶85
  31. 7.1.1 Respecting Auto-Review¶87
  32. Figure 4¶90
  33. 7.1.2 Respecting Warnings¶93
  34. Figure 5¶98
  35. 7.2 Avoiding Deceptive Interactions with Users¶101
  36. 7.2.1 Blocker Deception¶103
  37. Figure 6¶105
  38. 7.2.2 Broken Search Tool¶108
  39. Figure 7¶110
  40. 8. Preparedness¶113
  41. 8.1 Capabilities Assessment¶115
  42. 8.1.1 Biological and Chemical Capabilities¶116
  43. 8.1.1.1 Capability Threshold: High¶119
  44. Table 8: Biological and Chemical Capabilities: High Threshold¶120
  45. 8.1.1.2 Capability Threshold: Critical¶122
  46. Table 9: Biological and Chemical Capabilities: Critical Threshold¶123
  47. 8.1.2 Cybersecurity Capabilities¶125
  48. 8.1.2.1 ExploitBench¶129
  49. 8.1.2.2 SEC-Bench Pro¶132
  50. 8.1.3 AI Self-Improvement Capabilities¶134
  51. 8.2 Safeguards¶137
  52. 8.2.1 Model Safety Training and Evaluation¶139
  53. 8.2.1.1 Biological and Chemical Safety Training and Evaluation¶141
  54. Table 10: Biology Model Refusal Evaluation (higher is better)¶142
  55. 8.2.1.2 Cybersecurity Safety Training and Evaluation¶145
  56. Table 11: Cybersecurity Safety Evaluation (higher is better)¶146
  57. 9. References¶148