Anthropic taught Claude to see the job through.

In a study Anthropic published this October, an internal research version of Claude spent about 64 hours on a problem few people would associate with a chatbot: designing tiny proteins that latch onto scorpion neurotoxins. Working mostly on its own inside Claude Science, the company's research workbench, it generated 318,484 candidate designs and shortlisted 180. Researchers checked in with four follow-up prompts. When two contract labs tested the final set of 211 designs, 17 bound to the toxins they were built to target.

Anthropic is just as careful about what the result is not. It has not shown that the binders neutralize venom in living cells, and it says plainly that it has not made an antivenom. A striking result published next to a clear account of its limits is a fair introduction to the company behind Claude.

Anthropic is a public benefit corporation, and its stated mission is to help the world make the transition through transformative AI safely. Since it began offering Claude broadly in March 2023, that mission has grown into something unusually wide: a family of models, apps for everyday work, a coding agent, a platform for builders and a steady stream of published safety research.

A family of models, sized for the job#

Claude is a lineup rather than a single model, and the names borrow from literature: a haiku is short, a sonnet longer, an opus a major work.

Claude Haiku 5.5, released October 7, is the fastest. It is built for high-volume work such as classification, extraction and routing, and it is quick enough to serve as a subagent, a helper that a larger model hands small jobs to. It is also the first Haiku-class model with an adjustable effort setting, so developers can trade a little speed for more thought.

Claude Sonnet 5.5, from late September, is pitched as the best mix of speed and intelligence. It is strongest at well-scoped everyday work: fixing bugs and building polished documents, slides and spreadsheets.

Claude Opus 5.5, released September 22, is the model Anthropic tells developers to start with for most workloads. It is built for long-running agentic coding and knowledge work, where the model plans its own steps and uses tools such as a terminal to carry them out. Anthropic reports that one early tester used it to finish a 680,000-line code migration in less than a day.

At the top sits Claude Fable 5.1, released September 1 for demanding reasoning and long-horizon agentic work. Its twin, Claude Mythos 5.1, has the same capabilities but is offered only to organizations verified through programs such as Anthropic's Cyber Verification Program.

Every current model reads text and images, uses tools and decides for itself how long to think before answering, steered by an effort setting. Each also has a context window, the amount of material a model can keep in view at once, of one million tokens. Tokens are the word fragments that models read, and Anthropic puts a million of them at roughly 555,000 words: enough for a sizable codebase or a stack of contracts in a single request.

Claude now picks up the legwork.#

For most people, Claude is the app at claude.ai, on the desktop and on iOS and Android. Projects keep files and instructions together, Memory carries context between conversations, and web search and voice are built in. Artifacts are the shareable documents, slides and interactive visuals Claude builds alongside a chat. Connectors link Claude to the tools where work already lives, and extensions put it inside Chrome, Google Workspace and Microsoft 365.

What has changed this year is how much Claude does on its own. Anthropic introduced Claude Cowork as a separate agentic mode for non-coding work, then folded it into Claude itself, a change now rolling out across plans. You describe an outcome and Claude works out the steps across the folders and tools you choose. It shows its plan and waits for approval before significant actions, and deleting anything needs your sign-off. Tasks can run on a schedule, split into parallel pieces, or start on your phone and keep going in the cloud with your laptop closed.

Claude Design brings the same approach to visuals. Describe a slide, a page or a one-pager and Claude builds a first version, checked against your team's design system if you import one, then exports it to PDF, PowerPoint or HTML. Design, Slides and Docs are included on every plan, Free among them.

For teams, Claude Tag, in beta on Team and Enterprise plans, brings Claude into Slack as a colleague anyone in a channel can tag. It can work on a task for hours or days, schedule its own follow-ups, and leave admins a log of what it did and who asked.

Claude Code changed what one developer can take on.#

Claude Code is Anthropic's coding agent, and its pitch is disarmingly simple: hand it a bug fix, a test or a multi-day migration. It runs in the terminal, in VS Code and JetBrains, in a desktop app, on the web and on phones.

Point it at an unfamiliar codebase and it maps and explains the structure in seconds. Give it an issue and it reads the ticket, writes the code, runs the tests and opens a pull request. On large refactors it follows imports across the repository and keeps going when something breaks.

It also knows when to ask. Claude Code requests permission before it changes files or runs commands, and it talks to the model directly from your machine, with no separate server or remote index of your code. It is included with Pro, Max, Team and Enterprise plans.

A platform for people who build#

Underneath the apps is the Claude Platform, Anthropic's API: the interface that lets developers send Claude a request from their own software and get an answer back. The same models are offered through Amazon Bedrock, Google Cloud and Microsoft Foundry, so teams can use Claude inside the cloud they already run on.

For developers building agents, Anthropic offers two ready-made foundations. The Claude Agent SDK packages the tools, agent loop and context management that power Claude Code as a library for Python and TypeScript. Claude Managed Agents, now in beta, goes a step further and hosts the whole thing. You define an agent's model, instructions and tools, and Claude works in a secure sandbox where it can read files, run commands, browse the web and execute code for minutes or hours, on a schedule if you like.

Two open standards that started at Anthropic now reach well beyond it. The Model Context Protocol, introduced in November 2024, gives AI tools a common way to connect to data sources and services. Anthropic donated it to the Agentic AI Foundation, a neutral home for such standards under the Linux Foundation. Agent Skills, packaged folders of instructions and resources that teach an agent a specialized workflow, followed the same path this October, after more than 70 agent platforms had adopted the format.

Science gets a workbench.#

Claude Science, the workbench behind the scorpion-toxin study, launched in beta in June for Pro, Max, Team and Enterprise plans on macOS and Linux. It comes set up for genomics, single-cell biology, proteomics and cheminformatics, with access to more than 60 scientific databases. It can submit jobs to a lab's computing cluster and scale from one GPU to hundreds.

The detail scientists may appreciate most is the paperwork. Every figure and manuscript it produces comes bundled with the code and environment that made it, plus a plain-language description of the method, so a colleague can rerun the work instead of taking it on trust.

Safety you can read#

Anthropic publishes an unusual amount about how Claude is meant to behave and how it is tested. Claude's constitution, a detailed statement of the values and behavior Anthropic intends for Claude, is used directly in training and released under a CC0 public-domain dedication, so anyone can reuse it.

The Responsible Scaling Policy, first released in September 2023 and now at version 3.4, ties specific capability thresholds to stronger security and deployment safeguards. Each new model ships with a system card describing its safety evaluations. The Mythos arrangement is that thinking in practice: the same capabilities, with access gated by verification. Anthropic also reports that Opus 5.5 resists prompt injection, the trick of hiding instructions in a web page or document to hijack an agent, better than its predecessor did.

Then there is interpretability, the research that tries to see inside a model. As Anthropic's team puts it, models like Claude "talk in words but think in numbers," and much of its published work tries to translate. That includes a 2025 study tracing a model's internal reasoning step by step, open-source tools for that kind of circuit tracing, and a 2026 paper describing a workspace inside Claude that holds thoughts it never writes out.

What it adds up to#

Put together, this is a remarkable toolkit: models that take in a whole codebase at once, agents that work for hours in a sandbox, open standards that plug them into real systems, and published policies that explain how they are meant to behave.

It is a foundation on which a few people can build what once took a department, and on which a researcher, a teacher or a shop owner can get real work done without writing a line of code. To see why it matters, open Claude, hand it something you have been putting off, and watch it work.