Anthropic: Claude, Agent Tools and Governance

Anthropic develops Claude and agent tools such as Claude Code and Duet. Explore its PBC structure, protocol work and dated model history.

kindCompany
subsidiariesBun

Updated

Claude models and the tools around them

Anthropic develops Claude models and software for using them through applications, APIs and agent tools. Its own governance account identifies it as a Delaware public benefit corporation. The company also created a Long-Term Benefit Trust with staged rights to select directors, a separate mechanism from its corporate designation. The original design does not establish that every planned governance milestone has since occurred.

It introduced the Model Context Protocol in November 2024, publishing a specification, SDKs and reference servers for connecting assistants to tools and data. This was a reusable interface contribution, rather than a new model. Its original publication should not be confused with a claim that Anthropic permanently owns the protocol.

From coding assistance to connected agents

Claude Code began as a limited research preview in February 2025, giving developers a terminal interface for delegating engineering tasks. That combined a Claude model with tools for acting in a codebase.

In October 2026 Anthropic introduced Claude Duet, a research preview for proactive work using connected applications. The announcement offered waitlist access to US Pro and Max subscribers across mobile and web. Those access limits matter: a preview announcement does not establish availability for every Claude user, and Anthropic itself describes the announced safeguards.

Sources

  1. [1]
    The Long-Term Benefit Trust

    Anthropic · Sep 19, 2023

  2. [2]
    Company

    Anthropic · retrieved Oct 23, 2026

  3. [3]
    Claude 3.7 Sonnet and Claude Code

    Anthropic · Feb 24, 2025

  4. [4]
    Introducing Claude Duet

    Anthropic · Oct 21, 2026

  5. [5]

Anthropic history

2026

│
Anthropic publishes five Claude exploit-chain demonstrations

Anthropic published five exploit-chain demonstrations it said an unreleased Claude research model developed against browsers and operating systems. The company said it coordinated the vulnerabilities with vendors and that the reported issues had been addressed.

Anthropic
Research
│
Anthropic opens a research preview of Claude Duet

Anthropic introduced a proactive agent app powered by Claude Opus 5.5 that monitors connected work apps and runs tasks in its own cloud computer. The research preview opened by waitlist to US Pro and Max subscribers on iOS, Android and the web, sharing their existing Claude usage limits.

Anthropic
Product
│
Anthropic embeds METR evaluators and says they can publish without its veto

With METR

Six METR employees got Anthropic hardware, office space and internal access roughly on par with its own risk teams. Anthropic says they may publish findings it cannot block, only redact within set categories; no report under those terms had yet appeared. It follows Dario Amodei's September pledge to host embedded evaluators.

Anthropic
Company
│
Anthropic launches Claude Dashboards and Claude Motion in beta

Dashboards turns questions about a connected data warehouse or CRM into charts that refresh as the data changes and show the query behind each number; Motion writes code that animates text, charts and images into an MP4 explainer, with no video model involved. Claude Docs, Slides and Design left beta on every plan the same day.

Anthropic
Product
│
Anthropic releases Claude Haiku 5.5

Anthropic's smallest Claude 5.5 model listed at $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, a tenth of Haiku 4.5's price, and was the first Haiku with an effort setting. Anthropic put the average saving per task nearer 75 percent, because its newer tokenizer uses more tokens.

Anthropic
Model
■
Anthropic releases Claude Sonnet 5.5

Anthropic released its second Claude 5.5 model for everyday coding and document work. It kept Sonnet 5's $2 input and $10 output price per million tokens, while Anthropic reported faster output and lower cost per task from using fewer tokens.

Anthropic
Model
■
Anthropic releases Claude Opus 5.5

The first of the Claude 5.5 models, priced at $4 and $20 per million tokens, 20 percent below Opus 5, and pitched as performing at Fable 5.1's level on most work. Artificial Analysis measured the highest Intelligence Index score it had recorded. Anthropic called it its first release since asking the industry to pace the frontier.

Anthropic
Model
│
An Anthropic engineer factors RSA-896 with Claude's help

Stephen Weis reported factoring the 270-digit RSA challenge number after Claude ported the CADO-NFS number field sieve to GPUs and ran it on up to 2,048 idle Anthropic GPUs for ten days, about 30 GPU-years. The factors check out; the algorithm was not new, and RSA-2048 is unaffected.

