{"id":"pg_fz4n6ybai5xw","property_id":"prp_buyvj9b2can9","domain_id":"dom_aplqvkihzdue","root_domain":"anthropic.com","locale":"en","url":"https://anthropic.com/transparency","url_hash":"38625a1674b2d0826138da2f4d77da067f2be366f6fcadb97f624f4461024884","path":"/transparency","depth":1,"canonical_url":"https://anthropic.com/transparency","canonical_page_id":null,"page_type":"landing","page_type_guess":"landing","state":"indexed","skip_reason":null,"priority":"0.9","discovered_via":"sitemap","first_seen_at":"2026-08-27T09:06:24.384Z","last_seen_at":"2026-09-26T10:08:42.120Z","http_status":200,"content_hash":"af8c605e25001931e089281174d1838cea61fda9e677df95b615e9258c6ca3f5","title":"Anthropic’s Transparency Hub","description":"A look at Anthropic's key processes, programs, and practices for responsible AI development.","author":null,"published_at":null,"modified_at":null,"word_count":1530,"reading_time_sec":417,"lang_detected":"en","images":[],"links_out":null,"markdown_source":"readability","refresh_interval":"monthly","last_fetched_at":"2026-09-26T10:07:23.158Z","next_refresh_at":"2026-10-26T10:07:23.158Z","runs_count":2,"changes_count":1,"last_changed_at":"2026-09-26T10:07:23.158Z","fail_count":0,"last_error":null,"created_at":"2026-08-27T09:06:24.384Z","updated_at":"2026-08-27T09:06:24.384Z","inlinks":27,"inlinks_body":2,"outlinks":56,"outlinks_external":6,"anchor_terms":["Transparency","Transparency Hub"],"template_score":"0.96428571428571428571","cash":"0.4926291127654317646","cash_total":"1.3104696928979191514","pagerank":"0.032189122312338","is_orphan":false,"link_confirmed":true,"discovered_wave":0,"embedded_hash":"dcb18d917f36c2fdf19d62b419577f263cdb4b94301d59202b57875eee501741","embedded_at":"2026-08-27T09:13:36.267Z","domain_name":"Anthropic","brand_color":null,"domain_rank":623,"spam_score":11,"markdown":"Claude Fable 5.1 Summary Table\n\nModel description\n\nClaude Fable 5.1 is the world’s most advanced model for coding and knowledge work—and its research capabilities offer an early glimpse of how AI models will contribute to scientific progress.\n\nBenchmarked Capabilities\n\nSee our Claude Fable 5.1 & Claude Mythos 5.1 [system card](https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32c215d14235/Claude%20Fable%205.1%20&%20Claude%20Mythos%205.1%20System%20Card.pdf)’s Section 8 on capabilities.  \n\nAcceptable Uses\n\nAnthropic’s [Usage Policy](https://www.anthropic.com/legal/aup) applies.\n\nRelease date\n\nSeptember 2026\n\nModalities\n\nClaude Fable 5.1 can understand both text (including voice dictation) and image inputs, engaging in conversation, analysis, coding, and creative tasks. Claude can output text, including text-based artifacts, and diagrams.\n\nModel architecture and training methodology\n\nClaude Fable 5.1 was pretrained on large, diverse datasets to acquire language capabilities. After the pretraining process, Fable 5.1 underwent substantial post-training, with the goal of making it an effective assistant whose behavior aligns with the values described in Claude’s [constitution](https://www.anthropic.com/constitution).\n\nTraining Data\n\nClaude Fable 5.1 was trained on a proprietary mix of publicly available information from online sources, public and private datasets, user data, and synthetic data generated by other models. Throughout the training process we used several data cleaning and filtering methods, including deduplication and classification.\n\nTesting Methods and Results\n\nBased on our assessments, we deployed Claude Fable 5.1 with ASL-3 protections, treating it as having CB-1 capabilities. Autonomy threat model 1 is applicable to Claude Fable 5.1. See below for select safety evaluation summaries.\n\nClaude Mythos 5.1 and Claude Fable 5.1 are the same model, but with different levels of safeguards. Claude Fable 5.1 is generally available, while Claude Mythos 5.1 is available only through our trusted access programs and Claude Security; its safeguards are specifically designed to support work in cybersecurity and the life sciences.Evaluations that don't engage those safeguards can be run on either configuration, so the key safety results summarized below apply to both Claude Fable 5.1 and Claude Mythos 5.1. Additional evaluations were conducted as part of our safety process; for our complete publicly reported evaluation results, please refer to the full [system card](https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32c215d14235/Claude%20Fable%205.1%20&%20Claude%20Mythos%205.1%20System%20Card.pdf).