What is Claude? It is Anthropic’s family of AI models and the unified assistant built on top of them, used for writing, research, coding, analysis and, increasingly, for running multi-step tasks on your behalf. As of late September 2026 it spans five model tiers, from fast Haiku 4.5 up to frontier-class Fable 5.1, with the flagship Opus tier just refreshed to Opus 5.5 on September 22, and free and paid plans to match.
The product itself changed shape recently. On September 16, 2026, Anthropic folded its separate agentic product, Claude Cowork, back into the main Claude chat experience, so that chat and autonomous task execution now live in one assistant instead of two separate tools you had to choose between.
This Claude overview covers what it actually does, what each plan costs, who it suits, and where it still falls short, without the marketing gloss. If you are also comparing coding assistants specifically, see how Claude Code, Anthropic’s coding-focused product, against Cursor stacks up on that narrower question.
What Is Claude, Exactly?
Claude is a large language model family from Anthropic, an AI safety company founded in 2021 by former OpenAI researchers, plus the chat and agent product built around those models. Anthropic trains Claude using a method it calls Constitutional AI, meant to keep responses helpful while steering away from harmful outputs.
The name is not an acronym. Anthropic named the models after Claude Shannon, the mathematician and engineer widely credited as the father of information theory, according to Wikipedia’s release history of the Claude language model. The first Claude shipped in limited access in March 2023, and the lineup has gone through Claude 2, Claude 3’s Haiku/Sonnet/Opus split, Claude 3.5, Claude 4 and now the 2026 tiers described below.
Claude explained in one line: it is a text-first assistant that reads, writes, reasons over long documents and code, and now increasingly acts, rather than a tool that only answers one question at a time.
How Does Claude Work?
Claude works by predicting the next piece of text in a conversation, the same core mechanism as any large language model, but tuned through reinforcement learning and Constitutional AI so it follows instructions, cites its reasoning when asked, and refuses requests that conflict with its guidelines. Recent models add long context windows and tool use on top of that.
Under the hood, three things changed the most since 2024. First, context windows grew: current Sonnet 5, Opus 5.5 and Fable 5.1 all support up to 1 million tokens, roughly 750,000 words, in a single conversation, per Anthropic’s own product overview. Second, “extended thinking” lets a model spend more computation on harder problems before answering, at the cost of latency. Third, Claude can now call tools directly: it browses the web, runs code, reads files you connect, and edits documents in place through Claude Docs and Claude Slides rather than just describing what you should do.
None of that makes Claude infallible. Longer context does not mean perfect recall of everything in it, which matters more once you get to the limitations section below.
What Does Claude Do?
Claude drafts and edits writing, analyzes and summarizes documents, writes and debugs code, answers research questions with citations, and now creates working files directly: spreadsheets, slide decks, documents and simple charts, inside the same conversation where you asked for them.
Concretely, the assistant now bundles what used to be several separate Anthropic products. Claude Docs (still in beta) gives you real-time collaborative document editing with AI drafting and commenting, exporting to Word or Google Docs. Claude Slides builds and edits presentations you can download as PowerPoint or PDF. Claude Design turns a prompt into a visual layout. Claude Code, still a distinct install, handles software engineering from the command line, an IDE extension, or a browser. A memory system remembers project details and preferences across sessions unless you turn it off, and connectors let Claude read and act on files in Google Drive, Slack, GitHub and similar apps you authorize.
One thing Claude still will not do directly is generate photorealistic images or illustrations. Its official help center confirms it produces diagrams, charts and interactive HTML/SVG visuals instead, not artwork, so teams that need actual image generation still pair Claude with a dedicated image model.
What Is Claude Used For?
People use Claude for long-document analysis, first-draft writing, code generation and review, research summarization, and delegating recurring multi-step tasks such as compiling a weekly report from several data sources. It is also used inside other software through Anthropic’s API, which powers a large share of third-party AI features you never see branded as Claude.
Inside a business, the most common uses cluster into four buckets: drafting and editing (proposals, policies, marketing copy, course content), analysis (reading contracts, transcripts, survey data or specs and pulling out what matters), software development (Claude Code handles a meaningful share of day-to-day coding work at companies that have adopted it), and light automation (scheduled tasks that check a dashboard, draft a summary, or flag anomalies without a person kicking each one off manually).
Which Claude Models Are Available Right Now?
As of late September 2026, Anthropic offers four models most users can access: Haiku 4.5, Sonnet 5, Opus 5.5 and Fable 5.1, priced from $1 to $50 per million tokens depending on tier, plus Mythos 5.1, a restricted-access variant of Fable 5.1 for vetted cybersecurity and biomedical research partners only.
Fable and Mythos moved to version 5.1 on September 1, 2026, per Anthropic’s own announcement: roughly 25 percent cheaper for typical workloads, up to 45 percent cheaper for heavily agentic work thanks to a 75 percent cut in cache-read pricing, and with biology and cybersecurity safeguards that trigger far less often on benign requests. Three weeks later, on September 22, 2026, Opus itself was refreshed to Opus 5.5, succeeding the two-month-old Opus 5 at a lower price ($20 rather than $25 per million output tokens) while, according to Anthropic’s own product page, matching Fable 5.1 on most coding and reasoning work. Early developer reaction, tracked in Every’s vibe-check roundup, was largely positive: several engineers who had switched to OpenAI’s Codex described reconsidering that move, citing more idiomatic code and better instruction-following than the previous Opus. The same roundup and independent testing from Simon Willison both flag a real weakness, though: pushed to its maximum thinking-effort setting, Opus 5.5 tends to overthink simple tasks until it hits its own token limit, so the default or a middle effort setting is usually the better choice. Sonnet 5 remains the default the API and chat interface reach for unless you specify otherwise.