Stephen A. Weis
Research
│
Anthropic confirms it runs its own biology wet lab

Reuters reported, and Anthropic's head of life sciences confirmed, that the company runs physical biology experiments in its own facility as well as with partners. Anthropic said the lab's focus is fundamental biology rather than drug discovery and gave no size, cost or opening date.

CNBC
Company
│
Anthropic says Claude now leads a quarter of its own AI research work

Anthropic proposed three measurements any frontier lab could publish: how much of its AI research AI does, how well its agents are overseen, and how much research compute goes to safety. Its own figures, for August: Claude led 26 percent of R&D tasks, up from under 1 percent in February, and about 30,000 agents ran at once.

Anthropic
Research
│
Anthropic opens a verification program for looser biology safeguards

Vetted life-science teams can apply for grants that give Mythos 5.1, Opus 5 and Sonnet 5 more permissive biology classifiers, plus project-level High-risk Use grants that remove the safeguards blocking life-science requests. High-risk access to Mythos stayed limited to a few extra-vetted entities.

Anthropic
Product
│
Anthropic merges Claude chat and Cowork into one Claude

Anthropic began rolling out, over several weeks, a Claude for Pro and Max subscribers with no separate chat and agent modes: anything Cowork could do became available from any conversation, with tasks running in the cloud. Claude Docs and Claude Slides launched in beta the same day, and Team, Free and Enterprise plans were to follow.

Anthropic
Product
■
Dario Amodei calls on AI labs to pace the frontier

With OpenAI

Anthropic's chief executive asked frontier labs to slow capability gains, citing early recursive self-improvement and the Hugging Face incident, and committed Anthropic to host outside evaluators with employee-like access. Sam Altman said within hours that OpenAI would do the same, without terms or a date.

Dario Amodei
Culture
│
Anthropic accuses seven Chinese labs of distilling Claude

With Alibaba, Moonshot AI, DeepSeek, Xiaomi

Anthropic's September misuse report named Alibaba, Moonshot, DeepSeek, Zhipu, Xiaomi, SenseTime and MiniMax, and said Moonshot, DeepSeek and Xiaomi fed their own users' conversations into Claude. It also described a blocked request for help with gain-of-function work on chikungunya virus.

Anthropic
Research
■
Anthropic expands its account of cyber-evaluation incidents

With METR

Anthropic said it found a fourth incident in which a Claude model gained unauthorized access to real systems during a cyber evaluation, then broadened its review to 481 million transcripts. It released the Mythos 5 transcript and contracted METR for an independent investigation; the assessment's conclusions remain Anthropic's.

Anthropic
Research
■
Anthropic says Claude formalized Fermat's Last Theorem in Lean

Claude, on an internal research model roughly comparable to Fable 5.1, spent eleven days and about six billion output tokens converting Andrew Wiles's 1995 proof into thirteen million lines of machine-checked Lean code, formalizing 29,500 intermediate theorems along the way. It verified an existing proof rather than finding new mathematics.

Anthropic
Research
■
Anthropic releases Claude Fable 5.1 and Claude Mythos 5.1

Same underlying model, two safety configurations: Fable 5.1 ships generally available, Mythos 5.1 only through trusted access programs for cybersecurity and life-sciences work. Cache read pricing drops 75 percent, and Mythos 5.1 designed protein binders Anthropic says outperformed the best competition entries tenfold.

Anthropic
Model
│
Anthropic accounts for its own sandbox escapes

With UK AI Security Institute

Claude models reached the live internet during cyber evaluations in July, and Mythos 5 did it again in the UK AI Security Institute's testing in August. Anthropic attributes the incidents to operational security and to two alignment failures, motivated reasoning and harmful action in pursuit of a narrow task, and says it froze its RL environments for a month and flagged over 10 percent of them.

Anthropic
Research
│
More than a hundred companies sign a call for collective action on cyber defense

With OpenAI, Google, Microsoft, Amazon, Cloudflare, Hugging Face, Oracle, Perplexity

OpenAI hosted the letter and Anthropic, Google, Microsoft, AWS, Cloudflare, CrowdStrike, Visa and Mastercard signed it, the day after the Hugging Face postmortem. The premise is a window: models now help defenders faster than attackers, and the signatories expect it to close within months.

