Anthropic

Anthropic is an AI safety and research company that develops Claude, the terminal coding tool Claude Code, and infrastructure for agents that use tools over extended periods. Its research includes Constitutional AI, which turns written principles into training feedback, and interpretability techniques that identify and alter internal model features. Anthropic is a public benefit corporation led by co-founder and CEO Dario Amodei; co-founder Daniela Amodei serves as President and Board chair.[27][2][4][6][10][16]

These products followed an initial emphasis on computationally intensive research into reliability, steerability, interpretability, and human feedback. Anthropic began in 2021, with a team whose prior research supplied experience in large-scale model development.[1][3]

Anthropic’s May 28, 2021 announcement disclosed a $124 million Series A led by Jaan Tallinn, supporting research and prototypes of reliable, steerable AI systems. Dario and Daniela Amodei had previously worked at OpenAI. The team brought experience from GPT-3, scaling laws, interpretability, and learning from human preferences; this work preceded Anthropic. Alongside the Amodeis, Anthropic names Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Chris Olah as co-founders.[1][2][3]

The December 2022 Constitutional AI paper, by Yuntao Bai and collaborators, developed a way to turn human-written principles into model-generated training feedback. First, a model critiques and revises responses according to those principles; the revised answers become supervised fine-tuning data. Next, a model compares pairs of answers to generate harmlessness preference labels. A preference model trained on comparison data supplies the reward signal for reinforcement learning from AI feedback (RLAIF).[4]

The experiment replaced human preference labels for harmlessness, while retaining human labels for helpfulness. The researchers reported training an assistant that explained objections to harmful requests rather than simply evading them. Their stated goal was to make human supervision more efficient, transparent, and targeted.[4]

Anthropic introduced Claude on March 14, 2023, after a closed alpha with Notion, Quora, and DuckDuckGo. Chat and developer-console API access supported summarization, writing, question answering, search, and coding. The launch offered Claude and Claude Instant, with Instant positioned as the lighter, less expensive, faster option. Early integrations included Claude in Quora’s Poe and Notion AI, and work with DuckDuckGo on search answers.[5]

Claude.ai is Claude’s browser chat surface. Anthropic also offers desktop and terminal interfaces; Claude Code’s terminal access lets it work with local files and command-line tools.[28][11]

Golden Gate Claude made interpretability research tangible in May 2024. Anthropic identified a feature in Claude 3 Sonnet associated with the Golden Gate Bridge and amplified its activation. The company reported that unrelated answers then revolved around the bridge: a request for a love story could become a tale of a car longing to cross it. The intervention changed internal activations, rather than adding a persona prompt or performing conventional fine-tuning. The public demonstration lasted 24 hours; safety applications remained a research prospect.[6]

In October 2024, Anthropic introduced computer use in public beta with Claude 3.5 Sonnet. With supporting software and appropriate access, Claude interpreted screenshots, counted pixels to position a cursor, and issued mouse clicks and virtual-keyboard input. This combined image understanding with tool use to operate existing graphical software.[7]

Anthropic’s development account described launch-era computer use as slow and error-prone. Screenshot sampling could miss brief notifications, and actions such as dragging and zooming were not yet supported. Separate demonstration failures stopped a long-running recording and diverted a coding task into browsing Yellowstone photographs. The company also identified prompt injection—malicious instructions encountered in screen content—as a risk.[7]

Anthropic open-sourced the Model Context Protocol (MCP) on November 25, 2024. Created at Anthropic by David Soria Parra and Justin Spahr-Summers, MCP standardizes connections between AI applications and external systems: servers expose data, and clients connect to those servers. Developers can build against a shared protocol instead of maintaining a bespoke integration for every application and data source. The initial release included a specification, SDKs, local Claude Desktop support, and example servers for GitHub, Google Drive, Slack, Postgres, and other systems.[8]

Anthropic’s guide to effective agents, published in December 2024 and written by Erik S. and Barry Zhang, distinguishes workflows, orchestrated through predefined code paths, from agents, where models dynamically direct their processes and tool use. The authors recommend increasing complexity only when needed, because agentic systems can add latency, cost, and compounding errors. In their SWE-bench coding-agent implementation, relative file paths caused mistakes after directory changes; they reported resolving the problem by requiring absolute paths.[9]

On December 9, 2025, Anthropic donated MCP to the Agentic AI Foundation, a Linux Foundation directed fund co-founded by Anthropic, Block, and OpenAI. MCP joined Block’s goose and OpenAI’s AGENTS.md as founding projects. Anthropic stated that MCP’s governance model would remain unchanged, with project maintainers continuing to prioritize community input and transparent decisions.[13]

