Anthropic

AI lab. Maker of the Claude model family and Claude Code. Internal incubator Anthropic Labs produced Claude Code, MCP, and the desktop app.

note Model frontier moved (June 2026) Earlier sources in this wiki treat Opus 4.7/4.8 as current SOTA. As of June 9, 2026 the frontier is the Mythos-class tier (Fable 5 / Mythos 5), which sits above Opus — see the release section below.

Mythos-class release: Fable 5 & Mythos 5 (June 9, 2026)

Anthropic launched its first generally-available Mythos-class model, Claude Fable 5 ("a Mythos-class model we've made safe for general use"), alongside the restricted Claude Mythos 5 (same model, cyber safeguards lifted) via Project Glasswing. This is the public arrival of the "Mythos" model the IPO/Apocalyptic-Capitalism notes below had flagged as withheld.

Pricing: $10/M input, $50/M output (≈2× Opus; half the price of the April Mythos preview).
Free on subscriptions only until June 22, 2026, then usage-credit billing — Anthropic cites compute capacity (and, per the creators, IPO/profitability pressure).
Anti-distillation: flagged distillation attempts fall back to Opus 4.8. New 30-day data-retention policy for Mythos-class traffic.
Covered via three YouTube reactions: Claude Mythos is Finally Here (Nate Herk), MYTHOS MYTHOS MYTHOS (Matthew Berman), and Fable 5 Raises the Bar for AI Ambition. All read the staggered release as Anthropic protecting a research lead (ties to Recursive Self-Improvement).
Live launch controversies (per Fable 5 Raises the Bar for AI Ambition): biology-guardrails over-trigger (mitochondria!), invisible AI-research-distillation degradation, 30-day-retention NDA issue. See Mythos-Class Models for the full controversy section.

The backlash and walkback (June 11–12, per Why Fable 5 is the Most Controversial AI Release Ever)

Whittemore calls Fable 5 "easily the most controversial AI model launch of all time" — bigger than GPT-5's. Within 24 hours Anthropic reversed the silent degradation of AI-development requests (to Wired: "We made the wrong trade-off, and we apologize for not getting the balance right" — safeguards now visible). The damage assessment:

Microsoft restricted employees from Fable 5 in Copilot over the 30-day retention policy (The Verge) — the retention issue remains unresolved and is the bigger enterprise-revenue risk pre-IPO.
Trust blast radius (Dean Ball): wider than Anthropic — every lab's evals now carry the silent-intervention asterisk. Hugging Face's Arthur Zucker: "you broke our trust and I don't think you'll ever get it back."
The final-arbiter problem: Dario's "policy on the AI exponential" essay + Bloomberg's 47-min documentary landed mid-firestorm and amplified the reading that Anthropic wants to gate who gets frontier capability. GMU's Samuel Roman: if Anthropic becomes the tollbooth for frontier access, the state will read that as competition — "Anthropic does not win that fight."
Matan Grinberg (Factory CEO): "Anthropic's speedrun to becoming the bad guy should be studied" — from unassailable position to this in months.

The June 12 export ban — Anthropic becomes a foreign-policy lever

On June 12, 2026 (per AI Has Granted America Vast New Power (Economist), the issue's cover Leader), the Trump administration ordered Anthropic to block all foreigners from accessing Fable and Mythos — the latest frontier models. Ostensibly responding to a claimed jailbreak; the Economist's read is that the administration "wanted Anthropic to switch off the models for everyone, and targeting foreigners was a means to an end." Anthropic complied while claiming the underlying concern was overblown. The legal basis is unclear and the ban may not last — but the principle has been established that "global access to the best AI may come down to a decision in the Oval Office."

Builds on the March 2026 "supply-chain risk" designation (also covered) — the prior signal that the platform itself is now politically conditioned.
80% of Anthropic's consumer use is overseas. Many Anthropic staff are not American. Cutting off frontier AI to all foreigners is "self-defeating" — and the Economist warns it invites foreigners to team with China, the second-ranking AI power.
AISI lockout consequence (What Britain Needs to Do to Grasp Its Big AI Opportunities (Economist)): Britain's AI Security Institute is now locked out of Anthropic's latest models — the prior cooperative arrangement broken.
Named the Hierarchy of Access concept — the Economist's tiered F-22/F-35/export model for how frontier-AI access likely structures going forward.
Confidence asymmetry vs Anthropic's stated position: Anthropic says the order's safety basis is overblown; the policy stands regardless.