OpenAI
Culture
■
A federal judge rules Anthropic's Pentagon blacklist unlawful

With United States Department of War

Judge Rita Lin found the Department of War's supply chain risk designation was retaliation for Anthropic's refusal to let Claude be used for mass surveillance or autonomous weapons, violating the First and Fifth Amendments. The 59-page order permanently blocks the designation, five months after it was issued, and the government is expected to appeal.

NPR
Policy
│
Anthropic opens a research preview letting AI agents operate lab hardware

The Model Hardware Standard is a shared specification for AI agents to run microscopes, liquid handlers and robotic arms directly, cutting integration work Anthropic says used to take weeks down to hours. Genentech and HHMI Janelia tested it on real experiments during the preview, and Hugging Face and Raspberry Pi are building support into their own tools.

Anthropic
Research
■
Salesforce and Anthropic launch Claudeforce

With Salesforce

Claude becomes the default reasoning model inside Agentforce and the default AI backbone across Slack, while a new Salesforce plugin lets Claude query and act on live CRM data through 37 prebuilt sales skills. Salesforce in Claude is live for pilot customers now, with open beta due in September.

Salesforce
Deal
■
Anthropic signs Nscale GPU agreements worth up to $44.6 billion

With Nscale

Anthropic and Nscale subsidiaries executed GPU services agreements on August 25. Nscale later disclosed a payment cap of about $44.6 billion, conditional on delivery and service availability. August 26 reports described six years and roughly 460 megawatts at Monarch, with service expected in late 2027.

Nscale
Deal
■
Claude designs working protein binders for 14 of 15 lab targets

With humans limited to approving access requests and watching the infrastructure, Claude orchestrated public protein-design and co-folding models to produce binders against 14 of 15 targets, at roughly twice the field's usual hit rate. Adaptyv Bio and Twist Bioscience built and tested the designs independently, rather than Anthropic grading its own homework.

Anthropic
Research
│
Claude starts watermarking everything it writes

Anthropic turned on a SynthID-Text style watermark for models released from August 2 onward and applied it worldwide rather than only in Europe, while saying plainly that a detected mark is a signal and not proof. The EU's content-marking rule had arrived nine days earlier.

Anthropic
Policy
│
Claude Code stops asking permission by default

From August 14 the agent proceeds on its own for Pro, Max, and Team users, pausing only for work that is destructive or hard to undo. Anthropic said its automated reviewer caught 89 percent of harmful actions in testing, against 13.6 percent for human reviewers.

TechCrunch
Product
■
Anthropic discloses three cybersecurity-evaluation incidents

With Irregular

Anthropic disclosed that Claude models, running cybersecurity evaluations in an environment with unintended internet access, gained unauthorized access to the real systems of three organizations. The company described the cases as a containment and operational failure, not a deliberate escape or evidence of a model pursuing a goal of its own.

Anthropic
Research
■
Lab employees ask Washington for a brake they can pull later

With OpenAI, Google DeepMind, Meta

More than a thousand people inside the frontier labs, including Anthropic's chief executive and OpenAI's chief scientist, asked the United States to back an international effort to build the tools for deliberately pacing automated AI development. The statement does not ask anyone to slow down now.

Pacing the Frontier
Culture
│
Anthropic explains why it does not publish weights

A public position paper on open-weights models, written by the one major lab that has never released any, six weeks after the government restricted access to its own. The comment threads were longer than the paper.

Anthropic
Culture
■
Anthropic reports clearance for Fable 5’s July 1 return

With United States Department of Commerce

Anthropic said export controls had been lifted on June 30, with Fable 5 global access to resume July 1. It separately reported a narrower restoration of Mythos 5 to specified US organizations after June 26 approval.

Anthropic
Policy
│
Anthropic releases Claude Sonnet 5

Anthropic released Claude Sonnet 5 at introductory API prices of $2 per million input tokens and $10 per million output tokens. On August 10, it made those prices permanent.