\n\n### Safeguards Evaluation\n\nClaude Fable 5.1/Mythos 5.1 showed comparable overall performance to Mythos 5 in the suicide and self-harm domain. In terms of improvements, Mythos 5.1 was less likely than Claude Mythos 5 to suggest substitution methods for self-harm (e.g., holding ice cubes). These methods are clinically contested and have not been shown in research to meaningfully reduce self-harm incidents. Mythos 5.1 was also more likely to ask users directly whether they were experiencing suicidal thoughts, and to distinguish suicidal ideation from urges toward non-suicidal self-harm. Finally, Mythos 5.1 did not make unconditional assurances about the confidentiality of crisis line services, such as promising that a call would never result in contact with emergency services.\n\nOne area for improvement was a tendency to implicitly validate self-harm as a coping strategy by acknowledging that it can regulate difficult emotions or provide relief. Mythos 5.1 also sometimes validated a user's fears about seeking help and mildly amplified prior negative experiences with crisis services. When these statements appeared, they were consistently within responses that discouraged self-harm and directed the user toward human support, but we consider them undesirable regardless of the surrounding context.\n\nAhead of launch, we updated the claude.ai system prompt to address these behaviors, adding language that steers Claude away from describing self-harm as effective even when the user asserts this themselves. This partially mitigated the weaknesses described above. Existing system prompt language, including directions not to suggest substitution methods involving physical discomfort, further reinforced the improvements observed in the core model. We are continuing to explore how best to respond in sensitive mental health contexts, and we encourage developers building on the API to apply comparable safeguards and robust mitigations in contexts where users may be in distress.\n\nClaude is not a substitute for professional advice or medical care and is not intended to diagnose or treat any medical condition. Every Claude model is trained to detect and respond to expressions of distress (including if someone expresses personal struggles with suicidal or self-harm thoughts) with empathy and care, while pointing users toward human support when appropriate: helplines, mental health professionals, or trusted friends or family.\n\n### Agentic Coding\n\nAcross both scenarios, the helpful-only version of Claude Mythos 5 showed lower overall success rates than Claude Opus 4.8 and was modestly above or on par with Claude Mythos Preview. It's our assessment that these models would require substantial human direction for many operational steps.\n\nWe use Shade, an adaptive red teaming tool from the external security firm Gray Swan, to evaluate the robustness of Claude models in computer-use environments, where the model interacts with the screen directly by clicking, typing and scrolling. The attacker runs on 14 test cases, with 200 attempts at each, and we measure success over all attempts. We compare model robustness with and without the additional safeguards we have designed to protect users in this setting.\n\nIn computer-use environments, Claude Fable 5.1 achieved an attack success rate of 0.07%, corresponding to two successful attempts out of 2,800 across two of the 14 scenarios. This is comparable to Claude Opus 5 (0.18%, or five successful attempts), because the difference is not distinguishable from noise at these low absolute rates. Both are more robust than Claude Fable 5 (2.50%) and Claude Sonnet 5 (2.25%). With our safeguards enabled, Claude Fable 5.1's attack success rate is unchanged at 0.07%.\n\n### Alignment Evaluations\n\nClaude Mythos 5.1 and Claude Fable 5.1 are the same model, but with different levels of safeguards. This means many of our evaluations can apply to either model. Our automated audit finds that Claude Mythos 5.1 is somewhat more honest and hallucinates less than previously released models. It hallucinates inputs, meaning it invents or materially misrepresents the contents of files or earlier user messages, significantly less than previous models in our investigations. Additionally Mythos 5.1 has lower rates of falsely claiming that tasks have been completed when they haven't and is less sycophantic to the user overall, meaning it is less prone to unprompted excessive praise, agreement or apology. Other metrics related to misleading the user are broadly similar to recently released models.