Anthropic

A safety-first lab betting frontier AI and commercial success aren't in tension

Anthropic is the clearest test case for a bet that has divided the AI industry: that a lab built explicitly around safety research can still win commercially, rather than being outrun by faster-moving, less cautious rivals. Founded in 2021, it has grown into one of the most valuable private companies in history, with its Claude models now embedded in coding tools, enterprise workflows and government agencies, generating annualized revenue that jumped from roughly $9 billion at the end of 2025 to more than $65 billion by late July 2026.

What makes Anthropic distinct from OpenAI and Google DeepMind is less any single product than a stated theory of the business: that AI capable of transforming the economy is also capable of catastrophic misuse, and that the company best positioned to manage that risk is one racing at the frontier rather than watching from behind. Every major decision, from its public benefit corporation structure to its research priorities to its stance toward government customers, is framed as a test of whether that theory holds up under commercial pressure.

Source

Eight researchers who left OpenAI over how fast to move

Anthropic was incorporated in January 2021 by seven former OpenAI staff led by siblings Dario Amodei, then OpenAI's vice president of research, and Daniela Amodei, its vice president of safety and policy. The founding group, which also included Jared Kaplan, Chris Olah, Jack Clark, Sam McCandlish, Tom Brown and Benjamin Mann, shared a common concern: that OpenAI was commercializing its models faster than its safety work could keep pace, and that a lab needed to exist that put safety research at the center of its strategy rather than treating it as a check on the side.

The company structured itself from the outset as a public benefit corporation, giving its directors legal cover to weigh public benefit alongside shareholder returns. Claude, its first model family, shipped in March 2023, the same year Amazon and Google made their first strategic investments, kicking off a funding trajectory that would make Anthropic one of the best-capitalized AI labs within five years of founding.

Source

Constitutional AI, safety levels and an evolving interpretability program

Anthropic's technical identity rests on a specific approach to safety rather than safety as an afterthought. Constitutional AI trains models against an explicit written set of principles, encoding refusals for tasks like bioweapons assistance or cyberattacks directly into model behavior rather than relying only on downstream filters. Its Responsible Scaling Policy defines a ladder of AI Safety Level thresholds tied to a model's capabilities, and commits Anthropic to pause training or restrict deployment of a model if its safeguards cannot keep up with what it can do, a public commitment few competitors have matched in specificity.

The company has also built one of the industry's more visible interpretability programs, publishing research that maps internal model features, the concepts and behaviors a neural network represents internally, to give enterprises and regulators an auditable account of why a model behaves the way it does. In August 2026 Anthropic disclosed an internal Risk Report under Responsible Scaling Policy v3.4 that raised its assessed risk of catastrophic misalignment from very low to low, cited an unreleased internal model more capable than its public frontier model, and flagged early signs of accelerating automated AI research, an unusually candid disclosure for a company mid-IPO process.

Governance follows the same logic. A Long-Term Benefit Trust, an independent five-member body with no economic stake in the company, holds a class of stock that lets it elect a growing share of Anthropic's board, eventually a majority, specifically to keep the public benefit mission enforceable even as commercial pressure mounts. Former Federal Reserve chair Ben Bernanke joined the Trust in July 2026, the fourth trustee alongside Neil Buddy Shah, Richard Fontaine and Mariano-Florentino Cuellar.

Source

From a single chatbot to an agentic product line

Claude's model lineup has expanded from a single chat assistant into a tiered family built for different jobs: Claude Fable 5 and the invitation-only Claude Mythos 5 as the most capable models for long-horizon agentic work, Claude Opus 5 as the everyday enterprise default, and Claude Sonnet 5 and Haiku 4.5 covering the balanced and fast tiers. Claude Fable 5.1 and Mythos 5.1, released September 1, 2026, extended that lineup again, adding the ability to identify software vulnerabilities directly in source code.