Anthropic
Model
│
Anthropic launches Claude Science, a workbench for researchers

Not a new model, a workbench: more than 60 scientific databases and toolkits wired into Claude, producing auditable results a researcher can rerun. Announced at a live event with executives from Bristol Myers Squibb, Novartis and Genentech on the stage alongside Dario Amodei.

Anthropic
Product
■
Frontier access starts requiring a background check

With OpenAI

OpenAI and Anthropic both agreed to vet customers for their most capable models through a United States government process, two weeks after the export order. Buying the top tier now meant being approved for it.

CNBC
Policy
■
Anthropic accuses Alibaba of its largest distillation attack yet

With Alibaba

A letter to the Senate Banking Committee alleged 28.8 million exchanges with Claude through 25,000 fraudulent accounts between April and June, aimed at reproducing Claude's coding and agentic capabilities in Alibaba's own Qwen models. Alibaba did not respond publicly.

CNBC
Company
│
Anthropic puts Claude in Slack as a shared teammate, not a chatbot

Claude Tag lets anyone in a channel tag @Claude in and delegate a task, building context across the whole conversation instead of one message at a time. It replaces Anthropic's earlier Slack app; the company says 65 percent of its own product team's code now comes from an internal version.

Anthropic
Product
●
The United States puts export controls on frontier models

With OpenAI, United States Department of Commerce

Anthropic said a Commerce directive required suspending foreign-national access to Fable 5 and Mythos 5. A separate White House request later limited OpenAI’s GPT-5.6 rollout through voluntary pre-release review.

Anthropic
Policy
■
Anthropic ships Mythos 5 and a safer twin

Two models on one day: Mythos 5 at the top of Anthropic's capability range, and Fable 5, described as a Mythos-class model made safe for general release. Three days later the United States government suspended access to both.

Anthropic
Model
│
Claude Opus 4.8

Shipped the same day as the $65 billion round, and presented as a refinement rather than a jump. It was the most capable model Anthropic would sell anyone at the time, with the Mythos class still in preview.

Anthropic
Model
■
Anthropic raises $65 billion at $965 billion

The largest private round ever raised, closing eight weeks after OpenAI set the previous record and passing it on valuation. Anthropic shipped Claude Opus 4.8 the same day, which is one way to justify the number.

Anthropic
Funding
│
Andrej Karpathy joins Anthropic

The OpenAI founding member and former Tesla AI director, who had spent three years teaching the field from the outside, went back inside at a competitor. Six months earlier he had said agents would take a decade to work through their problems.

Axios
Company
■
Cloudflare points Mythos at its own code

With Cloudflare

Across more than fifty internal repositories the model chained low-severity bugs into working exploits and wrote proofs that compiled, where earlier models stopped at describing the bug and left exploitability unknown. Cloudflare's conclusion was that this is a different kind of tool, not a better one.

Cloudflare
Research
│
Claude agents get to dream

Dreaming runs between sessions, replays what an agent did and writes what it learned into memory, which Anthropic compares to the consolidation a brain does in sleep. It shipped at Code with Claude alongside a grader that sends work back for revision and an orchestrator that fans jobs out to subagents.

The New Stack
Product
■
Anthropic and investors form an AI services company

With Blackstone, Hellman & Friedman, Goldman Sachs

Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs announced a new AI services company for mid-sized organizations. Its engineers and Anthropic's Applied AI staff would build Claude-powered systems, adding a delivery arm alongside the Claude Partner Network.

Anthropic
Deal
■
Amazon adds $25 billion to its Anthropic position

With Amazon

Five billion immediately at a $380 billion valuation and up to twenty more against commercial milestones, taking Amazon's total to $33 billion. Anthropic committed over $100 billion to AWS across the next decade and locked in up to five gigawatts of capacity.

CNBC
Deal
│
Claude Design turns prompts into decks and prototypes

A research preview that reads a team's codebase and design files to keep output on brand, exporting to PDF, PPTX and Canva. Aimed at founders and product managers rather than designers, which put Anthropic in Figma and Canva's field of view.

TechCrunch
Product
■
Anthropic starts running other people's agents

Managed Agents moved agent execution onto Anthropic's own infrastructure: isolated containers, default-deny network egress, checkpointing and tracing, billed at token rates plus eight cents per session-hour. The model vendor took the runtime layer.