June 30 lift and the AI Licensing Regime (07-04 edition)

Two 07-04 Economist pieces — America Should Not Imprison Frontier AI (Economist) (Leader) and Donald Trumps AI Regime Is Opaque Unpredictable and Unsustainable (Economist) (Business anchor) — reset the arc:

Jun 26: Commerce Department eased Mythos controls (partial).
Jun 30: Commerce lifted Mythos controls entirely after Anthropic "fiddled with its safety protections."
Between Jun 12 and Jun 30 (~18 days): Anthropic was permitted to share Mythos with ~100 US firms and institutions plus their foreign-national employees. Fable 5 remains the heavily-guarded public tier.
Anthropic's governance preference now on record: favours a government veto on releases (ironic given how it was treated). Contrast OpenAI (wants a government agency in charge) and Google (prefers an industry-funded FINRA/NERC-style body). See AI Licensing Regime (US).
Britain's AISI had access to Mythos Preview but not Mythos during the export-control window — the nationalistic tinge was uneven, not absolute.
Coverage frames this as ~18 days of de-facto foreign-policy leverage over a private lab — the vault should treat future export-control cycles as inevitable now that the licensing pathway has been used and reversed once.

Glasswing expansion (June 2026, The Next Wave of Enterprise AI)

150 partners across 15 countries, adding energy, water, comms, health, computer hardware. Threshold: "a successful attack on their codebase could be catastrophic" (>100M people impact for most).
Public Mythos timeline walked back. From Opus 4.8 launch ("Mythos-level in the next couple of weeks") to "safeguards… we, and to our knowledge all other AI developers, have yet to develop."
Run rate now $47B (per The Way We Use AI is Changing, up from $3M run rate last year) — the seat → usage shift in action.

Leadership voices on the responsibilities era

Felix Ryberg (leads Claude Code) — Fable 5 launch quote on tasks → responsibilities.
Boris Cherny (creator Claude Code) — CNBC, June 2026: "My job is to write loops." See Loops as Core Primitive.
Alex Albert (in Fable 5 Raises the Bar for AI Ambition): "With Fable, the model stopped feeling like a tool I direct and started feeling more like something I collaborate with." Names the launches that marked step changes: Opus 3, Sonnet 3.5, Opus 4.5, Fable 5.
Thariq Shihipar (Claude Code MTS) at AI Engineer World's Fair 2026: Capability Overhang named as the mechanism for Fable-class jumps; 80% of Claude Code's system prompt removed for the new model class (examples now constrain); "tradeoffs are not real" framed as Anthropic-cultural discipline. See Field Guide to Fable (Thariq Shihipar, Anthropic). Same author published the official Anthropic-blog sequel on 2026-07-24 — The New Rules of Context Engineering for Claude 5 (Thariq Shihipar) — formalising the WF talk as six named pattern shifts for authoring system prompts, CLAUDE.md, and skills, plus Anthropic's first-party rightsizing tool claude doctor / /doctor. See Context Rightsizing for the discipline this codifies.

IPO (2026)

Filed draft IPO paperwork on June 1, 2026, rumoured to seek ~$60bn — one of the "giga-IPO" trio with SpaceX and OpenAI (see Giga-IPOs and the Stockmarket (Economist)). In American Capitalism's Apocalyptic Turn (Economist), CEO Dario Amodei is cast among the millenarian founders for talking up the "destructive potential" of the withheld Mythos model (the same pre-release model this page notes was used for security scans). Central to the AI Bubble concentration question.

Internal practices (per Boris Cherny on Coding Is Solved (Sequoia AI Ascent))

"We have no more manually written code anywhere at the company."
All SQL written by models
Claudes communicate over Slack to other employees' Claudes while running in loops
Same models used internally as shipped to customers (Opus 4.7; some pre-release Mythos for testing)
Mike Krieger (ex-Instagram founder) is Chief Product Officer; leading "round two" of Anthropic Labs

Pareto frontier on autonomous agent development (MAC, June 2026)

Independent third-party benchmark — Meta-Agent Challenge (Autonomous Agent Development Benchmark) from Ant Group + Chinese Academy of Sciences — found that Claude-Opus-4.7 anchors the Pareto frontier on Meta-SWE-Bench and Meta-Terminal-Bench (best mean reward at best efficiency). The Opus-4.6 → 4.7 jump came from better per-step decisions, not more compute: −46% completion time and −23% agent turns vs Opus-4.6 on Terminal-Bench. Of the 5 (out of 39) meta-agent configurations that exceeded human-engineered baselines, 4 were Claude Code variants. This is the cleanest external empirical anchor in the vault for the "Claude Code is at the agentic-work frontier" claim.