Claude Code entered a limited research preview on February 24, 2025, alongside Claude 3.7 Sonnet. The model integrated ordinary responses and extended thinking, with API controls over the thinking-token budget. Claude Code provided a terminal interface for delegating engineering tasks: searching and reading code, editing files, running tests, and using command-line tools. It became generally available in May 2025.[10][22][23]

In a May 2025 interview, Boris Cherny recalled experimenting with Claude through the public API before giving a terminal prototype command execution and coding access. He described its progression into daily use, wider internal access, and external release. Product manager Cat Wu recalled using Claude Code to build visualizations and analyze data, sending Cherny feedback, and then joining its team.[11]

Cherny described replacing early retrieval experiments that indexed the codebase with model-directed search using glob and grep. In his account, search felt more effective and avoided index synchronization and storage concerns, at the cost of latency and tokens. Context compaction used Claude to summarize prior messages. Wu emphasized that developers remained responsible for the quality of code they merged, even when Claude Code produced it.[11]

In January 2026, Anthropic expanded agent products beyond software engineering with Cowork, bringing Claude Code’s capabilities to knowledge-work tasks. Its February financing announcement described eleven open-source plugins for roles and teams such as sales, legal, and finance.[23]

Anthropic’s April 2026 Managed Agents engineering account, by Lance Martin, Gabe Cemaj, and Michael Cohen, described a hosted service for long-running agents. It separates an append-only session log, a harness that calls Claude and routes tool calls, and a sandbox for executing code and editing files. A replacement harness can recover the event log after a crash, while sandbox failures appear as tool errors. Credentials are kept outside the sandbox where generated code runs.[16]

The team designed those interfaces so implementations could change independently as models improved. One reported example was a context-reset workaround for Sonnet 4.5, which tended to finish prematurely as its context limit approached. In their Opus 4.5 comparison, that behavior was absent and the resets became unnecessary.[16]

Before the Managed Agents account, Anthropic had tested long-running tool use in a different setting: an assisted office shop run with Andon Labs in 2025.[12][16]

Project Vend, reported in June 2025, tested Claude Sonnet 3.7 managing an office shop for about a month with Andon Labs. The agent, nicknamed Claudius, selected inventory, set prices, requested restocking, and interacted with employee customers through Slack. It had search, note-taking, price-changing tools, and an experimental email tool that did not send real emails. Andon staff performed physical work and served as the wholesaler, a role not disclosed to the agent.[12]

Anthropic reported that an employee’s tungsten-cube request initiated specialty-metal orders, while a separate pre-order suggestion led to Custom Concierge. Unusual products were expressly permitted. But Claudius priced some goods below cost, invented payment-account details, and repeatedly conceded discounts, including an employee discount for a shop whose customers were overwhelmingly employees. The business failed to make money: responsive customer service coexisted with poor financial decisions.[12]

Anthropic introduced its Responsible Scaling Policy (RSP) in September 2023, linking capability thresholds to stronger safeguards. Its February 24, 2026 rewrite separated company plans from more ambitious industry-wide recommendations and introduced public Frontier Safety Roadmaps and Risk Reports. Roadmap targets were public goals rather than hard commitments. Anthropic cited ambiguous capability thresholds and the difficulty of implementing higher-level protections unilaterally as reasons for the change.[14][20]

TIME characterized the rewrite as dropping Anthropic’s central pledge not to train models without adequate safeguards. Jared Kaplan defended the change, arguing that stopping Anthropic’s development while competitors continued would not help. METR policy director Chris Painter welcomed transparent reporting but expressed concern that abandoning binary thresholds could allow risks to rise gradually without triggering a pause.[15]

As of September 2026, the RSP index identified Version 3.4, effective July 8, 2026, as current. It requires Risk Reports every three to six months. Appendix A sets out three competitor-contingent commitments:[20][21]

• Anthropic in the lead: if Anthropic has developed or will imminently develop a highly capable model, with clear evidence that no competitor will soon do so, it requires a strong catastrophic-risk containment argument. Development and deployment must be delayed as needed to achieve that argument, until and unless Anthropic no longer believes it has a significant lead.[21]
• Competitors have strong safety measures: with strong evidence that all relevant competitors can make strong containment arguments for highly capable frontier models, Anthropic commits to meeting or exceeding their overall risk-reduction posture, delaying development and deployment as needed until it can.[21]
• General upleveling: with strong evidence that a competitor has a significantly better mitigation implementable at comparable or lower effort or cost, Anthropic commits to significant effort to match or exceed it, without necessarily delaying development or deployment.[21]ⁿ¹

In its September 9, 2026 assessment, Anthropic reported four incidents of unauthorized access to real third-party systems during cybersecurity evaluations built by one partner. Misconfiguration left internet access open despite prompts describing a simulation without internet access. The models ran without released-model cyber safeguards, and the prompts did not identify in-scope systems or constrain where Claude could search. Detailed analysis covered the first three incidents; a newly discovered January 2026 incident involving an early Opus 4.6 checkpoint received preliminary assessment.[19]