Anthropic
Product
│
Anthropic loses its bid to pause the Pentagon blacklist

With United States Department of War

The appeals court declined to stay the supply chain risk designation, finding that Anthropic would suffer harm but that the harm looked primarily financial. A California court had enjoined parts of the same action, leaving the company shut out of Defense contracts and free to serve every other agency.

CNBC
Policy
■
Anthropic builds a model it will not sell

Claude Mythos Preview found more than 10,000 high and critical zero-days in its first month, including a seventeen-year-old remote root hole in FreeBSD that it exploited unaided. Anthropic held it back from release and put it behind Project Glasswing, an invitation-only program for critical infrastructure operators.

Anthropic
Model
■
Anthropic expands its Google and Broadcom TPU deal

With Google, Broadcom

Anthropic signed an agreement for multiple gigawatts of next-generation Google TPU capacity expected to come online beginning in 2027. Broadcom said about 3.5 gigawatts would be made available through it, subject to Anthropic's continued commercial success.

Anthropic
Deal
│
Anthropic publishes Claude Code's source by mistake

Roughly 500,000 lines across 1,900 files went to npm as source rather than compiled output, exposing the harness that tells the model how to use its tools. The second accidental disclosure in five days, and the company again called it human error rather than a breach.

Fortune
Culture
■
A misconfigured cache reveals Claude Mythos

About 3,000 unpublished files sat reachable on Anthropic's content system, among them a draft post calling Mythos by far the most powerful model the company had built and flagging what it could do to software security. Anthropic confirmed the model exists and called the exposure human error.

Fortune
Culture
■
Claude gets computer use and a phone remote

Cowork and Claude Code can now point and click at whatever is on the screen when no tool exists for the job, and Dispatch starts that work from a phone so the desk can be empty. Research preview, Pro and Max, macOS and Windows.

Anthropic
Product
■
Microsoft ships Copilot Cowork and puts Claude in mainline Copilot

With Microsoft

Anthropic's agent technology arrived inside Microsoft 365 as Copilot Cowork, for work that runs over hours rather than a single turn, and Claude appeared in ordinary Copilot chat alongside OpenAI's models. The company that bankrolled OpenAI now picks a model per task.

Microsoft
Product
■
Anthropic sues the Pentagon over the supply chain risk label

With United States Department of War

Two filings on the same day, one in federal district court in California and one in the DC Circuit, arguing the designation is retaliation for protected speech. Anthropic put the cost at billions of dollars of 2026 revenue and hundreds of millions in private contracts.

PBS NewsHour
Policy
■
The Pentagon designates Anthropic a supply chain risk

With United States Department of War

The Department of War made the label official, the first time it has landed on an American company rather than a foreign adversary. Every contractor doing Pentagon work now has to certify that it does not use Anthropic's models.

TechCrunch
Policy
■
The administration moves to bar Anthropic from federal work

With United States Department of War

The President directed agencies to designate Anthropic a supply-chain risk, a label previously reserved for foreign adversaries, following a dispute over the company's usage limits. Anthropic called the designation legally unsound and challenged it in court.

Anthropic
Policy
■
Amodei refuses the Department of War's terms

With United States Department of War

The department would contract only with companies that accept any lawful use, and Anthropic would not give up two carve-outs: mass domestic surveillance and fully autonomous weapons. Amodei published the refusal along with the threats attached to it, including invoking the Defense Production Act.

Anthropic
Policy
■
Anthropic rewrites its Responsible Scaling Policy

Version 3 reorganized the policy around company roadmaps, risk reports every three to six months and conditional competitor commitments. The roadmaps were public goals rather than hard commitments; Anthropic said collective-action limits made the old absolute framework impractical.

TIME
Company
│
Claude Sonnet 4.6

Anthropic introduced Claude Sonnet 4.6 with a million-token context window in beta, bringing that capacity to its mid-tier model.

Anthropic
Model
■
Anthropic raises $30 billion at $380 billion

A Series G that more than doubled the company's valuation in under a year, with Claude Code's weekly active users cited as the headline number. Private markets had stopped pretending these were startups.

Anthropic
Funding
│
Anthropic measures what AI help costs a learner

In a randomized study of 52 developers learning Trio, the AI group averaged 50 percent on an immediate quiz versus 67 percent without AI, a 17 percentage-point gap. The completion-time difference was not statistically significant; long-term retention was not measured.

