Anthropic

Type: AI Safety Research Lab

Anthropic is an AI safety company known for Claude, their large language model. In this knowledge base, they are represented through their pioneering work on mechanistic interpretability and circuit tracing — the effort to understand how AI models produce their outputs at a mechanistic level.

Anthropic developed a "microscope" that can identify features inside Claude and trace sequences of features from prompt to response, revealing how the model reasons. This progressed from identifying individual features (2024) to tracing complete reasoning paths (2025-2026), making it possible to understand why a model produces a particular output.

Their Transformer Circuits Thread (2021-2026) is the most sustained research program in mechanistic interpretability, progressing through: a mathematical framework for transformer circuits (2021), discovery of induction heads as the basis for in-context learning (2022), toy models of superposition revealing polysemanticity (2022), sparse autoencoders extracting monosemantic features (2023), scaling to millions of features in Claude 3 Sonnet including the Golden Gate Bridge feature (2024), and circuit tracing with attribution graphs revealing end-to-end computational paths (2025). They open-sourced circuit tracing tools for the research community.

They are part of the cross-lab coalition of 40 researchers warning about the growing difficulty of understanding advanced AI models.

Financials (Attributed, Unverified)

Dylan Patel (SemiAnalysis), on Podcast Alpha (2026-07-10), claims: "Anthropic is free cash flow positive" and is "past $50B ARR with 70%+ gross margins." These are Patel claims/estimates, unverified — no primary Anthropic financial disclosure exists in this KB to confirm them. Patel is described in the source as "a large customer with monthly-close specifics," implying direct commercial visibility, but SemiAnalysis is also a disclosed active Anthropic enterprise customer (reported ~$7M/year on Claude Code) — a material conflict of interest on any Anthropic-specific claim from this source. Evidence: weak. (Dylan Patel / Podcast Alpha)

This 70%+ margin figure is notably higher than the ~40-50% industry-wide inference gross margin (training-cost-excluded) estimated by Ben Evans the same week — see Token Pricing & the Inference-Margin Question for the full comparison. Treat the gap as an open question, not a contradiction to resolve by picking a side.

Platform & Model Cadence, Jun 25 – Jul 15 2026 (first-party)

Anthropic's own Claude Platform release notes are the highest-authority source in this KB's July-2026 batch. They are authoritative for what shipped and say nothing about how good any of it is — no benchmark disclosure, no efficacy claim.

Models

• Jun 30 — Claude Sonnet 5 (claude-sonnet-5): $2/$10 per Mtok introductory through 2026-08-31, $3/$15 standard after. 1M-token context, 128k max output. Priority Tier not available. Three migration-relevant behaviour changes: adaptive thinking is on by default with manual extended-thinking budgets removed (400 on attempt); non-default temperature/top_p/top_k now return 400; and a new tokenizer produces ~30% more tokens for the same text.
• Jul 1 — Claude Fable 5 and Claude Mythos 5 access restored after a prior suspension, with a published statement on the redeployment. The Fable 5 / Mythos 5 generation (launched Jun 9) shares the 1M context, always-on adaptive thinking, and the same +30% tokenizer shift; Fable 5 runs safety classifiers that can return stop_reason: "refusal" with no billing on zero-output refusals, plus an opt-in fallbacks parameter re-routing refused requests to another (billed) model.
• Jun 25 — Fast mode for Claude Opus 4.7 deprecated, removal scheduled 2026-07-24 (migrate to Opus 4.8 fast mode).

Agent / harness surface

• Jul 2 — agent-memory-2026-07-22 beta header replaces managed-agents-2026-04-01 on memory-store endpoints (sending both → 400): server-defined stable ordering, depth restricted to 0/1/omitted, segment-exact path_prefix required to end in /, prior page cursors invalidated. All major SDKs send it by default; old behaviour retires 2026-07-22.
• Jul 10 — Dreams (managed-agent research preview) expanded to Claude Fable 5 and Claude Sonnet 5.
• Managed Agents sessions gained opt-in event deltas, backward pagination, per-session config override (model, system prompt, tools, MCP servers, skills), and vault-credential injection_location.

Platform economics & admin

• Jul 14 — Admin API user management (beta, all Claude Enterprise orgs): member lookup/role changes/removal, invites, groups, custom roles (anthropic-beta: ce-user-management-2026-07-13 for group/custom-role calls).
• Jul 8 — API key expiration with preset/custom/"Never" durations and pre-expiry email reminders; expires_at exposed on Admin API key listings.
• Jun 26 — Rate-limit consolidation: Sonnet and Haiku limits now match Opus at every tier; usage tiers collapsed to three — Start, Build, Scale. Anthropic states no org received lower limits.

The number that matters commercially: the ~30% tokenizer expansion. A quoted per-token price that holds flat while ~30% more tokens are billed for identical text is an effective price increase, and it makes per-token comparisons across tokenizer generations — and across vendors — invalid as stated. See Inference & Token Economics.

Key Contributions

• Mechanistic interpretability microscope: Identifies features and traces reasoning paths inside Claude (Mechanistic Interpretability)
• Transformer Circuits Thread (2021-2026): Running research program — mathematical framework, superposition, monosemanticity, circuit tracing (Anthropic Circuit Tracing)
• Circuit Tracing with attribution graphs (2025): Directed graphs revealing step-by-step computational paths in language models, open-sourced for the community (Anthropic Circuit Tracing)
• Sparse autoencoders at scale: Extracted millions of interpretable features from Claude 3 Sonnet; discovered steerable features (Golden Gate Bridge) (Anthropic Circuit Tracing)
• Feature → path progression: Advanced from individual feature identification (2024) to complete reasoning path tracing (2025-2026) (Mechanistic Interpretability)
• Cross-lab interpretability advocacy: Part of 40-researcher coalition calling for CoT investigation (Mechanistic Interpretability)

Mentioned In

• Mechanistic Interpretability — Primary innovator in feature tracing
• Circuit Tracing — Developer of attribution graphs and open-sourced tools
• Chain-of-Thought Reasoning — CoT monitoring for safety; also the source of the adaptive-thinking-only datapoint
• Agent Harness Evolution — Managed Agents / Dreams as commercial harness infrastructure; Claude Opus 4.6 as a tested backbone
• Agent Memory Architectures — the agent-memory beta-header migration
• Inference & Token Economics — Sonnet 5 pricing and the ~+30% tokenizer shift

Related Entities

• Google DeepMind — Collaborator on interpretability research
• OpenAI — Collaborator on interpretability research
• Moonshot AI — Kimi K3 prices into the Sonnet tier and is benchmarked against Claude Fable 5

Changelog

• 2026-07-14 — Added attributed, unverified Financials section (Patel/Podcast Alpha: FCF-positive, $50B+ ARR, 70%+ gross margin) with disclosed-conflict note; cross-linked to the new inference-economics topic's margin-comparison concept page.
• 2026-07-22 — Compiled first-party platform release notes (Jun 25 – Jul 15): added the platform/model cadence section — Sonnet 5 launch and pricing, the ~+30% tokenizer shift, Fable 5 / Mythos 5 restoration, the agent-memory beta-header migration, Dreams expansion, per-session harness override, Admin API user management, key expiration, rate-limit consolidation.