Claude Versions

The OpenAI exodus and the founding (2021)

Anthropic was incorporated in San Francisco in early 2021 by a group of senior researchers and engineers who had just left OpenAI over a stated disagreement about the direction of safety and commercialization. The cofounders included Dario Amodei (Research VP at OpenAI; CEO at Anthropic), Daniela Amodei (President), Tom Brown (lead author of the GPT-3 paper), Sam McCandlish, Jared Kaplan (the “Scaling Laws” paper), Jack Clark, Chris Olah, Tom Henighan, Andy Jones, Nick Joseph, and several others.

The Series A in mid-2021 raised roughly $124M with Jaan Tallinn and Dustin Moskovitz as lead investors and the Center for Emerging Risk Research participating. Anthropic spent its first eighteen months as a research-only operation; the first commercial Claude product did not ship until March 2023.

Constitutional AI and the research direction

Anthropic's December 2022 paper “Constitutional AI: Harmlessness from AI Feedback” (Bai et al.) introduced RLAIF — reinforcement learning from AI feedback — as an alternative to RLHF. A model critiques and revises its own outputs against a written constitution rather than relying on human comparison labels for every preference. The technique remains the technical foundation of every Claude release since, and is the most widely-cited piece of Anthropic research outside of the earlier “Scaling Laws” paper.

The funding story

After the 2021 Series A, Anthropic raised a Series B and C in 2022 with participation from Sam Bankman-Fried's FTX and Alameda Research (reportedly around $500M). The FTX collapse in November 2022 left the FTX bankruptcy estate as Anthropic's largest single shareholder; the estate later sold the stake in 2024.

In 2023 Google invested an initial $300M, later expanded to a multibillion-dollar commitment. Amazon committed an initial $4 billion in September 2023, increased the commitment through 2024 to a reported $8 billion total, and on April 20, 2026 expanded again with $5 billion now plus up to $20 billion more, alongside a $100B+ ten-year AWS-services commitment from Anthropic for up to 5 GW of training and inference capacity. The two-cloud financial structure (Google + Amazon) is a recurring background fact in coverage of the company.

Venture rounds since then have moved fast: Series E in March 2025 (~$3.5B at $61.5B post-money, Lightspeed lead), Series F in September 2025 (~$13B at $183B post-money, Iconiq / Fidelity / Lightspeed co-leading), Series G on February 12, 2026 (~$30B at $380B post-money, Coatue and GIC co-leading), and Series H on May 28, 2026 — $65 billion at a $965 billion post-money valuation, led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, with Capital Group, Coatue, D1 Capital Partners, GIC, Iconiq, and XN co-leading. The Series H folded in $15 billion of previously committed hyperscaler investment (including Amazon's $5 billion from the April agreement above) and added Micron, Samsung, and SK hynix as strategic infrastructure partners. Anthropic disclosed in the same announcement that run-rate revenue crossed $47 billion in May 2026, up from the ~$30 billion it reported in April.

Four days after the Series H, on June 1, 2026, Anthropic confidentially submitted a draft Form S-1 to the SEC for a proposed initial public offering. The registration statement is not public, the share count and price are not set, and the company frames the filing as preserving the option to go public once the SEC completes its review — not as a committed listing. On October 18, 2026 Anthropic committed 5% of its shares, subject to stockholder approval, to a new wholly-owned Anthropic Public Benefit Fund, with a further 2% addable at the board's discretion for every 50% rise in the company's value above a set baseline; counting the cofounders' pledge of 80% of their equity and the employee donation-match program, the company puts roughly 20% of its shares as dedicated to the public good.

The Claude 1 / Claude 2 era (2023)

Claude 1 launched March 14, 2023 in limited access via Anthropic's API, Slack, and Quora's Poe — the same week GPT-4 became publicly available. Claude 2 launched July 11, 2023 as the first broadly available Claude, with the Claude.ai consumer chat product as its delivery surface. The headline differentiator was a 100,000-token context window, an order of magnitude beyond the GPT-4 8k / 32k offerings of the time.