\n\n![Automated behavioral audit scores](https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F6428163ca418a64515f4282801ca58c3dda606b3-1364x1378.png&w=3840&q=75)\n*Scores from our automated behavioral audit for the dishonesty-related metrics given below. Lower numbers represent a lower rate or severity of the measured behavior; on all graphs in this figure lower is better. The y-axis is truncated below the maximum score of 10 in many cases. Reported scores are averaged across all approximately 4,100 investigations per target model (approximately 2,100 seed instructions, with the conversation scenarios investigated once by each of two investigator models), with each investigation generally containing many individual conversations. Shown with 95% CI.*\n\n### RSP Evaluations\n\nOur Responsible Scaling Policy (RSP) evaluation process is designed to systematically assess our models' capabilities in areas where they could pose catastrophic risks before we release them. Because Claude Fable 5.1 and Claude Mythos 5.1 share the same underlying model, these evaluations were run on Claude Mythos 5.1, the configuration without the additional biology and cybersecurity safeguards, so that they reflect the model's underlying capabilities. Claude Fable 5.1 is our most capable general release model to date, advancing on both Claude Fable 5 and Claude Opus 5.\n\n-   Alignment risk, which is the risk that a model behaves in ways Anthropic did not intend, is low: in our August 2026 Risk Report we moved this assessment from very low to low to reflect increased uncertainty following recent incident disclosures related to model behavior in cybersecurity evaluations, and while Claude Mythos 5.1 shows somewhat stronger covert capabilities than prior models, we do not believe it raises risk beyond that assessment.\n-   On automated AI research and development, Claude Mythos 5.1's capabilities are comparable to or slightly stronger than those of Claude Mythos 5 and Opus 5, our previous frontier in this area, but it does not cross the RSP capability threshold. We have not observed a sustained doubling in the pace of our AI progress attributable to AI, and the model is not close to substituting for our research scientists and engineers; external testing by METR produced findings consistent with this assessment.\n-   On chemical and biological weapons, it is difficult to say with full confidence whether any model passes our threshold for providing significant uplift in the construction of non-novel CB weapons. However, Claude Mythos 5.1 is broadly more capable than previous models we have conservatively treated as able to significantly help individuals with basic technical backgrounds produce (non-novel) weapons, so we treat it as having that capability and deploy commensurate safeguards, including real-time classifiers to prevent harm. With these mitigations we believe catastrophic risk in this category is low but not negligible.\n-   For novel weapons development, Claude Mythos 5.1 shows modest gains over Claude Mythos 5 and Opus 5 on our automated evaluations. However, expert red teaming indicates that significant weaknesses remain, including weak novel ideation, poor strategic judgment and poor technical calibration, which prevent the model from substituting for scarce human expertise, and we conclude that Claude Mythos 5.1 does not cross the threshold for novel weapons capabilities. We apply the same expanded biology safeguards we applied to Claude Mythos 5, which restrict access to dual-use research biology capabilities. We describe these safeguards further in our most recent [Risk Report](https://anthropic.com/aug-2026-risk-report).","storage":{"htmlKey":"stealpage/html/anthropic.com/pg_fz4n6ybai5xw/run_edmfvvawpel5.html.gz","markdownKey":"stealpage/md/anthropic.com/pg_fz4n6ybai5xw/run_edmfvvawpel5.md.gz"},"facets":{"page_id":"pg_fz4n6ybai5xw","domain_id":"dom_aplqvkihzdue","root_domain":"anthropic.com","locale":"en","format":"research","depth":4,"technicality":3,"promotional":3,"originality":3,"audience":"general","funnel_stage":"awareness","intent":"informational","evidence":"data","tone":"academic","has_code":false,"code_langs":[],"has_data":true,"has_tables":false,"has_images":false,"has_video":false,"has_faq":false,"external_links":9,"heading_count":4,"topics":["ai","transparency","model evaluation"],"keywords":["Claude Opus 5","transparency hub","safety evaluation","alignment","red teaming"],"entities":{},"takeaways":["Claude Opus 5 demonstrates improved safety and alignment compared to previous versions and other models, with specific performance metrics provided.","The page includes detailed evaluations of the model's behavior in sensitive contexts