AI-Wiki

Anthropic

13 min read

Anthropic

Confidence 0.95 · 34 sources · last confirmed 2026-10-03

AI safety and research company; publisher of the Claude family of models and the Anthropic Economic Index research series. Founded 2021 by Dario Amodei and Daniela Amodei (with a cohort of OpenAI departures) following the Amodei team’s exit from OpenAI “two or three OpenAI crises ago” per Ries 2026. Public Benefit Corporation with a two-tier governance structure (see §Governance structure below). Partnered with Amazon Web Services (~$8B total invested as of 2024 per 2026-04-28-werner-lebrun-octopus-organization). A new (not-yet-fully-released) Anthropic model, Mythos, surfaces in How I AI (June 2026) as the model credited with Mozilla’s Firefox security-bug spike — though Grinstead splits the credit ~50/50 between the model and the harness/pipeline.

Governance structure

Per Ries 2026 (who advised the Amodeis on the structure during Anthropic’s founding):

“Anthropic has directors on its for-profit board who are appointed by and are accountable to an outside group of trustees who are AI safety experts who do not have equity in Anthropic.”

The mechanism: Long-Term Benefit Trust (LTBT) as the mission guardian entity. Two-tier:

  1. For-profit board — the standard corporate governance layer; directors operate under PBC fiduciary obligations.
  2. Long-Term Benefit Trust — outside trustees who (a) are AI safety experts, (b) hold no equity in Anthropic, and (c) appoint and hold accountable the for-profit board’s directors.

The novel structural property: trustee accountability flows the right direction with respect to mission-relevant decisions — directors are accountable to the trust (not to shareholders) on the mission axis. Ries’ first-hand framing of what this buys: “Whenever you see Anthropic do the right thing, like when they refuse to release a model because they think it’s too dangerous, think about how much that’s costing them.”

Ries (who plays “no important role” but “a very big role” in the founding-advice phase) was approached by “one of their investors” and told the Amodeis “if you don’t get this right, here’s what’s going to happen” — the early framing that became the LTBT. Per Ries’ framework, PBC + LTBT is strictly stronger than PBC alone (OpenAI’s post-2025 structure) because the LTBT introduces outside trustee accountability that PBC-alone does not.

Research initiatives appearing in this wiki

Platform / product engineering

  • Claude Platform — Anthropic’s developer-facing platform.
    • Claude Managed Agents — hosted service for long-horizon agent work; brain/hands/session decoupled architecture; published April 2026 (Engineering blog); by July 2026 the service carries a shipped outcomes primitive (rubric + iteration/spend budget), a “dreaming” memory-distillation concept, and an evolving agent identity / service-account model — per the Claude Platform team’s own panel discussion (Building the future of agentic infrastructure, the wiki’s first first-party Claude Platform interview: Jess Yann — PM, Claude Managed Agents; Katelyn Lesse — Head of Engineering, Claude Platform; Angela Jiang — Head of Product, Claude Platform; all three dangling, single-source, deferred).
    • Claude Agent SDK — the packaging of Claude Code’s building blocks for third-party agents. Described first-hand by Isabella He (Applied AI) in [[2026-08-19-he-databricks-anthropic-primitives-to-production-agents|From Primitives to Production]] (recorded April 2026, published August 2026) as “the same building blocks as Claude Code, but allows you to put in your custom tools and custom system prompts and custom skills.” The same talk is the wiki’s fullest first-party account of Anthropic’s agent design intent: the LLM → code-orchestrated-workflow → agent progression; the domain-specialisation reversal (“we used to think you might need almost an entire new harness for something like a research agent… instead of breaking down agents by domain, we actually see that it’s more effective if we just think about agents as this really almost general purpose”) with code execution, file system, web search and to-do list named as the domain-invariant primitives; skills as progressive disclosure against context pollution; subagents as context isolation rather than delegation; hooks as “a way for you to inject a little bit of determinism into your agent”; sandboxing as permission-prompt reduction; and evals framed as the model-upgrade mechanism rather than as a quality mechanism. Includes a live SRE-agent demo that closes the loop end to end (detect → diagnose → edit → redeploy → verify → resolve the page → publish a postmortem via MCP). No measurements of any kind — see the source page’s scope warning.
    • Claude Code — agentic coding harness; described as “an excellent harness” in the Managed Agents post; its agentic architecture (splits coding work into smaller API calls labeled as distinct tasks) is the empirical signature of agent-mediated work in the 5th Economic Index report. Engineering leadership: Boris Cherny (10–15 concurrent Claude instances + CLAUDE.md as in-workflow learning capture per Kiron-Schrage 2026) and Fiona Fung (Director of Engineering; documents the Claude Code team-norms rewrite — JIT planning, code-wins-over-whiteboard debate, manager-starts-as-IC dogfooding, “Claudify everything”, “explicit permission to kill processes” — in 2026-05-08-running-an-ai-native-engineering-org). Founder-vantage worked examples on Claude Code as substrate: at $100M-ARR product scale by Jha at Emergent (multi-agent Kubernetes harness, system rewritten 4× in 9 months); at 2-FTE-startup-internal-ops scale by Garg at AnswerThis (Claude Code CLI wrapped in Python with a self-extending coding sub-agent and an agent-editable instructions.md).
  • Claude Cowork — the non-coding surface on the Claude Code runtime, not a separate product line. First-party origin and architecture from Lydia Hallie in The Agent Factory, July 2026: Anthropic noticed “back in December” (2025) that “a lot of our employees were using Claude Code for non-coding tasks — the marketing team was using it to create marketing files, data science, and so on”, so “we wanted to build a better harness for these types of tasks. And that’s how Cowork came out in a week, because we of course use Claude Code to build Claude Cowork as well.” It “still uses the Claude Code runtime underneath, so it can still do coding… we just have better connectors and better system prompts in Cowork to focus on that non-technical stuff.” Two things follow: the split is coding-vs-not by context configuration, not by engine, which is why Blum’s “it’s not Cowork in itself… Codex could be great for this as well” is architecturally literal; and Anthropic uses the harness vocabulary in its own voice, for a term it did not coin. (An earlier characterisation of Cowork on this page as a “collaborative-design product” was wrong and has been corrected.) Also referenced by Spiegel 2026 in the designers-shipping-code discussion (alongside the named team member Jenny Wen, head of design at Claude, ex-Figma director — the move from Figma director to Claude IC designer is narrated by Spiegel as a case study in the crits-as-core-skill dimension of post-AI design work).