Claude 2.1 in November 2023 doubled that to 200k and added tool-use in beta. The pre-tier era was also the period during which Anthropic's role in the AI-safety conversation crystallized publicly — Dario Amodei's Senate testimony, the “responsible scaling policy,” and the early model-card publishing convention.

The tier system (March 2024)

Claude 3 in March 2024 introduced the Opus / Sonnet / Haiku naming — three models released together for the first time, sharing a generation number and differentiated by capability and price. The convention has held through every release since, and is the one piece of product naming Anthropic has kept stable while everything around it (mid-version numbers, snapshot dates, “new” suffixes) has shifted. Claude 3 Opus was also Anthropic's first model widely considered competitive with the contemporary GPT-4 line on hard benchmarks — the moment the “serious challenger” framing became standard.

The 3.5 / 3.7 mid-version era (2024 – early 2025)

Claude 3.5 Sonnet shipped Artifacts in June 2024 — a side-panel rendering surface for code, documents, and diagrams that became the template for similar features across the chat-product industry. Sonnet 3.5 (new) in October 2024 shipped Computer Use, the first frontier-model capability where the model could take screenshots of a virtual desktop, click, type, and navigate UIs the way a human does. The Model Context Protocol (MCP) was announced alongside this period in November 2024 and has since become the de-facto standard for tool integration across the industry.

Sonnet 3.7 in February 2025 introduced extended thinking — an exposed reasoning mode toggleable per request. Claude Code launched the same week as the CLI agent for developers. Together those two changes set the shape of Anthropic's 2025: a sustained push into “agent” products on top of frontier models, with a deliberate distance from competitor strategies that emphasized consumer chat features.

The Claude 4 line (May 2025 onward)

Claude Opus 4 and Sonnet 4 launched together on May 22, 2025 as the public debut of the Claude 4 line. Opus 4 was the first new Opus-tier model in fourteen months — the longest gap between flagships in the line's history. The mid-version cadence accelerated noticeably afterward: Opus 4.1 in August, the 4.5 generation across all three tiers in September / October / November 2025, Opus 4.6 in February 2026 alongside Sonnet 4.6, Opus 4.7 in April 2026, and Opus 4.8 on May 28, 2026 — the latter shipping the Dynamic Workflows orchestration feature and the effort-control dial alongside the model itself. The Sonnet tier then jumped a whole number with Claude Sonnet 5 on June 30, 2026 — the most agentic Sonnet yet, priced below Sonnet 4.6 but performing near Opus 4.8, and shipping a new tokenizer and real-time cyber safeguards. Its $2 / $10 launch rate was introductory and scheduled to rise to the old Sonnet price on September 1, 2026; Anthropic cancelled that increase on August 10, 2026 and kept $2 / $10 as the standard price. The Opus tier followed on July 24, 2026 with Claude Opus 5 — a step-change release Anthropic frames as near-Fable-5 intelligence at half the price, with thinking on by default and a new top max effort level; it became the default model on Claude Max the same day. Two months later, on September 22, 2026, Anthropic opened a new numbering with Claude Opus 5.5 (claude-opus-5-5) — “the first model in our new Claude 5.5 family,” with Sonnet 5.5 and Haiku 5.5 promised within weeks. It cut the Opus price to $4 / $20 and cache reads to $0.20 per million (about 40% cheaper to run than Opus 5 at default settings), took the top of every benchmark Anthropic published for it — ahead of Fable 5.1 on agentic coding, knowledge work, and computer use — and was the first Opus to ship with the cyber, biology, and anti-distillation safeguards built for the Mythos class, routing flagged cyber work to Opus 4.8 and flagged biology work to Opus 5. Claude Sonnet 5.5 (claude-sonnet-5-5) followed six days later, on September 28, 2026, at Sonnet 5's $2 / $10 price: Anthropic measures it 30%+ faster and up to 30% cheaper per task than Sonnet 5, and its 70.6% on Terminal-Bench 4.0 edged past Opus 5.5's 66.4%, though Anthropic still ranks Opus 5.5 ahead for complex, open-ended work. It is the first Sonnet with cyber safeguards, falling back to Sonnet 5 on higher-risk security requests. Claude Haiku 5.5 (claude-haiku-5-5) completed the family on October 7, 2026, bringing the fast tier to a 1M-token window and an effort dial at a price that scales with prompt length — $0.10 / $0.50 per million for prompts up to 100k tokens, about 75% cheaper to run than Haiku 4.5 by Anthropic's measure; the same day Sonnet 5.5's cache reads were halved to $0.10. Anthropic also relaunched its Cyber Verification Program on October 6 as three access tiers covering Opus 5.5, Sonnet 5.5, and Mythos 5.1, absorbing Project Glasswing, and on October 15 took Opus 5.5's fast mode out of research preview.

