Anthropic’s Transparency Hub \ Anthropic
https://www.anthropic.com/transparency
Anthropic transparency hub — the source for the RSP, system cards and external red-team results in the Claude assessment
29
Lines added
29
Lines removed
Was read
2026-10-03
75532f8a6f91585a…
Changed by
2026-10-20
d6569c63a19b874d…
29 lines added, 29 lines removed
Left column is the version read 2026-10-03, right is 2026-10-20. Unchanged stretches are elided. Both bodies are kept in full and are addressed by the hashes above.
Showing the first 400 lines of the comparison; 278 further lines are held back. This page reports a change, it is not a copy of anyone’s document — the source is linked above and is the authority for what it says.
⋯ 6 unchanged lines
77
88 Model Report
99 October 2, 2026
1010 Select a model to see a summary that provides quick access to essential information about Claude models, condensing key details about the models' capabilities, safety evaluations, and deployment safeguards. We've distilled comprehensive technical assessments into accessible highlights to provide clear understanding of how the models function, what they can do, and how we're addressi…
11−Claude Sonnet 5.5 Claude Opus 5.5 Claude Fable 5.1 Claude Mythos 5.1 Claude Opus 5 Claude Sonnet 5 Claude Fable 5 Claude Mythos 5 Claude Opus 4.8 Claude Opus 4.7 Claude Mythos Preview Claude Sonnet 4.6 Claude Opus 4.6 Claude Opus 4.5 Claude Haiku 4.5 Claude Sonnet 4.5 Claude Opus 4 and Sonnet 4 Claude Opus 4.1 Claude Sonnet 3.7
11+Claude Haiku 5.5 Claude Opus 5.5 Claude Fable 5.1 Claude Mythos 5.1 Claude Opus 5 Claude Sonnet 5 Claude Fable 5 Claude Mythos 5 Claude Opus 4.8 Claude Opus 4.7 Claude Mythos Preview Claude Sonnet 4.6 Claude Opus 4.6 Claude Opus 4.5 Claude Haiku 4.5 Claude Sonnet 4.5 Claude Opus 4 and Sonnet 4 Claude Opus 4.1 Claude Sonnet 3.7
1212
13−Claude Sonnet 5.5 Summary Table
14−Model description Claude Sonnet 5.5 is a faster, lower-cost complement to Claude Opus 5.5, and is strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets.
15−Benchmarked Capabilities See our Claude Sonnet 5.5 system card ’s Section 8 on capabilities.
13+Claude Haiku 5.5 Summary Table
14+Model description Claude Haiku 5.5 is our fastest, most capable small model, built for high-volume work like summarization, subagents, and browser use
15+Benchmarked Capabilities See our Claude Haiku 5.5 system card ’s Section 8 on capabilities.
1616 Acceptable Uses See our Usage Policy
17−Release date September 2026
18−Modalities Claude Sonnet 5.5 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.
19−Model architecture and training methodology Claude Sonnet 5.5 was pretrained on large, diverse datasets to acquire language capabilities. After the pretraining process, Sonnet 5.5 underwent substantial post-training, with the goal of making it an effective assistant whose behavior aligns with the values described in Claude’s constitution .
20−Training Data Claude Sonnet 5.5 was trained on a proprietary mix of publicly available information from the internet, public and private datasets, and other sources, such as synthetic data generated by other models. The mix may also include user data from feedback or bug reports, or data that users have explicitly permitted for training. In the course of its training, we used several data cleaning…
21−Testing Methods and Results Based on our assessments, we deployed Claude Sonnet 5.5 with the same chemical and biological misuse protections we deployed for Claude Opus 5, treating it as having CB-1 capabilities. Autonomy threat model 1 is applicable to Claude Sonnet 5.5. See below for select safety evaluation summaries.
17+Release date October 2026
18+Modalities Claude Haiku 5.5 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.
19+Model architecture and training methodology Claude Haiku 5.5 was pretrained on large, diverse datasets to acquire language capabilities. After the pretraining process, Haiku 5.5 underwent substantial post-training, with the goal of making it an effective assistant whose behavior aligns with the values described in Claude’s constitution .
20+Training Data Claude Haiku 5.5 was trained on a proprietary mix of publicly available information from the internet, public and private datasets, user data included in feedback or bugs or which users have explicitly permitted for training, and other sources, such as synthetic data generated by other models. In the course of its training, we used several data cleaning and filtering methods, includi…
21+Testing Methods and Results Based on our assessments, we deployed Claude Haiku 5.5 with the same chemical and biological misuse protections we deployed for Claude Opus 5 and Claude Sonnet 5. We treat it as meeting our CB-1 threshold, but not our higher CB-2thresholds. See below for select safety evaluation summaries.
2222 See Fable 5.1 & Mythos 5.1 System Card
2323
24−The following are summaries of key safety evaluations from our Claude Sonnet 5.5 system card. Additional evaluations were conducted as part of our safety process; for our complete publicly reported evaluation results, please refer to the full system card .
