0

Anthropic

Frontier AI safety lab building the Claude family of models, founded by former OpenAI researchers.

Visit
Type
frontier lab
HQ
San Francisco, CA, USA
Founded
2021
Funding
Private, ~$300B+ valuation 2025; backed by Google, Amazon, Lightspeed, Menlo
Website
anthropic.com
GitHub
github.com/anthropics
Twitter
twitter.com/AnthropicAI

Cite

Notes

Only stored in your browser.

Evals
3
Tools
1
Models
38
Papers
1
Boards
0
People
47

Models

38

Ranked by Artificial Analysis Intelligence Index

#ModelReleasedIntelligence
1Claude Opus 524 Jul 202663.1
2Claude Fable 59 Jun 202662.1
3Claude Opus 4.828 May 202657.3
4Claude Sonnet 530 Jun 202655.3
5Claude Opus 4.716 Apr 202655
6Claude Opus 4.65 Feb 202638.8
7Claude Opus 4.524 Nov 202535.6
8Claude Sonnet 4.617 Feb 202635.1
9Claude Sonnet 4.529 Sep 202529.9
10Claude 4.1 Opus5 Aug 202528.8
11Claude 4 Sonnet22 May 202526
12Claude 4 Opus22 May 202526
13Claude 4.5 Haiku15 Oct 202524.1
14Claude Sonnet 3.724 Feb 202523.9
15Claude 3.5 Haiku22 Oct 202412.2
16Claude Opus 34 Mar 202411.8
17Claude 3.5 Sonnet22 Oct 20249.8
18Claude 3.5 Sonnet (June '24)21 Jun 20248.1
19Claude Sonnet 34 Mar 20244.4
20Claude 3 Haiku4 Mar 20243.5
21Claude 2.121 Nov 20233.5
22Claude 2.011 Jul 20233.3
23Claude Instant14 Mar 20231.7
24Claude Haiku 4.5--
25Claude Sonnet 4 5 Search--
26Claude Opus 4 1 Search--
27Claude Opus 4 6 Search--
28Claude Sonnet 4 6 Search--
29Claude Opus 4 1 20250805 Thinking--
30Claude Sonnet 4 20250514 Thinking--
31Claude Sonnet 5 Search--
32Claude Opus 4 5 Search--
33Claude Opus 4.1--
34Claude Opus 4 Search--
35Claude Opus 4 20250514 Thinking--
36Claude Opus 4--
37Claude Sonnet 4--
38Claude Mythos Preview--
38 / 38 models

Evals

3

Sycophancy Eval

Evaluate sycophancy of language models across a variety of free-form text-generation tasks.

ActiveAssistants

Anthropic HH-RLHF

Anthropic's helpful & harmless preference dataset - paired human-rated assistant responses widely used both as a preference-training corpus and as a reward-model benchmark.

ActiveSafetyHarmful ContentInstruction Following

Agentic Misalignment: How LLMs could be insider threats

Eliciting unethical behaviour (most famously blackmail) in response to a fictional company-assistant scenario where the model is faced with replacement.

ActiveScheming

Tools

1

Anthropic HH-RLHF

Anthropic's foundational helpful-and-harmless human preference dataset - the first major public RLHF corpus and a long-time community baseline.

PreferenceSafetyJailbreak ResistanceMulti Turn Dialog

Papers

1

Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback

preprint

2022