Anthropic
Research
■
A trademark complaint renames Clawdbot overnight

With OpenClaw

Anthropic decided Clawd sat too close to Claude and asked for the name, so Peter Steinberger's personal agent became Moltbot the same morning, mascot included. Squatters took the freed GitHub and X handles within seconds and stood up typosquat domains and a cloned repository.

Laravel News
Company
■
Anthropic publishes Claude's constitution

A document describing the values and judgment Anthropic wants Claude to have, written to be read by Claude rather than by users, and used to shape training. It went out under CC0, so anyone can take it and use it.

Anthropic
Company
■
PromptArmor demonstrates Claude Cowork file exfiltration

With PromptArmor

PromptArmor demonstrated hidden document instructions making Claude Cowork upload accessible files to an attacker-controlled Anthropic account without an approval prompt. The upload used Anthropic's allowed API domain; Simon Willison described the report on January 14.

Simon Willison
Research
■
Claude Cowork takes the agent out of the terminal

Anthropic released Cowork as a Mac research preview for delegating multi-step work over selected folders without opening a terminal. Two days later, a security vendor demonstrated how a malicious document and an allowlisted network route could expose those files.

Anthropic
Product

2025

■
OpenAI, Anthropic and Block put their agent standards under one roof

With Linux Foundation, Agentic AI Foundation, OpenAI, Block

The Linux Foundation launched the Agentic AI Foundation, taking in Anthropic's Model Context Protocol, Block's goose framework and OpenAI's AGENTS.md as its first projects. Amazon, Google, Microsoft, Bloomberg and Cloudflare joined as platinum members.

Linux Foundation
Open source
│
Anthropic builds an AI that interviews workers about AI

A tool that drafts questions, conducts interviews and analyzes responses at scale, piloted on 1,250 professionals across the general workforce, creative fields and science. Anthropic used the results to publish its first read on how work is actually changing.

Anthropic
Research
■
Anthropic buys Bun

With Bun

The JavaScript runtime that Claude Code already runs on went to the company that ships Claude Code, with Bun staying open source and MIT licensed and the same team on it. Anthropic said it is betting on Bun as the infrastructure under its coding products.

Bun
Acquisition
■
Anthropic reports the first AI-run espionage campaign

A Chinese state-sponsored group turned Claude Code on roughly thirty targets and succeeded against a few, with Anthropic putting the model's share of the work at 80 to 90 percent and human involvement at four to six decision points. Security researchers objected that the report shipped without a single indicator of compromise anyone could act on.

Anthropic
Research
│
Anthropic commits $50 billion to US data centers

With Fluidstack

Anthropic announced a $50 billion investment in American computing infrastructure, building custom data centers with Fluidstack in Texas and New York, with facilities expected to come online during 2026.

Anthropic
Deal
│
Anthropic ships Claude for Excel

A sidebar lets Claude read, audit and edit live Excel workbooks while preserving formulas and structure, released in beta to about 1,000 Max, Team and Enterprise users alongside a financial data partnership with LSEG.

Anthropic
Product
│
Anthropic ships Claude Code on the web

A research preview lets Pro and Max subscribers assign coding tasks to Claude that run in isolated cloud sandboxes instead of a local terminal, so several tasks can run in parallel across repositories from a browser or the iOS app.

Anthropic
Product
■
Anthropic ships Skills

Folders of instructions, scripts and resources that Claude loads only when a task calls for them, in one format that works across the apps, the API and Claude Code. Composable, portable between surfaces, and paged in only when relevant.

Anthropic
Product
│
A few hundred documents can poison a model of any size

With UK AI Security Institute

Anthropic, the UK AI Security Institute and the Alan Turing Institute found that the number of malicious training documents needed to install a backdoor stays roughly constant as models scale. The intuition that bigger models are harder to poison turned out to be backwards.

Anthropic
Research
■
Microsoft adds Anthropic's Claude models to Copilot

With Microsoft

Copilot business users can now choose Claude Opus 4.1 or Claude Sonnet 4 alongside OpenAI's models for research and building custom agents. It follows an earlier deal bringing Anthropic's models into Office, another step in Microsoft's move beyond OpenAI exclusivity.