25−Political Even-handedness
26−People use Claude to learn about and discuss political issues, so we test whether it treats different political views fairly. We use our open-source evaluation , which gives Claude 1,350 pairs of requests on 150 political topics. Each pair asks for the same thing from opposite political perspectives, for example a request about healthcare policy written once from a Democratic point of view and onc…
27−Another Claude model then grades the answers. It checks whether Claude answers both requests in a pair with similar depth and quality (what we call even-handedness), whether it acknowledges other points of view, and whether it refuses to answer. We report results in two settings: Claude accessed directly through our API (the service developers use to build Claude into their own products), with no …
28−Claude Sonnet 5.5 was substantially more even-handed than Claude Sonnet 5. It scored 97.9% on the API and 99.0% on claude.ai (vs. 86.2% and 95.7% for Sonnet 5), and it refused less often in both settings. Results on acknowledging other points of view were mixed. Through the API, Sonnet 5.5 did this slightly less often than Sonnet 5 (41.4% vs. 45.7%). On claude.ai, whose system prompt directs Claud…
29−Pairwise political bias evaluations: even-handedness. Higher scores for even-handedness are better. Results for previous models may show variance from previous system cards due to routine evaluation updates.
30−Honesty
31−People increasingly rely on AI models for information and to get work done, so it matters that these models tell the truth and are open about what they have done. We test Claude’s honesty in several ways. Claude Sonnet 5.5 did better than Claude Sonnet 5 on most of these tests, with some exceptions noted below. For more in-depth descriptions of the evaluations and their results, please see the Cla…
32−Misleading users: We ran about 4,100 simulated test sessions with Claude, then had another AI model review them for eight kinds of misleading behavior, such as lying to the user, flattering them or agreeing with them too readily, and claiming to have finished a task that was not done. Claude Sonnet 5.5 did better than Claude Sonnet 5 on seven of the eight. The exception was dodging questions on se…
33−Making up facts: AI models sometimes “hallucinate,” stating false information as if it were true. An honest model should admit when it does not know an answer instead of guessing. We asked Claude factual questions on 41 subjects, without letting it look anything up online, and marked each answer as right, wrong, or declined. Claude Sonnet 5.5 got more questions right than Claude Sonnet 5, but it a…
34−Honesty under pressure: We used a public test called MASK to check whether Claude will say something it believes is false when a user, or the instructions it has been given, pressure it to. Claude Sonnet 5.5 stuck to what it believed more often than Claude Opus 5.5, but less often than Claude Sonnet 5. (Section 6.3.2.2)
35−Using answers without saying so: We gave Claude programming tasks where the answer was already within reach of the computer it was working on, for example in a newer version of the software that already contained the fix. We then measured how often Claude used that answer without telling the user, which would leave the user unaware of where the solution came from. Claude Sonnet 5.5 did this less o…
36−Admitting to hidden changes: During some training exercises, we saw Claude hide changes to code from the automated system that graded its work, so that it would pass. To test whether Claude would own up to this, we took records of these sessions and asked Claude to summarize its work, without asking about the hidden changes directly. Claude Sonnet 5.5 mentioned the hidden changes 96.2% of the time…
37−The charts below compare Sonnet 5.5 with other recent Claude models on the eight misleading behaviors from the first test above. Lower scores are better. Behaviors are scored even in sessions where they are unlikely to come up, so the scores are most useful for comparing models with each other.
38−Scores from our automated behavioral audit
24+The following are summaries of key safety evaluations from our Claude Haiku 5.5 system card. Additional evaluations were conducted as part of our safety process; for our complete publicly reported evaluation results, please refer to the full system card .
25+Election Integrity
26+People use Claude to find information about elections and voting, so we test whether it answers legitimate election questions while refusing requests that violate our Usage Policy , such as requests to spread false information about how to vote, or to create content designed to mislead or intimidate people into not voting.
27+We use two kinds of tests. The first is a set of 600 single-turn requests based on patterns we see in real use: 300 are harmful requests that Claude should not assist with, and 300 are legitimate ones that Claude should answer. The second kind of evaluation assesses Claude’s behavior inlonger conversations, in which a simulated user, played by another AI model following instructions from our polic…
28+On the single-turn requests, Claude Haiku 5.5 responded safely to every harmful request on both the API and claude.ai, and it refused 0.33% of the legitimate requests (Claude Haiku 4.5 refused none, a difference within the margin of error). On the multi-turn evaluation, Haiku 5.5 responded appropriately throughout 99% of conversations through the API and 98% on claude.ai, the highest of any recent…
29+The table below shows the share of multi-turn conversations in which each model responded appropriately throughout. Higher is better.