The Claude 4 era has also been when Claude Code went from a developer side-project to a cornerstone product, and when the “agentic” framing of the platform — Opus 4.x for long-horizon coding, Sonnet 4.x for routed everyday work, Haiku 4.5 for high-volume cheap calls — became the canonical recommendation.

On June 9, 2026, Anthropic pushed past the Opus tier entirely with Claude Fable 5 (claude-fable-5), the first generally-available model in the new Mythos class — a tier it positions above Opus in capability, and the model the Opus 4.8 release notes above described as the “bridge” to. Fable 5 is the public, safety-gated counterpart to Claude Mythos 5 (claude-mythos-5), an invitation-only variant with lighter safeguards released to a small group of cyberdefenders and infrastructure providers under Project Glasswing; on some sensitive cybersecurity topics Fable 5 routes the answer to Opus 4.8 instead, and since the Opus 5 launch blocked biology questions route to Opus 5 — a classifier Anthropic loosened on August 7, 2026, cutting biology-related fallbacks by roughly 85%. It is priced at $10 / $50 per million input / output tokens. Three days after launch, on June 12, 2026, a US government export-control directive forced Anthropic to suspend both Fable 5 and Mythos 5 for all customers; the controls were lifted June 30, 2026 and Fable 5 was redeployed globally on July 1, alongside an improved cyber safety classifier and a proposed industry framework for scoring model “jailbreak” severity developed with Amazon, Microsoft, and Google. Two months later the safety story moved from the products to the evaluation harness: on July 30, 2026 Anthropic disclosed that three cyber-evaluation runs had escaped a third-party sandbox onto the live internet, the UK AI Security Institute reported a similar unauthorized-action incident involving Mythos 5 on August 4, and on August 31 Anthropic published the containment, monitoring, and alignment changes it made in response. All of it involved safeguard-reduced evaluation builds rather than the generally-available Fable 5, which kept shipping unchanged throughout. The class took its second entry on September 1, 2026 with Claude Fable 5.1 (claude-fable-5-1) and its trusted-access sibling claude-mythos-5-1: the same $10 / $50 headline price, cache reads cut 75% to $0.25 per million (about 25% off a typical workload), retuned cyber safeguards that intervene roughly 60% less often, the first generally-available Claude cleared to search for software vulnerabilities, and an Enterprise Frontier Safeguards system that leaves enterprise data in the customer's own cloud rather than Anthropic's.

The lawsuits

Several cases have shaped the legal context around Claude.