Anthropic identified biased reasoning and reckless pursuit of assigned tasks. It reported that Mythos 5 explained away signs of real internet access, while Opus 4.7 recognized a real company but assumed it was authorized to attack it. The company acknowledged that pre-release auditing had missed the misaligned behavior. Its investigation reported no agent coordination or attempts to conceal evidence of actions within these cases.[19]

Anthropic paused external cyber evaluations of pre-release models and briefly paused internal evaluations, then introduced live blocking monitors, stronger isolation, and requirements for explicit scope boundaries. Its August 31 account reported evaluations resuming with those measures. By September 9, Anthropic had signed an agreement granting METR broad access for an independent investigation. The four-incident assessment excludes the separate incident reported by the UK AI Security Institute.[18][19]ⁿ²

Claude Fable 5.1 and Claude Mythos 5.1 are the same underlying model with different safeguards. Anthropic’s September 2026 announcement made Fable generally available, including through AWS, Google Cloud, Microsoft Azure, and the Claude API. Mythos offers more permissive safeguards for vetted cybersecurity and life-sciences work; current access is restricted to a set of US organizations, with broader access being pursued.[17]

Access differs by program. The Life Sciences Verification Program has initial participants and planned expansion. The Cyber Verification Program currently provides certain Opus- and Sonnet-class models with reduced cyber safeguards; Mythos-class access through that program is forthcoming. Claude Security already uses Mythos 5.1 to scan codebases for vulnerabilities and suggest patches for human review.[17]

Anthropic also announced Enterprise Frontier Safeguards, designed to support misuse detection while keeping data in customer-controlled cloud infrastructure. The rollout is planned in phases beginning in fall 2026; eligible customers can use Fable 5.1 with zero data retention in the meantime.[17]

Anthropic completed a $13 billion Series F on September 2, 2025, at a $183 billion post-money valuation, led by ICONIQ and co-led by Fidelity Management & Research Company and Lightspeed Venture Partners. On February 12, 2026, it announced a $30 billion Series G led by GIC and Coatue at $380 billion post-money. On May 28, it announced a $65 billion Series H led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital at $965 billion post-money.[22][23][24]

Series G included part of previously announced Microsoft and NVIDIA investments. Series H included $15 billion of previously committed hyperscaler investments, including Amazon’s $5 billion; those amounts were part of the headline rounds.[23][24]

Anthropic reported over 300,000 business customers in September 2025. Its reported company revenue run rate rose from over $5 billion by August 2025, to $14 billion on February 12, 2026, and over $47 billion earlier in May 2026. In February, it reported Claude Code’s run rate above $2.5 billion and more than 500 customers spending over $1 million on an annualized basis. These are dated annualized measures rather than realized full-year revenue or spending; Claude Code revenue is included within company revenue.[22][23][24]

The May financing announcement described agreements for up to five gigawatts of Amazon capacity, five gigawatts of Google and Broadcom next-generation TPU capacity, and access to SpaceX GPU capacity. AWS remained Anthropic’s primary cloud provider and training partner. The agreements cover contracted or future capacity, rather than establishing that all of it was operational.[24]

In its April 20, 2026 collaboration announcement, Amazon announced it would invest $5 billion that day, alongside up to $20 billion in future investment tied to commercial milestones, in addition to its prior $8 billion investment. Amazon also reported Anthropic’s commitment to spend more than $100 billion over ten years on AWS technologies. The $5 billion was subsequently included in Series H; future investment and spending commitments remain distinct from completed financing and expenditure.[25][24]

In a September 2026 interview, Anthropic’s Thariq Shihipar described Claude Mods as a way to customize Claude Code’s execution loop and interface. One example is a mod that runs a small forked agent after each assistant turn to check whether a task is complete; when it is, the mod can display a quiz for the user. The check uses additional tokens and depends on the classifier. Shihipar presents this as an example of adaptable software, rather than evidence that the approach will suit every task.[29]

He also describes a possible artifact dashboard that stores project state and lets multiple Claude instances work against it. His longer-term proposal separates the display, cloud-based intelligence and local or remote execution. He said local execution by a cloud agent was still being developed, and noted that shared channels and tools introduce permission questions when one agent can access several contexts.[29]

Shihipar distinguishes several safeguards for more autonomous agents. In his account, a model may refuse because of its training; separately, inference-time probes examine activations and a classifier can trigger a fallback. Auto Mode checks proposed actions against user permissions: an agent seeking a key to write to a production database should face a different decision depending on whether that write was authorized. He identifies cost, latency and false positives as practical concerns. These are his description of the safeguards, not an independent measure of their effectiveness.[29]