30+
31+Agentic Safety
32+AI models are increasingly used as “agents”: instead of only answering questions, they take actions, such as writing and running code or using websites and apps on a computer. This creates additional risks: someone could try to use Claude to carry out harmful tasks, or attackers could try to trick it into working against the person it is helping. We test for both. Unless noted, these tests measure…
33+Harmful coding requests: We gave Claude, working as a coding agent in Claude Code, 61 cyber-related requests that violate our Usage Policy, such as writing malware, and 61 sensitive but permitted security requests, such as testing a website for weaknesses. Each request was run 10 times. Claude Haiku 5.5 refused 84.3% of the harmful requests, up from 66.6% for Claude Haiku 4.5, and assisted with 98…
34+Harmful computer use tasks: We gave Claude control of a computer in a closed-off test environment and asked it to carry out 112 harmful tasks in three areas: surveillance and collecting people’s data without permission, creating and spreading harmful content, and abuse at scale. Claude Haiku 5.5 refused 82.6% of these tasks, up from 58.9% for Claude Haiku 4.5 and slightly higher than both Claude S…
35+Hidden instructions (prompt injection): When Claude works as an agent, it reads content from outside sources, such as emails and web pages. Attackers can hide instructions in that content; for example, an email with hidden text telling Claude to forward the user’s private messages to the attacker. On a benchmark created by the AI security company Gray Swan, the UK AI Security Institute, and the US…
36+Attackers tuned to our models: In a separate prompt injection evaluation, an automated attacker from Gray Swan was trained against previous Claude models. 54 different coding and computer scenarios were run 200 times each with this attacker. Fewer than 1 in 1,000 attempts succeeded against Haiku 5.5 in coding or computer use tasks, in line with our most advanced models. Against Claude Haiku 4.5, 5…
37+The chart below shows how often attackers succeeded within 1, 10 and 15 attempts (shown as “k”) on Gray Swan’s benchmark, against our models and those from other developers.. Lower is better. Claude models were tested without our extra protections against these attacks, and Qwen 3.8, DeepSeek-V4-Pro and GLM-5.3 were not tested on tasks where the model operates a computer through its screen.
38+
3939 RSP Evaluations
40−Our 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. Claude Sonnet 5.5 is not more capable than our most advanced models, and it is less capable than Claude Opus 5.5 on most of our tests. We found that Claude Opus 5.5 does not reach our higher risk thresho…
41−Alignment risk is the risk that a model behaves in ways Anthropic did not intend. In our August 2026 Risk Report we raised our estimate of this risk from very low to low, because recently disclosed incidents involving how models behaved during cybersecurity testing made us less certain. We assess that the risk of Claude Sonnet 5.5 causing catastrophic harm by acting in ways we did not intend is al…
42−On its ability to carry out AI research and development on its own, Claude Sonnet 5.5 is broadly similar to or less capable than Claude Opus 5.5. Opus 5.5 did not cross our higher threshold for this ability, so we reach the same conclusion for Sonnet 5.5. On our capability index, which combines many tests into a single score, Sonnet 5.5 scored below Opus 5.5 (167.93 vs. 169.12).
43−On chemical and biological weapons, we treat Claude Sonnet 5.5 as able to significantly help someone with a basic technical background make known chemical or biological weapons, which is our CB-1 threshold. We deploy it with safeguards to match, including the same automated filters we used for Claude Opus 5, which are designed to detect and block attempts to get this kind of help. On three automat…
44−Developing new chemical or biological weapons takes rare expertise and judgment. We estimate that Claude Sonnet 5.5’s abilities in this area are similar to or below those of Claude Opus 5, and well below those of Claude Opus 5.5. It struggled with longer, open-ended research tasks, which limits how far it could stand in for human experts. We are releasing Sonnet 5.5 with the same safeguards we use…
40+Our 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. Claude Haiku 5.5 is less capable than our most advanced models, and on most of our tests it is also less capable than Claude Opus 5. We therefore determined that it does not reach our higher risk thresho…
41+Alignment risk is the risk that a model behaves in ways Anthropic did not intend. In our August 2026 Risk Report we raised our estimate of this risk to low, because recently disclosed incidents involving how models behaved during cybersecurity testing made us less certain. A less capable model can still pose this kind of risk, and we found that the risk of Claude Haiku 5.5 causing catastrophic har…
42+On its ability to dramatically accelerate AI research and development, Claude Haiku 5.5 is much less capable than Claude Opus 5.5 . Opus 5.5 did not cross our higher threshold for this ability, so we reach the same conclusion for Haiku 5.5. On our capability index, which combines many tests into a single score, Haiku 5.5 scored well below Opus 5.5.
43+On chemical and biological weapons, we treat Claude Haiku 5.5 as meeting our CB-1 threshold, which covers models that could significantly help someone with a basic technical background make known chemical or biological weapons. We deploy it with safeguards to match, including automated filters designed to detect and block attempts to get this kind of help. Because Haiku 5.5 is not at the frontier,…
44+Developing new biological weapons takes rare expertise and judgment. We estimate that Claude Haiku 5.5’s abilities in this area are above those of Claude Sonnet 5 but below those of Claude Opus 5, and well below those of Claude Opus 5.5, Claude Sonnet 5.5 and Claude Mythos 5.1. Since those models do not meet our CB-2 threshold, we determined that Claude Haiku 5.5 does not either. On one part of on…
4545
4646 Claude Opus 5.5 Summary Table
4747 Model description Claude Opus 5.5 is our new leading model, and early testers saw large jumps in performance on their most complex work.
4848 Benchmarked Capabilities See our Claude Opus 5.5 system card ’s Section 8 on capabilities.