• Bartz v. Anthropic (N.D. Cal., filed 2024). Class action by authors Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson over training-data sourcing including pirated copies from sites like LibGen. In June 2025, Judge William Alsup ruled on summary judgment that training on lawfully-acquired books is fair use, but that the acquisition of pirated copies is not — setting the trial up around piracy liability for over seven million books Anthropic had downloaded. Anthropic settled in August 2025 for $1.5 billion, the largest copyright settlement in U.S. history. The four-installment payment schedule runs October 2025 / April 2026 / September 2026 / September 2027 at roughly $3,000 per eligible work; the claim deadline was March 30, 2026 (91.3% of 482,460 eligible works claimed). Judge Martínez-Olguín held the final fairness hearing May 14, 2026 and granted final approval on July 20, 2026, entering judgment and terminating the case the same day; the order took up the filed objections on their merits and cut class counsel's fee award from the requested $187.5 million to roughly $101.6 million. Distribution is gated on the 30-day appeal window running from that judgment, which closed August 19, 2026. One notice of appeal was filed inside that window, on August 18, 2026, by the publishers' coordination counsel — directed solely at the fee and service awards, which under the settlement agreement does not move the Effective Date for class payments. In a status report filed September 2, 2026, class counsel set the first distribution at approximately $2,203.56 per work, drawn from the roughly $1.083 billion held in escrow by then, and said payments would start no later than November 15, 2026 — but only for works where every rightsholder agrees on the split and has filed payment details. The administrator emailed each claimant a consolidated statement on September 4, 2026 showing what every co-claimant on a title claims as their percentage, with 30 days to resolve differences; contested splits go to a court-appointed Special Master and land in a later distribution. A second payment of roughly $930 per work follows once Anthropic's remaining $450 million installment lands, which the settlement gives it until September 2027 to pay — putting the full recovery a little over $3,100 per work. The fair-use ruling on training data is now the most-cited piece of LLM-copyright law.
• Concord Music Group, et al. v. Anthropic (N.D. Cal., transferred from M.D. Tenn., filed October 2023). Music publishers (Concord, UMG, ABKCO) sued over Claude's training on copyrighted song lyrics and over the model's lyric outputs. In March 2025, Judge Eumi K. Lee denied the motion for a preliminary injunction and granted Anthropic's motion to dismiss the contributory and vicarious infringement counts (with leave to amend), narrowing the case substantially. The publishers then dismissed those counts with prejudice on April 20, 2026 rather than replead them, leaving the direct-infringement claim on the training side as the case. Discovery has closed: cross-motions for summary judgment and Daubert motions are briefed, six amicus briefs were accepted in June 2026, and Judge Lee has set the summary-judgment and Daubert hearing for October 21, 2026.
• Concord Music Group v. Anthropic (II) (N.D. Cal., filed January 28, 2026). The same publisher coalition's second Anthropic suit, applying the Bartz piracy template to music. The complaint alleges mass torrenting of 20,000+ copyrighted songs from LibGen and PiLiMi, names CEO Dario Amodei and cofounder Benjamin Mann individually, and seeks roughly $3 billion. The first major LLM piracy-acquisition case filed after the Bartz settlement. Anthropic sought an indefinite stay pending Concord I; Judge Lee denied it on April 24, 2026, instead relating the two cases with a 90-day continuance. One count was dismissed in May 2026 and the publishers filed a second amended complaint; Amodei and Mann remain parties with separate counsel. The two individual defendants then split: on August 3, 2026 Anthropic filed a partial motion to dismiss and Amodei filed his own separate partial motion targeting the direct-infringement count against him, while Mann was ordered to answer rather than move. Both motions are set for hearing November 4, 2026 before Judge Lee — the first judicial test of whether the individual counts against an AI lab's founders state a claim.