Third-party uses of the Anthropic API

  • WikiZZ / LLM WikiZZ (Mysore 2026) — single-author open-source browser-only extension of Karpathy’s LLM Wiki pattern that routes API requests via a Cloudflare Worker CORS proxy to NVIDIA NIM, Anthropic, and Gemini as user-selectable provider backends. The wiki’s first third-party-developer-tool-uses-Anthropic-API mention in the LLM Wiki cluster. Surfaced here as a one-line context note; the substantive treatment is on the source page.

  • Khan Academy (Sal Khan interview, July 2026) — a $1.2M/year run rate; per Khan, Anthropic told the nonprofit it is “the top of the stack” among organizations leaning into agentic code review, with engineers “running five, six, seven, eight, nine, 10 agents simultaneously writing code, reviewing code.”

  • Polsia (Bloomberg Originals (Sep 2026)) — a San Francisco agent-automation startup whose founder chose Anthropic at launch (“I need to give my customers the absolute best”) and left when viral growth pushed the bill to $1–1.5M a month. Moving to Chinese open models cut it to about $100k. The corpus’s clearest case of the scale point at which a frontier API stops being affordable. Bloomberg’s cost test priced one task at about $50 on Claude Fable 5, against about $12 on Kimi K3 and about $4 on OpenAI’s GPT-5.6 Luna.

Models referenced in this wiki

  • Claude Sonnet 4.5 — predominant model in the November 2025 Economic Index sample. Referenced as exhibiting “context anxiety” (premature task wrap-up) in the Managed Agents engineering post.
  • Claude Opus 4.5 — released between the 4th and 5th Economic Index sample windows. The Managed Agents engineering post notes that the “context anxiety” seen on Sonnet 4.5 was not present on Opus 4.5.
  • Claude Opus 4.6 — released coincident with the 5th Economic Index sample window (Feb 2026).
  • Family-level cost/speed/performance tradeoff: Haiku (fast, cheap) → Sonnet (default) → Opus (most capable, higher per-token price). The 5th Economic Index report quantifies that users select Opus differentially for higher-value tasks: +1.48 pp Opus per +$10/hour task value (Claude.ai); +2.79 pp per +$10 (1P API — about twice as steep).

The Mythos restriction sequence (added 2026-09-15)

Mythos enters the wiki twice, and the two appearances are a year apart in tone. How I AI (June 2026) credits it with Mozilla’s Firefox security-bug spike, splitting the credit roughly evenly between model and harness. WEF Radio Davos (September 2026) treats the same capability as the reason the model became a governance object, and reads the sequence as two phases with different governors:

  • Phase one — self-restraint. The code-generating capability “posed a threat to the computer security of not just the country but of the globe, in terms of this capability could be harnessed to hack into any system,” and “the platform proactively said, as a responsible actor, we are not going to release it.”
  • Phase two — state control. “The US government has decided that it may be a threat to national security for non-US citizens — not just people in other countries but non-US citizens — to have access to it, and Anthropic has made the choice to simply not make it available to anyone… in part because they have a number of non-US citizens who work for them.”