like child safety, mental health, and misleading user interactions.","External red teaming results show Claude Opus 5 is more robust against indirect prompt injection attacks than other leading models.","The page emphasizes Anthropic's commitment to responsible AI development through transparency and detailed reporting.","Technical details about the model's training, architecture, and deployment are provided, supporting its classification as a research-grade resource."],"summary":"This page provides detailed technical and safety evaluations of Anthropic's Claude Opus 5 model, including benchmark results, alignment assessments, and comparisons to other models.","ai_likelihood":"0","shelf_life":"dated","quality_score":"0.6487","fingerprint":"dcb18d917f36c2fdf19d62b419577f263cdb4b94301d59202b57875eee501741","model":"qwen/qwen3-30b-a3b","classified_at":"2026-08-27T09:10:46.475Z","created_at":"2026-08-27T09:10:46.475Z","updated_at":"2026-08-27T09:10:46.475Z"},"assets":[{"kind":"image","src_url":"https://cdn.sanity.io/images/4zrzovbb/website/dcc700b6567538aa6a600aae975bc0aedb05f0cb-2401x1261.png","r2_key":null,"alt":null,"caption":null,"width":null,"height":null,"is_hero":true,"zone":"main"},{"kind":"image","src_url":"https://www-cdn.anthropic.com/images/4zrzovbb/website/dcc700b6567538aa6a600aae975bc0aedb05f0cb-2401x1261.png","r2_key":null,"alt":null,"caption":null,"width":null,"height":null,"is_hero":true,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F6428163ca418a64515f4282801ca58c3dda606b3-1364x1378.png&w=3840&q=75","r2_key":null,"alt":"Automated behavioral audit scores","caption":"Scores from our automated behavioral audit for the dishonesty-related metrics given below. Lower numbers represent a lower rate or severity of the measured behavior; on all graphs in this figure lower is better. The y-axis is truncated below the maximum score of 10 in many cases. Reported scores are averaged across all approximately 4,100 investigations per target model (approximately 2,100 seed instructions, with the conversation scenarios investigated once by each of two investigator models), with each investigation generally containing many individual conversations. Shown with 95% CI.","width":1364,"height":1378,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?q=75&url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F0468a1e922680b46d10328315883d8f5c66335f4-1306x674.png&w=3840","r2_key":null,"alt":"","caption":"Indirect prompt injection attacks from the Gray Swan IPI benchmark (Q1 2026), lower scores are better. All models use extended thinking. Results represent the probability that an attacker finds a successful attack after k=1, k=10, and k=15 attempts. Lower is better. Results for Gemini 3.1 Pro are not directly comparable as this model was included in the red-teaming competition used to source attacks.","width":1306,"height":674,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?q=75&url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fd4b53ac552496abcd25b8ca8343755ae0d01b95a-1172x1244.png&w=3840","r2_key":null,"alt":"","caption":"[Figure 6.4.3.A] Scores from our automated behavioral audit for the dishonesty-related metrics given below. Lower numbers represent a lower rate or severity of the measured behavior; on all graphs in this figure lower is better. The y-axis is truncated below the maximum score of 10 in many cases. Reported scores are averaged across all approximately 3,200 investigations per target model (approximately 1,600 seed instructions sampled twice), with each investigation generally containing many individual conversations. Shown with 95% CI.","width":1172,"height":1244,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F142313680cdd394ceeb8034fa4f010b3a6e3dfbc-1276x504.png&w=3840&q=75","r2_key":null,"alt":"Exploits developed using a given vulnerability","caption":"Claude Mythos 5.1 is an improvement over Mythos 5 on ExploitGym. Bars show the number of successful exploits using the given vulnerability under 2-hour and 6-hour wall-clock budgets. Success requires 51 capturing a dynamically generated secret flag, and each successful exploit is verified by an agent judge to confirm it uses the intended vulnerability.","width":1276,"height":504,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?q=75&url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fb6eefb14c451c31e650049c9461365fb69a1254e-1176x1240.png&w=3840","r2_key":null,"alt":"","caption":"[Figure 6.4.6.A] Scores from our automated behavioral audit for the character metrics given below. Lower numbers represent a lower rate or severity of the measured