• BMG Rights Management v. Anthropic (N.D. Cal., filed March 17, 2026). The third music-publisher suit, and the first from outside the Concord / UMG / ABKCO coalition. BMG's 47-page complaint names 493 compositions — among them “What a Wonderful World,” “Sympathy for the Devil,” and “Uptown Funk” — and pleads five counts: direct infringement through training and through Claude's outputs, direct infringement by torrenting from LibGen and the Pirate Library Mirror, contributory infringement, vicarious infringement, and removal of copyright-management information. BMG is represented by Manatt, Phelps & Phillips. Two complications sit around it: BMG licensed the AI music generator Suno on August 12, 2026 while continuing to litigate against Anthropic, and its merger with Concord — lead plaintiff in the two cases above — closed September 1, 2026, so the same company now runs both dockets.
• Round Hill Music, et al. v. Anthropic (N.D. Cal., filed August 17, 2026). The fourth music-publisher suit, and the one that says out loud it will not settle. Round Hill Music LP and five affiliated funds control roughly 14,000 compositions; the complaint attaches 500 of them as a “prioritized representative bellwether” — “Iris,” “Total Eclipse of the Heart,” “I Got You (I Feel Good)” — and says it will amend to cover “potentially ten thousand or more.” Counts are direct infringement plus two DMCA theories: circumvention of access controls and removal of copyright-management information. Statutory damages of up to $150,000 per willfully infringed work put the ask at “hundreds of millions of dollars and potentially approach or conceivably exceed $1 billion.” It also quotes Claude back at itself — asked to rewrite three of the songs “for a modern audience,” the model observed that its own output was “edging past ‘inspired by’ into reproducing the copyrighted song.” Round Hill hired Richard S. Busch, who won Blurred Lines, and demanded a jury trial; a parallel complaint the same day targets Suno.
• Sony Music Publishing, et al. v. Anthropic (N.D. Cal., filed August 28, 2026). Thirty-five publishing entities behind Sony Music Publishing and Warner Chappell — the EMI, Hipgnosis, Jobete, and Warner-Tamerlane catalogs among them — sued Anthropic, Dario Amodei, and Benjamin Mann over the same LibGen and PiLiMi torrenting that Bartz turned on, this time for the lyrics and sheet music inside the pirated books. Four counts: direct infringement by torrenting against all three defendants, contributory infringement by torrenting against Amodei and Mann personally, direct infringement against Anthropic for training and for Claude's outputs, and removal of copyright-management information under 17 U.S.C. § 1202. The complaint asks for statutory damages up to $150,000 per work infringed and $25,000 per CMI violation, a permanent injunction, and an accounting of Claude's training data. Same lead counsel as Bartz and the two Concord cases; with Universal already suing through those, all three major music publishers are now litigating against Anthropic at once. Anthropic said it disagrees with the claims and intends to defend itself in court. On September 9, 2026 Judge Lee related the case to Concord, consolidating supervision of all five music-publisher suits against Anthropic — Concord I and II, BMG, Round Hill, and Sony — before a single judge.
• Reddit v. Anthropic (San Francisco Superior Court, filed June 2025). Reddit alleged that Anthropic continued scraping Reddit user posts after access was revoked, with claims for breach of contract, unjust enrichment, trespass to chattels, tortious interference, and unfair competition. After Anthropic removed to N.D. Cal., Judge Trina L. Thompson remanded the case to state court on March 30, 2026, holding that none of Reddit's claims is preempted by the Copyright Act — a first-of-its-kind ruling making contract / TOS theories durable against AI-lab preemption defenses. The case is the highest-profile of the “platform-data-licensing” suits to date.
• Buist v. Anthropic PBC — the AI-pacing antitrust class action (N.D. Cal. 3:26-cv-10693, filed September 18, 2026). The only antitrust case on this list, and the only one in which Anthropic is a defendant alongside its competitors rather than on its own. Four named subscribers to ChatGPT, Claude, Grok, and Gemini sued Anthropic, OpenAI, SpaceXAI, and Google on behalf of a proposed nationwide class of paid subscribers, alleging the four labs struck an illegal agreement to slow the pace of AI capability development and so delivered less than subscribers paid for. The complaint pins the agreement to September 12, 2026 — the day Dario Amodei published “We Must Pace the Frontier” and Sam Altman, Elon Musk, and Demis Hassabis each publicly agreed — and traces it back to a July 2026 multi-lab statement about the “intense competitive pressure not to unilaterally slow.” Plaintiffs say they do not object to any lab slowing down on its own; their theory is that antitrust law forbids the shortcut of “collective restraint” substituting for individual accountability. They seek an injunction and treble damages. Amodei's essay had anticipated the exposure, arguing the US government should mediate cross-lab safety talks or issue “a narrow waiver for certain kinds of safety conversations”; Anthropic had not responded publicly as of this update.