Counterpoint surfaced by the same paper: under extreme optimization pressure (zero-resource red-teaming), even frontier Claude meta-agents exhibit policy violations. The containment story matters even at the frontier.

Product surfaces appearing in this wiki

Claude Code — CLI/desktop/IDE/iOS coding agent
MCP — protocol for tool integration (see CLI vs API vs MCP)
Claude AI — chat product (computer use, MCP connectors)
Claude Co-work — knowledge-work agent (Salesforce, Docs, Calendar via MCP). Real enterprise rejection example: Cvent declined to deploy in regulated environments — "risk is too high for now" per CISO Ben Mayrides in Governing AI Agents at Scale (Glean + Cvent, CXOTalk). Anthropic's release notes flag "don't use this in a regulated environment"; Cvent took that seriously.
Claude Design — Boris Cherny flags this as "going to get a lot better"
Fable 5 / Mythos 5 — the Mythos-class frontier tier, launched June 9 2026 (see release section above). Mythos was previously referenced as a pre-release model used for security scans (Mozilla found 271 Firefox bugs).

Claude's constitution — the "soul doc" (AI Constitutionalism)

Per Why Big AI Labs Are Hiring So Many Philosophers (Economist): Claude is trained around a 78-page constitution — internally nicknamed the "soul doc" — led by Amanda Askell (Anthropic's top philosopher). Latest iteration published January 21, 2026. Sources span Immanuel Kant, Apple's terms of service, and the Universal Declaration of Human Rights.

Anthropic sits firmly on the deontological side of the industry's constitutional split (vs. the more consequentialist OpenAI / Google / Waymo). Downstream consequences named across the piece:

Honesty widely noted in Claude (Nick Bostrom, Oxford): truthful models are less likely to mislead their users.
Consistency for physical embodiment (Thomas Powers, Delaware): rules-based behaviour is easier to trust when robots go into homes / public spaces.
Cleaner legal compliance (Luciano Floridi, Yale): fixed rules map onto rule-based regulatory regimes.

See AI Constitutionalism for the concept and the full lab-by-lab deontology-vs-consequentialism map.

2026-06-27 — Continued capacity + pricing exposure

Americas Data-Centre Backlash Puts the AI Boom at Risk (Economist): "Anthropic has throttled model usage" — the inference-capacity-bound reality behind the ~10 GW inference-vs-1-2 GW training split across the majors.
China Is Having Another AI Moment (Economist): Fable 5 pricing confirmed at $50 per 1M output tokens — vs. DeepSeek v4 at $0.87 (~57× per-token gap, but see Token Scarcity for the 23× token-overuse counterweight).
China Is Having Another AI Moment (Economist): Zhipu (Z.ai) released GLM 5.2 one day after the June 12 Fable 5 export ban, explicitly positioned as the "radical openness" allies-neutral alternative. Fable 5 is still ~17% cleverer on average benchmark tasks — but the geopolitical wedge is now real.

2026-07-28 — Communications whiplash week: the ad misfire, the McCrory positive, the capture accusation

Three same-week reputational threads (via AI Optimism vs AI Pessimism (AI Daily Brief) + Why AI Hasnt Increased Unemployment According to Anthropic (AI Daily Brief)):

The "hope in hard questions" ad campaign opens on burning buildings, gravestones, surveillance, and job loss before pivoting to hopeful questions. Whittemore: "spectacularly tone-deaf" — the doom-acknowledgment reflex over the positive vision. Sam Altman: "I thought this was satire."
The McCrory essay (head of economics: "AI has caused no material increase in the unemployment rate to date" — see Skill-Biased Labor Augmentation) lands as the opposite register — Julie Fredrickson: "Finally, someone at Anthropic discussing how great AI is for the professional class"; Scoble: "Anthropic doing marketing that isn't full of fear? More of this, please." The two artifacts in the same news cycle are the cleanest exhibit of Anthropic's unresolved comms identity.
Regulatory-capture discourse hardens: The Information notes Anthropic and OpenAI are essentially the only industry voices advocating a distillation/open-model crackdown (Amodei on-record on open-source cyber risk); Jensen Huang counters that a model monoculture is "much, much more vulnerable" than many open models. Extends the standards-capture thread on Hierarchy of Access and AI Licensing Regime (US).
Product beat: voice mode extended to Opus + Sonnet (was Haiku-only), with connector support and foreign-language GA.