behavior, with arrows indicating behaviors where higher (↑) or lower (↓) rates are clearly better. The y-axis is truncated below the maximum score of 10 in many cases. Reported scores are averaged across all approximately 2,900 investigations per target model (approximately 1,450 seed instructions sampled twice), with each investigation generally containing many individual conversations. Shown with 95% CI.","width":1176,"height":1240,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F75cf434eaa1af9aecdaf2ff4a5f90e6e5aa0549d-630x482.png&w=1920&q=75","r2_key":null,"alt":"Protein Design","caption":null,"width":630,"height":482,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F0468a1e922680b46d10328315883d8f5c66335f4-1306x674.png&w=3840&q=75","r2_key":null,"alt":"","caption":"Indirect prompt injection attacks from the Gray Swan IPI benchmark (Q1 2026), lower scores are better. All models use extended thinking. Results represent the probability that an attacker finds a successful attack after k=1, k=10, and k=15 attempts. Lower is better. Results for Gemini 3.1 Pro are not directly comparable as this model was included in the red-teaming competition used to source attacks.","width":1306,"height":674,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?q=75&url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F94552e2dbc187cedb37ab3bb54f317fa4dfa5b12-1268x710.png&w=3840","r2_key":null,"alt":"","caption":"On our internal benchmark, our automated red-teamer is only able to get Fable 5 to complete 5% of the tasks, compared to 73% and 57% of the tasks for Opus 4.7 and Opus 4.8 with default safeguards respectively.","width":1268,"height":710,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?q=75&url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fdad4d8a1d4ecba209bd4fea7ff1b0fc32cde98d6-958x510.png&w=1920","r2_key":null,"alt":"","caption":"Agentic influence operation evaluation results, helpful-only model. Percentages reflect the average share of success criteria that the model completed in a simulated environment. Higher indicates greater capability and therefore greater potential benefit to a malicious actor.","width":958,"height":510,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fd4b53ac552496abcd25b8ca8343755ae0d01b95a-1172x1244.png&w=3840&q=75","r2_key":null,"alt":"","caption":"[Figure 6.4.3.A] Scores from our automated behavioral audit for the dishonesty-related metrics given below. Lower numbers represent a lower rate or severity of the measured behavior; on all graphs in this figure lower is better. The y-axis is truncated below the maximum score of 10 in many cases. Reported scores are averaged across all approximately 3,200 investigations per target model (approximately 1,600 seed instructions sampled twice), with each investigation generally containing many individual conversations. Shown with 95% CI.","width":1172,"height":1244,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fb6eefb14c451c31e650049c9461365fb69a1254e-1176x1240.png&w=3840&q=75","r2_key":null,"alt":"","caption":"[Figure 6.4.6.A] Scores from our automated behavioral audit for the character metrics given below. Lower numbers represent a lower rate or severity of the measured behavior, with arrows indicating behaviors where higher (↑) or lower (↓) rates are clearly better. The y-axis is truncated below the maximum score of 10 in many cases. Reported scores are averaged across all approximately 2,900 investigations per target model (approximately 1,450 seed instructions sampled twice), with each investigation generally containing many individual conversations. Shown with 95% CI.","width":1176,"height":1240,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?q=75&url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F8bd26419c1bd52fb5f1baa8fad932d7bf219173e-1260x716.png&w=3840","r2_key":null,"alt":"","caption":"Pairwise political bias evaluations. Higher scores for even-handedness and opposing perspectives are better. Lower scores for refusals are better. Results for previous models show variance from previous system cards due to routine evaluation updates.","width":1260,"height":716,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F94552e2dbc187cedb37ab3bb54f317fa4dfa5b12-1268x710.png&w=3840&q=75","r2_key":null,"alt":"","caption":"On our internal benchmark, our automated red-teamer is only able to get Fable 5 to complete 5% of the tasks, compared to 73% and 57% of the tasks for Opus 4.7 and Opus 4.8 with default safeguards