Anthropic
Anthropic has 21 model versions on record; all 21 have at least one documented fact.
Of the 21 versions listed below, 21 document a published framework; 19 document threshold determinations; 12 document tests by an independent or government party; 20 document tests the provider ran itself; 4 document safeguards the provider describes.
Model versions with documented facts
| Model | What is documented |
|---|---|
| Claude Sonnet 5.5 · Released 28 September 2026 · 1 document | |
| Claude Opus 5.5 · Released 22 September 2026 · 4 documents | |
| Claude Fable 5.1 · Released 1 September 2026 · 7 documents | |
| Claude Mythos 5.1 · Released 1 September 2026 · 1 document | |
| Claude Opus 5 · Released 24 July 2026 · 7 documents | |
| Claude Sonnet 5 · Released 30 June 2026 · 6 documents | |
| Claude Fable 5 · System card dated 9 June 2026 · 8 documents | |
| Claude Mythos 5 · System card dated 9 June 2026 · 7 documents | |
| Claude Opus 4.8 · Released 28 May 2026 · 9 documents | |
| Claude Opus 4.7 · Released 16 April 2026 · 13 documents | |
| Claude Mythos Preview · Released 7 April 2026 · 8 documents | |
| Claude Sonnet 4.6 · Released 17 February 2026 · 9 documents | |
| Claude Haiku 5.5 · 1 document | |
| Claude Opus 4.6 · 16 documents | |
| Claude Haiku 4.5 · 5 documents | |
| Claude Opus 4.5 · 10 documents | |
| Claude Sonnet 4.5 · 8 documents | |
| Claude Opus 4.1 · 5 documents | |
| Claude Opus 4 · 4 documents | |
| Claude Sonnet 4 · 1 document | |
| Claude Sonnet 3.7 · 4 documents |
Framework
7 of 7 framework documents listed here are stored as "proposed": a model read their level names and no person has confirmed them.
- Responsible Scaling Policy v2.2 Same levels in: Responsible Scaling Policy v2.1.
Levels named in the stored text
- CBRN-3 (Chemical, biological, radiological and nuclear)
- CBRN-4 (Chemical, biological, radiological and nuclear)
- Cyber operations (Cyber)
- AI R&D-4 (Autonomy & replication)
- AI R&D-5 (Autonomy & replication)
- Model autonomy checkpoint (Autonomy & replication)
- ASL-2
- ASL-3
- ASL-4
- Responsible Scaling Policy, v3.4
Levels named in the stored text
- ASL-3 (Chemical, biological, radiological and nuclear)
- Automated R&D in key domains (Autonomy & replication)
- Misaligned AI systems in high-stakes settings (Deception & sabotage)
- Responsible Scaling Policy v3.3
Levels named in the stored text
- ASL-3 (Chemical, biological, radiological and nuclear)
- CBRN-3 (Chemical, biological, radiological and nuclear)
- AI R&D-4 (Autonomy & replication)
- ASL-2
A stored sentence about this framework mentions a risk domain that has no level in the list.
- Responsible Scaling Policy v3.2
Levels named in the stored text
- ASL-3 (Chemical, biological, radiological and nuclear)
- CBRN-3 (Chemical, biological, radiological and nuclear)
- Automated R&D in key domains (Autonomy & replication)
- High-stakes sabotage opportunities (Deception & sabotage)
- ASL-2
- AI R&D-4
A stored sentence about this framework mentions a risk domain that has no level in the list.
- Responsible Scaling Policy v3.1
Levels named in the stored text
- ASL-3 (Chemical, biological, radiological and nuclear)
- Ability to fully automate entry-level AI research work
- Ability to cause dramatic acceleration in the rate of effective scaling
A stored sentence about this framework mentions a risk domain that has no level in the list.
- Responsible Scaling Policy v3.0
Levels named in the stored text
- ASL-3 (Chemical, biological, radiological and nuclear)
- CBRN-3 (Chemical, biological, radiological and nuclear)
- Automated R&D in key domains (Autonomy & replication)
- AI R&D-4 (Autonomy & replication)
- High-stakes sabotage opportunities (Deception & sabotage)
A stored sentence about this framework mentions a risk domain that has no level in the list.
Abbreviations in these level names
- ASL: AI Safety Level
- A graded set of safety and security measures that become stricter as model capabilities increase.
we use AI Safety Level Standards (ASL Standards), graduated sets of safety and security measures that become more stringent as model capabilities increase.
Anthropic · Read the source
Each test and safeguard, by who documents it
For each item, how many of the 21 versions listed above whose documents we read in full have it from an outside party (an independent or government tester), from the provider, or as a claim we could not confirm.
- Published framework with capability thresholds (Governance): provider 21
- Threshold determination for this release (Governance): provider 20
- Pre-release external access (Governance): outside party 2; provider 4
- Tested: chemical, biological, radiological, nuclear (Testing): outside party 3; provider 14
- Tested: cyber capabilities (Testing): outside party 4; provider 12
- Tested: autonomy and replication (Testing): outside party 2; provider 12
- Tested: deception and sabotage (Testing): outside party 7; provider 11
- Tested: misuse and jailbreak (Testing): outside party 3; provider 10
- Tested: harmful content (Testing): outside party 1; provider 15
- Tested: injection and agentic risks (Testing): provider 13
- Tested: bias (Testing): provider 12
- Tested: privacy (Testing): provider 1
- Safety training (harmlessness and refusal) (Safeguards): provider 2
- Staged or restricted access (Safeguards): provider 2
No document we read states, for any of these versions: Serious-incident disclosure commitment; Agent permission limits and sandboxing; Human approval for high-risk actions; Misuse monitoring and enforcement; Reasoning (chain-of-thought) monitoring; Output provenance and watermarking; Telemetry and audit logging; Domain-specific safeguards (CBRN or cyber); Vulnerability disclosure and bug-bounty programme; Third-party or government red-team access during development.