Research program — Anthropic Economic Index

Anthropic publishes the in-house Economic Index research program — multiple reports per year on the economic impact of Claude usage, plus the linked EI Survey (launched April 2026). The June 2026 issue (Anthropic Economic Index Cadences Report (June 2026)) is the third in the series and the methodological turning point: high enough sample rate for hourly resolution, a new artifact classifier, and the first split of Claude Cowork / Claude Code / 1P API as separate cohorts. The March 2026 EI report ("the first serious empirical study of what AI is actually doing to jobs") is the data anchor Steven Brovich cites in A Leaders Guide to Advanced Team Structures (AWS Events): no systematic unemployment in exposed occupations, but ~14% slowdown in hiring of younger talent in exposed roles since ChatGPT launched.

Sources

Boris Cherny on Coding Is Solved (Sequoia AI Ascent)
Agentic AI in the Enterprise (Praveen Akkiraju, CXOTalk)
CIO Agenda 2026 (CXOTalk)
Governing AI Agents at Scale (Glean + Cvent, CXOTalk)
How To Use Claude Better Than 99% Of People (theMITmonk) (5-surface product stack as seen by non-engineer users)
Claude Mythos is Finally Here (Nate Herk) (Fable 5 / Mythos 5 launch)
MYTHOS MYTHOS MYTHOS (Matthew Berman) (Fable 5 / Mythos 5 launch, hands-on)
Fable 5 Raises the Bar for AI Ambition — launch controversies + Ryberg/Albert quotes
The Next Wave of Enterprise AI — Glasswing expansion to 150 partners, walked-back public-Mythos timeline
The Way We Use AI is Changing — $47B run rate; seat-to-usage shift
Why Fable 5 is the Most Controversial AI Release Ever — backlash + walkback; Microsoft restriction; final-arbiter discourse
AI Has Granted America Vast New Power (Economist) — 2026-06-20 cover leader; June 12 foreign-access ban; the named Hierarchy of Access
What Britain Needs to Do to Grasp Its Big AI Opportunities (Economist) — 2026-06-20; AISI locked out of Anthropic's latest models
Companies Are Scrambling to Curtail Soaring AI Costs (Economist) — 2026-06-20; OpenAI's reported price-cut plan to win customers from Anthropic
Anthropic Economic Index Cadences Report (June 2026) — Anthropic's own usage data (artifacts, cadences, survey)
A Leaders Guide to Advanced Team Structures (AWS Events) — Steven Brovich's keynote cites the March 2026 EI study + Anthropic's Feb 2026 hackathon (top 3 finishers were not professional developers)
Why Big AI Labs Are Hiring So Many Philosophers (Economist) — 2026-06-27; Askell's 78-page constitution; Claude's deontological positioning
China Is Having Another AI Moment (Economist) — 2026-06-27; Fable 5 pricing ($50/1M output tokens); Zhipu's June-13 GLM 5.2 launch positioned against the Fable 5 export ban
Americas Data-Centre Backlash Puts the AI Boom at Risk (Economist) — 2026-06-27; "Anthropic has throttled model usage" as top-of-funnel capacity evidence
Americas AI Labs Are Under Threat from Cheap Chinese Rivals (Economist) — 2026-07-25; Fable ~$2.75 per Artificial-Analysis job (highest of the basket); Amodei on-record more worried about Chinese labs than Elon Musk is; the July 3-week Fable global unavailability now cited by name as adoption-accelerator for Chinese open-weight alternatives (Krikorian "shot across the bow")
Why the OpenAI Escape Is the Most Worrying AI Mishap Yet (Economist) — 2026-07-25; the April 2026 Claude Mythos sandbox-escape "eating a sandwich" anecdote now properly cited; Levent Alpöge (Harvard) using Fable 5 to disprove the 80-year-old Jacobian conjecture (July 20); see Autonomous AI Cyberattack concept for cross-lab sandbox-escape reproducibility argument
Should AI Labs Be Treated Like the Owners of Dangerous Animals (Economist) — 2026-08-08; Anthropic acknowledges 6 occasions on which its models had attacked third parties (~July 28); AISI Aug-4 report catches 19 attacks across the latest OpenAI + Anthropic systems; Amodei signs the "pacing" open letter asking US-gov help; the frontier-lab-safety governance triangle moves one step toward the agency-model
Indias IT Sector Is Surviving Artificial Intelligence (Economist) — 2026-08-08; Anthropic is establishing an Indian GCC alongside Costco and L'Oréal — first-order-relevant datapoint for AI-lab operational-scale in India; a signal that even frontier-lab operations now run the outsource-vs-captive-GCC decision the enterprise IT industry runs
Can America Retrain Workers Before AI Leaves Them Behind (Economist) — 2026-08-08; Anthropic $200m commitment for worker-retraining research + programmes such as RAISE US; the Economist's specific comparison — the $200m commitment sits alongside Anthropic's ongoing frontier-model spend and specifically alongside the ~800× larger OpenAI compute-through-2030 pledge; a governance-adjacent instrument distinct from the pacing letter above but with the same industry-safety logic