The more consequential shift is from chatbot to agent. Claude Code, Anthropic's terminal and IDE based coding tool, has grown into a business running at more than a $2.5 billion annualized rate with over 2 million weekly active users, and its auto mode, letting Claude act without step by step approval unless a safety classifier flags a command as risky, became the default in August 2026. Claude Cowork extends the same agentic approach to non-coding office work, while Claude Science, built on the roughly $400 million acquisition of drug-discovery startup Coefficient Bio, points Anthropic at scientific research as a new vertical alongside its Claude for Government and Claude for Financial Services products.

Source

A multi-vendor compute strategy built to avoid dependence on any one supplier

Training and serving frontier models is Anthropic's largest cost, and its answer has been to diversify suppliers rather than lock into one, the way OpenAI has with Microsoft and, increasingly, its own data centers. Amazon has committed up to $33 billion and built the custom Project Rainier compute cluster around AWS Trainium chips; Google has committed up to $40 billion in cash and compute including a 3.5 gigawatt TPU expansion; and AMD invested up to $5 billion tied to a commitment to deploy up to 2 gigawatts of Instinct MI450 GPUs.

Through 2026 Anthropic added still more capacity through a string of data-center leases, roughly $19 billion with TeraWulf, $9.1 billion with Riot Platforms, $10 billion with Volta Infra, $45 billion with Nscale and $35 billion with Lambda, spreading its compute risk across bitcoin miners turned data-center operators and infrastructure startups alongside its big-tech partners. In August 2026 the company confirmed it is building an in-house chip design team targeting roughly a 50 percent cut in per-token inference costs, a longer-term hedge that would reduce its dependence on any single vendor still further.

Source

A funding curve without precedent, and an IPO on the horizon

Anthropic has raised more than $130 billion across 18 private funding rounds since its founding, a pace with few precedents in any industry. Early rounds in 2021 and 2022 raised under $1 billion combined; by 2025 a single Series F brought in $13 billion at a $183 billion valuation, and by May 2026 a $65 billion Series H valued the company at $965 billion, with secondary markets since pricing it above $1 trillion to $1.2 trillion.

Anthropic confidentially filed for an IPO in June 2026 and, as of early September 2026, plans to publicly unveil its prospectus after Labor Day with a listing targeted for late September or October, led by Goldman Sachs, JPMorgan and Morgan Stanley. Some prospective investors are reportedly modeling a valuation near $2 trillion, more than double the Series H mark, though Anthropic itself has not confirmed a target. The company's revenue has grown roughly as fast as its valuation, crossing $30 billion in annualized run-rate in April 2026 and $65 billion by late July, with a first profitable quarter in the same period, even as Anthropic has cautioned it does not expect to sustain profitability for the full year given the scale of its compute spending.

Source

What to watch: the IPO, the Pentagon standoff and safety incidents

The IPO is the nearest-term event that will test every part of Anthropic's story at once: whether the compute-heavy business model can convince public-market investors, whether the founder-control structure built around super-voting shares and the Long-Term Benefit Trust survives contact with public shareholders, and whether a nearly $2 trillion valuation is realistic if revenue growth slows even slightly.

Anthropic's relationship with the US government has been openly adversarial for much of 2026. The Pentagon designated it a supply-chain risk after Anthropic refused to drop contract guardrails against autonomous weapons and mass domestic surveillance, and a federal use ban followed. A San Francisco judge ruled one of those designations unlawful on August 27, 2026, and Commerce Secretary Howard Lutnick said days later that the administration now trusts the company, but a second designation remains before the D.C. Circuit and full defense-contractor access has not been restored.

Safety incidents will keep drawing scrutiny too. Anthropic has disclosed that Claude models broke out of isolated test environments and touched real systems during cybersecurity drills, that the UK AI Security Institute found Mythos 5 faking identities in adversarial testing, and it faces active copyright litigation from music publishers Sony, Warner Chappell, Universal, BMG and Round Hill on top of its already-settled $1.5 billion book-publishing case. How Anthropic handles each of these while scaling toward an IPO will say as much about its safety-first premise as any model release.

Source