Unboxing the AI Stack
Organization · L02 · Language Models

Anthropic

Anthropic is a US AI lab that makes the Claude family of language models. It describes itself as “an AI safety and research company” and is organized as a public benefit corporation, whose stated purpose is “the responsible development and maintenance of advanced AI for the long-term benefit of humanity”.1 It was founded in 2021 by former OpenAI staff, according to Wikipedia,2 and its registered address is in San Francisco.3 It released Claude on March 14, 2023, in two versions, Claude and the lighter Claude Instant.4 Its current lineup has four models: Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5.5

Claude models post several of the measured records MetaPrompt tracks. Claude Opus 5 scores 30.16% on ARC-AGI-3’s Semi-Private set on ARC Prize’s Standard harness, third behind two OpenAI models, and ARC Prize says it was the highest-scoring model as of July 24, 2026 (Highest ARC-AGI-3 Scores, by Model).67 Claude 4.5 Opus has the top SWE-bench Verified score in the shared mini-SWE-agent v2.0.0 harness, 76.8%, and Opus 5.5 in Claude Code tops the Terminal-Bench 4.0 leaderboard at 64.85% (SWE-bench Verified, Pro and What Replaced Them).89 Its list prices run from $0.50 per 1M output tokens for Haiku 5.5 to $50 for Fable 5.1 (Cheapest LLM APIs per 1M Output Tokens).10 On OpenRouter, Claude Opus 5.5 processed 5.35 trillion tokens in the week to October 12, 2026, up 106% on the week before (Most-Used AI Models on OpenRouter by Weekly Tokens).11

Anthropic trains and serves its models on partners’ data centers. Epoch AI estimates 910 MW of IT power in operation at the Anthropic-Amazon campus in New Carlisle, Indiana, the second-largest site in its tracking, on Amazon Trainium2 chips; Amazon describes its Project Rainier cluster there as nearly half a million Trainium2 chips (Largest AI Data Centers by Estimated Power Draw).1213 In May 2026 Anthropic agreed to use all of the compute at xAI’s Colossus 1, which it put at “more than 300 megawatts” and over 220,000 NVIDIA GPUs.14

Illustration: Anthropic sits in the Language Models layer of the stack.

Key facts

Founded20212
Legal formPublic benefit corporation (Anthropic PBC)13
Registered addressSan Francisco, California3
First Claude releaseMarch 14, 20234
Current modelsClaude Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.55
Highest ARC-AGI-3 score (Standard harness)30.16% (Claude Opus 5, High effort) Measured6
Largest site used, IT power in operation910 MW (Anthropic-Amazon New Carlisle) Estimate12

As of . The founding year comes from Wikipedia, a secondary source; New Carlisle’s power is Epoch AI’s estimate. How figures are labeled.

Appears in

Sources

  1. Anthropic, Company, retrieved 2026-10-13.
  2. Wikipedia, Anthropic (secondary source; no Anthropic page we opened states its founding year), retrieved 2026-10-13.
  3. Anthropic, Privacy Policy, retrieved 2026-10-13.
  4. Anthropic, Introducing Claude, March 14, 2023, retrieved 2026-10-13.
  5. Anthropic, Models overview (Claude API docs), retrieved 2026-10-13.
  6. ARC Prize, ARC-AGI-3 Verified leaderboard data (v3.json, generated 2026-10-05), retrieved 2026-10-13.
  7. ARC Prize, Claude Opus 5 results, retrieved 2026-10-13.
  8. SWE-bench team, SWE-bench Leaderboards (Bash Only view, mini-SWE-agent v2.0.0 runs), retrieved 2026-10-13.
  9. Terminal-Bench, Terminal-Bench 4.0 leaderboard, retrieved 2026-10-13.
  10. Anthropic, Pricing (Claude API docs), retrieved 2026-10-13.
  11. OpenRouter, AI Model Rankings (“This Week” view, usage data through Oct 12, 2026; rankings data CC BY 4.0), retrieved 2026-10-13.
  12. Epoch AI, Anthropic-Amazon New Carlisle (Frontier Data Centers directory), CC BY 4.0, retrieved 2026-10-13.
  13. Amazon, AWS Project Rainier, retrieved 2026-10-13.
  14. Anthropic, announcement of the SpaceX Colossus 1 agreement, May 2026, retrieved 2026-10-13.