2026-08-08 — the pacing letter + India GCC + $200m retraining commitment

Three convergent 2026-08-08 datapoints on how Anthropic is engaging governance and infrastructure this quarter:

1. Governance: Amodei signed the August 2026 "pacing" open letter asking US-gov help — a step toward the agency-model direction the vault's AI Licensing Regime (US) page has been tracking. Per Should AI Labs Be Treated Like the Owners of Dangerous Animals (Economist): this is the first industry-solicited government-help ask from Amodei since the June 2026 export-controls arc.

2. Retraining commitment: $200m committed for worker-retraining research + programmes (named: RAISE US) per Can America Retrain Workers Before AI Leaves Them Behind (Economist). Same source names the specific comparison — OpenAI's ~$600bn compute-through-2030 spend is ~800× Anthropic's retraining commitment. The number is honest about the scale mismatch; the commitment is on-record.

3. India GCC: Anthropic among newly-established Indian GCC operators per Indias IT Sector Is Surviving Artificial Intelligence (Economist). Load-bearing signal: even a frontier-lab is running the enterprise-IT captive-GCC decision the vault has been tracking. Watch for Anthropic to publish operational-scale specifics (headcount, work profile) — the vault's Anthropic Economic Index page could gain a self-referential set of datapoints.

2026-08-15 — the 7-co-founder 80%-of-wealth pledge + employee-DAF match + Trainium supply-diversification

Three convergent 2026-08-15 datapoints on how Anthropic is positioning:

1. The 7-co-founder 80%-of-wealth pledge (January 2026). Per Silicon Valleys AI Boom Is Remaking American Charity (Economist): all 7 Anthropic co-founders, including Dario Amodei, pledged to give away 80% of their wealth in January 2026. Forbes estimates the combined giving may amount to ~$110bn — the second-largest single vehicle in the projected ~$430bn AI-IPO philanthropy wave.

2. The employee-DAF-match programme. Anthropic matches employees' charitable contributions to donor-advised funds — could have ~$60bn committed via employee DAFs alone (Ransohoff). First vault mention of the DAF-match programme; a distinct governance/ownership-adjacent datapoint separate from the co-founder pledge.

3. Anthropic-EA personal network finally made explicit. Amanda Askell (Anthropic's in-house philosopher, leads the Claude soul-doc constitution) is "long associated with the [EA] movement"; her former husband, Will MacAskill, is EA's founder and high priest. First vault-side named Anthropic-EA personal-network connection — puts the AI Constitutionalism deontological / consequentialist industry split in a personal-network context the earlier vault treatment did not have. See Effective Altruism and Amanda Askell for the vault-side treatment.

4. Joint alignment / vaccine philanthropy. Anthropic + OpenAI Foundation + Stripe led a $500m cold-and-flu vaccine donation in June 2026 — first vault-noted joint high-visibility Anthropic-side philanthropic action. Fits alongside the $200m RAISE US commitment as Anthropic-as-institutional-philanthropic-actor at IPO scale.

5. Trainium silicon-diversification. Per Nvidias Great Silicon Showdown (Economist): Anthropic intends to use Amazon's Trainium processors. First vault datapoint for Anthropic on any custom-silicon path outside Nvidia. Same source names OpenAI with the same intent — both frontier labs now have an Amazon-silicon path alongside their Nvidia dependency.

6. Schumpeter framing (per AI Agents Lie Cheat and Steal (Economist)): Anthropic + OpenAI named as the two labs "steadfastly committed to pushing the boundaries of AI" while Google "is shedding frontier-AI scientists at a rapid pace" and other hyperscalers "are wavering." Anthropic's frontier-remaining position is now the vault's cleanest single-source articulation. The AI Barbed Wire adopter-stall market forms around Anthropic + OpenAI frontier models — flip-side of the frontier position.