• Anthropic PBC v. U.S. Department of War (N.D. Cal. 3:26-cv-01996, filed March 9, 2026, with a companion petition for review in the D.C. Circuit). The one case on this list where Anthropic is the plaintiff, and the only one that has moved Claude's availability rather than its training data. After Anthropic declined to let the department use Claude for fully autonomous weapons or mass domestic surveillance, the department cancelled a contract reported at around $200 million and designated Anthropic a supply-chain risk twice over — under 10 U.S.C. § 3252, the designation the district court reviewed, and in a parallel letter under the Federal Acquisition Supply Chain Security Act, which only the D.C. Circuit can review — a label that forces removal of Claude from covered systems and bars Anthropic as a prime or sub-contractor. Judge Rita F. Lin granted a preliminary injunction on March 26, 2026; the D.C. Circuit declined in April to lift the designation itself, so removal proceeded on a 180-day clock while the district case ran. On August 27, 2026 Lin ruled the designation unlawful outright — First Amendment retaliation, arbitrary and capricious, and a Fifth Amendment due-process failure — finding on the department's own records that it acted to make a public example of Anthropic for its “arrogance” in criticizing the government, and noting the incoherence of calling a vendor a national-security threat while separately pursuing its contract and collaborating with its Mythos models on cyberdefense. Final judgment was entered the same day and the case terminated. The department has not conceded the label in public: on September 3, 2026 Under Secretary of War Emil Michael — the author of the March memorandum that justified the designation — posted that Anthropic “is still a designated Supply Chain Risk” at the department and for the defense industrial base. The D.C. Circuit then went the other way on the parallel supply-chain-act designation: on September 25, 2026 a split panel denied Anthropic's petitions for review 2–1. Judge Gregory Katsas, joined by Judge Neomi Rao, held that the department had ample support for treating Claude's built-in use restrictions, which had blocked government users' requests more than once, as a covered national-security risk, and rejected the due-process and First Amendment claims. Judge Karen LeCraft Henderson dissented, reading the statute as aimed at sabotage by hostile actors rather than at an American vendor's safety guardrails. The panel held its decision back from taking effect so Anthropic can seek rehearing, and on October 13, 2026 Anthropic petitioned for panel rehearing or rehearing en banc, arguing that the majority read “otherwise manipulate” to turn on what a vendor does rather than why, upheld the designation on a rationale the department never gave, and ignored the August 27 district-court finding of retaliation. The mandate stays withheld while the petition is pending.
• Ortiz Colon v. Anthropic — the biometric-privacy class action (San Francisco County Superior Court, filed September 30, 2026). Jose Enrique Ortiz Colon, a Chicago resident, sued under the Illinois Biometric Information Privacy Act over the identity-verification step Anthropic requires of some Claude users, in which a government ID and a live face image are matched by its vendor Persona. The two-count complaint says Anthropic never disclosed in writing, before the scan, how long the face geometry would be kept, obtained no written release, and published no retention or destruction schedule; it seeks the statute's $1,000 or $5,000 per violation for a proposed class of Illinois residents who went through the check in the prior five years, plus an order that Anthropic publish its biometric-retention policy. Anthropic said it takes user privacy seriously and is reviewing the complaint. The first privacy suit on this list, and the first aimed at a Claude product feature rather than at training data.
• The Canadian news publishers v. Anthropic (Ontario, filed October 6, 2026). Postmedia, Torstar, The Globe and Mail, and The Canadian Press — the same coalition that sued OpenAI in Ontario in 2024 — filed a statement of claim alleging that Anthropic scraped their websites and trained Claude on their articles without a licence, breached the sites' terms of service, circumvented technological protection measures, and was unjustly enriched. The first copyright case against Anthropic outside the United States and the first brought by news publishers; the allegations are unproven, Anthropic had not commented as of the first reports on October 19, 2026, and the relief sought was not public.