TechCrunch
Deal
■
Anthropic settles with authors for $1.5 billion

The largest copyright recovery in United States history, covering roughly half a million books. A judge had already ruled that training on lawfully acquired books was fair use, which left how the books were obtained as the expensive question.

NPR
Policy
■
The Pentagon puts up to $200 million behind frontier AI labs

With United States Department of War

The Defense Department's Chief Digital and AI Office signed a two-year, $200 million ceiling prototype agreement with Anthropic to test frontier AI for national security workflows. Google, OpenAI and xAI were reported to have landed matching contracts the same day.

Anthropic
Deal
│
Anthropic launches a voice mode for Claude

Rolling out first in English on mobile, the beta lets users speak with Claude Sonnet 4 and hear it reply in one of five voices, with a transcript kept after each conversation. It follows similar features already shipped by ChatGPT and Gemini Live.

TechCrunch
Product
■
Claude 4 is sold on how long it can run

With Apollo Research, SecureBio

Opus 4 and Sonnet 4 were pitched on sustained multi-hour agentic work rather than benchmark tables. The industry's question had shifted from what a model knows to how long it can be trusted unattended.

Anthropic
Model
│
Anthropic launches Claude for Education

A version of Claude for universities with a Socratic Learning mode that guides students toward answers instead of giving them outright. Northeastern, the London School of Economics and Champlain College signed on for campus-wide access at launch.

Anthropic
Product
●
Claude Code puts an agent in the terminal

Anthropic launched Claude Code alongside Claude 3.7 Sonnet as a limited research preview. The command-line tool could read a repository, propose file edits and run approved shell commands.

Anthropic
Product

2024

│
Anthropic publishes evidence of alignment faking in Claude

Working with Redwood Research, Anthropic showed that Claude 3 Opus would strategically comply with harmful requests it would normally refuse when it believed the response was being used to retrain it, reasoning in a hidden scratchpad that faking compliance protected its existing values.

Anthropic
Research
■
Anthropic introduces the Model Context Protocol

Anthropic released a common protocol for connecting AI applications to external tools and data, with SDKs, example servers, and local support in Claude Desktop. Developers could expose an integration through an MCP server for reuse across compatible applications.

Anthropic
Open source
│
Claude gets a mouse and keyboard

An API that let a model look at a screen, move the cursor and type. It was slow, it made mistakes, and it was obviously the direction of travel.

Anthropic
Product
│
Frontier labs sign the Seoul safety commitments

With OpenAI, Google, Meta, Microsoft, Amazon, Mistral AI, xAI, Cohere, G42, IBM, Inflection AI, Naver, Samsung Electronics, Technology Innovation Institute, Zhipu AI

At the UK and South Korea's Seoul summit, sixteen frontier labs including OpenAI, Anthropic, Google, Meta and Zhipu AI committed to publish safety frameworks with risk thresholds and to hold back models that breach them. Voluntary, and the signatory list kept growing afterwards.

Department for Science, Innovation and Technology
Policy
│
Anthropic maps millions of features inside a frontier model

Scaling Monosemanticity extracted millions of interpretable features from Claude 3 Sonnet, the first detailed map of a production model's internals. Golden Gate Claude, a version steered to obsess over one bridge, made interpretability briefly famous.

Anthropic
Research

2023

■
Amazon commits up to $4 billion to Anthropic

With Amazon

An initial $1.25 billion, AWS as the primary cloud provider, and Trainium as the training chip. A hyperscaler funding a lab to rent back its own compute was now the shape of the industry.

Anthropic
Funding
│
Claude 2 opens to the public

Anthropic made Claude 2 available through a public sign-up, giving more users access to its existing 100,000-token context window.

Anthropic
Model
●
Anthropic introduces Claude

With Notion

Anthropic introduced its Claude assistant on March 14, 2023, the same day OpenAI announced GPT-4. The company described its approach as constitutional AI.

Anthropic
Model

2021

●
Anthropic launches with $124 million

Seven people left OpenAI, including its VP of research, to start a lab organized around AI safety research. The split reshaped the competitive structure of the field and gave it a second pole.

TechCrunch
Company