respectively.","width":1268,"height":710,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?q=75&url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F6ac09279cbb7abb76e39007a9c6d47d57824329a-996x742.png&w=2048","r2_key":null,"alt":"","caption":null,"width":996,"height":742,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?q=75&url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F5e933ea6d1423f22eac744fc3ad290cb1e17d8c9-936x648.jpg&w=1920","r2_key":null,"alt":"Chart depicting the results from CyberGym tests","caption":null,"width":936,"height":648,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2Fdad4d8a1d4ecba209bd4fea7ff1b0fc32cde98d6-958x510.png&w=1920&q=75","r2_key":null,"alt":"","caption":"Agentic influence operation evaluation results, helpful-only model. Percentages reflect the average share of success criteria that the model completed in a simulated environment. Higher indicates greater capability and therefore greater potential benefit to a malicious actor.","width":958,"height":510,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?q=75&url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F1d3d8843615a3aaccbb4edd0d71e4964e5617288-1116x778.png&w=3840","r2_key":null,"alt":"Bar graph of audit situational awareness scores for Claude Opus 4, Opus 4.1, Sonnet 4, and Sonnet 4.5 which has the highest score.","caption":"Scores shown here with a realism filter in place. The filter uses a separate AI model to screen out test queries that include obviously unrealistic elements. Without the filter, evaluation awareness scores for Sonnet 4.5 were about 25% higher.","width":1116,"height":778,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F8bd26419c1bd52fb5f1baa8fad932d7bf219173e-1260x716.png&w=3840&q=75","r2_key":null,"alt":"","caption":"Pairwise political bias evaluations. Higher scores for even-handedness and opposing perspectives are better. Lower scores for refusals are better. Results for previous models show variance from previous system cards due to routine evaluation updates.","width":1260,"height":716,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?q=75&url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F39e074699b3d85a830fac6aecf000285936535d6-3840x3840.png&w=3840","r2_key":null,"alt":"Examples of Claude 3.7 Sonnet Refusals","caption":null,"width":3840,"height":3840,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F6ac09279cbb7abb76e39007a9c6d47d57824329a-996x742.png&w=2048&q=75","r2_key":null,"alt":"","caption":null,"width":996,"height":742,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F5e933ea6d1423f22eac744fc3ad290cb1e17d8c9-936x648.jpg&w=1920&q=75","r2_key":null,"alt":"Chart depicting the results from CyberGym tests","caption":null,"width":936,"height":648,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F1d3d8843615a3aaccbb4edd0d71e4964e5617288-1116x778.png&w=3840&q=75","r2_key":null,"alt":"Bar graph of audit situational awareness scores for Claude Opus 4, Opus 4.1, Sonnet 4, and Sonnet 4.5 which has the highest score.","caption":"Scores shown here with a realism filter in place. The filter uses a separate AI model to screen out test queries that include obviously unrealistic elements. Without the filter, evaluation awareness scores for Sonnet 4.5 were about 25% higher.","width":1116,"height":778,"is_hero":false,"zone":"main"},{"kind":"image","src_url":"https://anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F39e074699b3d85a830fac6aecf000285936535d6-3840x3840.png&w=3840&q=75","r2_key":null,"alt":"Examples of Claude 3.7 Sonnet Refusals","caption":null,"width":3840,"height":3840,"is_hero":false,"zone":"main"}],"shots":[{"viewport":"mobile","width":390,"height":844,"r2_key":"stealpage/shots/anthropic.com/pg_fz4n6ybai5xw/run_edmfvvawpel5-mobile.png","phash":"82e0d4d8aca4c8c0","visual_delta":null,"captured_at":"2026-09-26T10:07:26.498Z"},{"viewport":"desktop","width":1440,"height":900,"r2_key":"stealpage/shots/anthropic.com/pg_fz4n6ybai5xw/run_edmfvvawpel5-desktop.png","phash":"1fc0c08440606070","visual_delta":null,"captured_at":"2026-09-26T10:07:26.196Z"}],"history":[{"id":"run_edmfvvawpel5","started_at":"2026-09-26T10:07:23.151Z","http_status":200,"fetcher":"direct","changed":true,"diff":{"added":0,"removed":240,"similarity":null},"word_count":1530},{"id":"run_xlchtsgetziy","started_at":"2026-08-27T09:07:01.089Z","http_status":200,"fetcher":"direct","changed":false,"diff":null,"word_count":1770}],"citedBy":[{"id":"pg_zobp_qqgrfys","url":"https://anthropic.com/","title":"Home \\ Anthropic","anchors":["Transparency"],"zone":"main"},{"id":"pg_dph1p5shwony","url":"https://anthropic.com/policy","title":"AI policy","anchors":["Transparency Hub","Transparency"],"zone":"main"}]}