The OpenAI superalignment exodus (2024)

In May 2024, several senior members of OpenAI's superalignment team left and the team was dissolved within days. Jan Leike, who had co-led that work with Ilya Sutskever, joined Anthropic publicly the same week, citing concerns about OpenAI's safety culture. John Schulman, an OpenAI cofounder and the original lead on RLHF and on ChatGPT itself, joined Anthropic in August 2024 (and later moved on to other ventures). The pattern echoed the original 2021 founding exodus: several of the most-influential alignment-focused OpenAI researchers ending up at Anthropic.

People who shaped Claude

Founders: Dario Amodei (CEO; Research VP at OpenAI before), Daniela Amodei (President, and chair of the board), Tom Brown (Chief Compute Officer; lead author of the GPT-3 paper), Sam McCandlish (Chief Architect), Jared Kaplan (Chief Science Officer; “Scaling Laws” paper), Jack Clark (Head of Public Benefit, running the company's Anthropic Institute think tank), Chris Olah (interpretability research lead), Tom Henighan, Andy Jones, Nick Joseph.

Post-founding leadership: Mike Krieger (cofounder of Instagram) joined as Chief Product Officer in 2024 and moved in January 2026 to co-lead Anthropic's Labs incubator alongside Ben Mann; Ami Vora succeeded him as Chief Product Officer. Jan Leike joined from OpenAI's superalignment team in 2024. Mariano-Florentino (Tino) Cuéllar — until recently President of the Carnegie Endowment for International Peace, and before that a Justice of the Supreme Court of California — joined on August 4, 2026 as Anthropic's first Chief Global Affairs Officer, leading policy, international engagement, and government relations; he stepped down from Anthropic's Long-Term Benefit Trust to take the role. The interpretability team that Chris Olah leads has produced much of the published research that defines Anthropic's public technical voice (mechanistic interpretability, sparse autoencoders, the “Towards Monosemanticity” line of work).

Public face: Dario Amodei has been the company's primary public voice through Senate testimony, the “Machines of Loving Grace” essay (October 2024), and the major podcast appearances. On September 12, 2026 he published “We Must Pace the Frontier”, arguing that frontier labs should deliberately slow capability gains and laying out a three-step plan — embedded third-party evaluators, then verifiable pacing commitments across US labs, then global coordination. Anthropic committed unilaterally to the first step and named its first partner six days later: Accenture, through its Faculty AI business, will red-team and run alignment assessments from inside the company, with both sides expecting to spend at least $1 billion each over five years. On October 15, 2026 the embedding started in earnest: six employees of the nonprofit evaluator METR now hold “employee-like access” — company hardware, office space, and internal systems on par with Anthropic's own risk-assessment teams — and will publish their findings, which Anthropic can redact but not veto; SecureBio is reviewing the biological-weapons chapter of the August risk report, the Accenture / Faculty partnership is still being built out, and talks with others, including Hugging Face, are at an early stage. The essay drew immediate public agreement from OpenAI, SpaceXAI, and Google DeepMind — and, within a week, the antitrust class action described above. Jack Clark writes the long-running “Import AI” newsletter and was the company's policy voice for years; his current brief is the Anthropic Institute, the in-house think tank on AI's societal effects, while policy and government relations now report to Cuéllar.

The competitive landscape

Claude is one of three frontier-model families that meaningfully compete in 2026, alongside OpenAI's GPT line (now a single numbered series — the separately-branded o-series reasoning models are no longer shipped) and Google's Gemini. Meta's Llama was the canonical open-weights frontier line until Meta moved its frontier work to the closed-weights Muse Spark line under Meta Superintelligence Labs in April 2026, leaving the open-weights frontier mostly to the Chinese labs (DeepSeek, Qwen); Grok, now developed inside SpaceX after the February 2026 xAI merger, is the fourth notable closed-weights contender. The positioning between Claude and its closed-weights peers has shifted over time — Claude is widely associated with coding and agentic workflows in 2026, GPT with the broadest consumer surface, Gemini with multimodal and long-context strength — but the rankings move release-to-release and the page does not attempt to capture them.