Two details worth recording about Anthropic specifically. First, the compliance route — withdrawing the model from everyone — was determined by the company’s own workforce composition rather than by any property of the risk: a restriction framed around who may access the model was satisfiable only by making it available to nobody. Second, the episode places Anthropic on both sides of the governance line within months: the actor exercising voluntary restraint in phase one, and the object of an involuntary instrument in phase two.

Sundararajan’s reading of what it portends is explicitly provisional — “we’re going to be making it up as we go along, and it’s going to be both the government doing that and the platforms doing it.” Caveat: this is an interview account, recalled from memory, with no document cited. The wiki has not ingested primary reporting on the restriction, and Anthropic’s own account of it is not represented here at all. See responsible-ai for the governance-authority framing and ai-sovereignty for the state-capacity one.

After the restriction: an open-weight match, as relayed (added 2026-10-03). Chow (Metalware), at YC Root Access on 30 September 2026, told a defense audience that Anthropic “published a report literally yesterday saying that Mythos-level capabilities are now available for download, open source, for all of our adversaries to use, with similar performance to Mythos.” The report he means appears to be Anthropic’s post of 29 September 2026, GLM-5.3 and the spread of advanced cyber capabilities (read on the web 2026-10-03, not ingested). It compares Zhipu AI’s open-weight GLM-5.3 with Mythos Preview on exploit benchmarks and argues for giving defenders wider access to Claude’s cyber capabilities. If that identification holds, the sequence on this page has a third step: the company that withheld Mythos is now the one documenting a comparable capability in someone else’s downloadable model. The wiki holds a forecast of this, from Jones on 1 May (“in six months, in a year”), and an earlier version of the claim, from Loukides’s May digest.

A model release that triggered an architecture demolition (added 2026-09-16)

Kavak, August 2026 supplies an unusual datapoint about what a model release does downstream. Kavak had “tens of thousands” of function-decomposed agents running its business at scale. Then, in Maza’s account, Opus 4.5 changed the right architecture:

“Opus 4.5 came out and I realized this isn’t the right paradigm anymore — the intelligence now doesn’t need the graph and the multi-agent lattice work and harness, because it will constrain this level of intelligence.”

Kavak scrapped two years of working infrastructure and rebuilt around one long-running agent per customer. This is the wiki’s first account of a named Anthropic release causing a customer to remove scaffolding rather than add it — the capability-overhang argument made concrete and expensive. Anthropic is otherwise absent from that episode; the claim is the customer’s. See agent-harness and multi-agent-failure-modes.

As a flow-state organisation

Two recent sources name Anthropic specifically as a flow-state organisation — a place where structural pressure + lack of hierarchy strips away traditional thinking:

  • Ries 2026 frames it from the governance layer: PBC + LTBT make the structural mission-protection real.
  • Chamath 2026 names it from the operating-structure layer: “if you find one of these places — Anthropic is such a place, OpenAI is such a place, Facebook was such a place, Google was such a place, SpaceX is such a place” — alongside no-org-chart and chronic-under-hire as deliberate organisational design.

The two framings are at different layers (governance vs operating-structure) but converge on Anthropic as the AI-era exemplar of the category.

Named from outside, September 2026 (added 2026-09-21)

  • 2026-09-15-krishna-bain-winning-with-ai-era-of-experimentation-is-over — IBM’s CEO agrees with Anthropic that LLMs are good at translating legacy code such as COBOL, but disputes the inference that this threatens the mainframe: mainframes are used for transaction volume, not because of language lock-in.

  • 2026-09-18-swisher-lafrance-atlantic-festival-tech-economy-is-ai-reshaping-america — Kara Swisher predicts the seven frontier labs will consolidate to two, and names Anthropic as “probably” one of the survivors alongside Google. Opinion, no figures.

  • 2026-09-30-michael-miller-department-of-war-ai-adoption — US Under Secretary of War Emil Michael names “Anthropic’s leadership” among those who signed a White House compact “for the safe development of superintelligence” the day before the YC event (livestream published 30 September 2026), committing, in his account, to “auditing and independent evaluations”. His description only; the compact’s text is not in the wiki. It is the wiki’s second account this quarter of Anthropic in US AI governance, after the Mythos sequence above.

Open questions

  • The wiki has multiple references to Claude as a measurement substrate but no primary source on Anthropic itself yet; this entity page is a stub awaiting first-party Anthropic source ingestion.

Cited as its own customer (added 2026-09-28)

Garry Tan’s “20x company” episode opens with an Anthropic engineer’s post that Claude wrote Claude Cowork and that developers each manage “anywhere between three and eight Claude instances”. Tan uses it as the example that sets his whole argument up.

Graph View

Backlinks