| Model | Context window | Price per million tokens (input / output) | Best for |
|---|---|---|---|
| Claude Fable 5.1 | 1M tokens | $10 / $50 | The hardest reasoning, engineering and research problems |
| Claude Opus 5.5 | 1M tokens | $4 / $20 | Complex agentic coding, large refactors, deep research |
| Claude Sonnet 5 | 1M tokens | $2 / $10 | Default choice for most day-to-day production work |
| Claude Haiku 4.5 | 200K tokens | $1 / $5 | High-volume, latency-sensitive, simple tasks |
Pick the Cheapest Model That Passes
Start every recurring workflow on Sonnet 5 or even Haiku 4.5, and only escalate to Opus 5.5 or Fable 5.1 for the specific steps that actually fail, since output tokens, not input tokens, drive most of the cost gap between tiers.
Who Is Claude Built For?
Claude is built for knowledge workers doing document-heavy reasoning, developers building or shipping software, and organizations that need admin controls, audit logs and compliance features layered on top of a general assistant. Anthropic markets it across marketing, sales, finance, HR, legal, product and engineering teams, not just to individual power users.
In practice, the strongest fit is people whose work involves reading or writing a lot of text under real constraints: contracts, specs, curricula, transcripts, code. A solo course designer drafting learning objectives and assessment items, a compliance officer summarizing a policy change, and a developer refactoring a legacy codebase are all squarely in Claude’s lane. A social media manager who mainly needs on-brand images, or a team that lives entirely inside Google Workspace’s native AI, will find less of a gap to close.
Is Claude Free to Use?
Yes, Claude has a genuinely usable free tier with chat, web search, memory, file creation and artifacts, but message limits reset often and heavy users hit them quickly. Paid plans start at $17 to $20 a month for an individual and scale up through Max, Team and Enterprise tiers for higher usage and admin controls.
| Plan | Price | Who it fits |
|---|---|---|
| Free | $0 | Occasional chat, web search, memory and artifacts |
| Pro | $17/mo billed annually, or $20/mo | Individuals wanting Claude Code, Docs, Slides and more usage headroom |
| Max | From $100/mo | Power users needing 5x to 20x the usage and higher output limits |
| Team | $20-$125 per seat/mo | Groups of 2-150 needing SSO and shared admin controls |
| Enterprise | $20/seat/mo plus usage at API rates | Organizations needing SCIM, audit logs and HIPAA-ready deployment |
Those figures come straight from Anthropic’s current pricing page, which is worth checking directly since Anthropic has adjusted usage caps on the Pro tier multiple times in 2026. For a fuller breakdown of what each tier includes and how the API costs compare, see our guide to current Claude pricing.
What Can Claude Not Do?
Claude cannot generate photorealistic images, cannot guarantee perfect recall across a full 1-million-token context, and gets measurably less reliable the longer and more autonomous a task chain runs. It also still hallucinates, more on open-ended questions than on grounded ones, and its published benchmark scores tend to overstate real-world performance on unfamiliar work.
The gaps worth planning around, in order of how often they bite:
Long-context retrieval is not perfect. Independent testing summarized by Digitortoise’s 2026 roundup of documented Claude limitations found that even Opus-class models miss roughly one in four items on multi-needle retrieval tests inside a supposedly 1-million-token window, and Claude Code specifically showed circular reasoning creeping in around 20 percent context usage.
Benchmark scores drop hard on unfamiliar work. The same roundup notes Opus 5 falls from 96 percent on the standard SWE-bench Verified coding benchmark to roughly 79 percent on a harder “Pro” variant designed to reduce task familiarity, a 17-to-20-point gap that shows up across model generations, not just one.
Agentic reliability decays with task length. At 80 percent reliability, current Opus-class models handle roughly 70 minutes of autonomous work; stretch that to 50 percent reliability and the ceiling rises to about 12 hours, meaning long unsupervised runs need a human checkpoint built in, not blind trust.
It does not generate images. For anything beyond charts, diagrams and HTML/SVG visuals, you still need a dedicated image model alongside Claude, not instead of it.
It is not free of usage friction. Pro-tier rate limits have shifted several times in 2026, and the practical jump from Pro to Max, from around $20 to $100-plus a month, catches teams off guard when a single heavy user outgrows the entry plan.
Checkpoint Long Agent Runs
For any task you let Claude run unsupervised for more than an hour, build in an explicit checkpoint where it reports back before continuing, since reliability drops well before the stated context or task-duration limits are reached.
How Is Claude Different From the Alternatives?
Claude’s main differentiators against ChatGPT and Gemini are its long context windows, its emphasis on safety training and steerability, and its stronger reputation for coding through Claude Code, while it trails on native image generation and deep integration with a single ecosystem like Google Workspace. None of the three is strictly better across every task.
Reviewers who have used it for extended periods tend to land in a similar place: strong at document reasoning and coding, honest about uncertainty rather than confidently wrong, but pricier at the high end and firmer about refusing certain requests than some users want. If your comparison is specifically about coding tools rather than general assistants, the more detailed matchup of Claude Code against Cursor is the better next read.
How Much Does Claude Cost for Teams and Enterprise Deployments?
Team plans run $20 to $25 a month per standard seat, or $100 to $125 for premium seats with 5x the usage, both billed for groups of 2 to 150 people with SSO and admin controls included. Enterprise deployments start at $20 per seat monthly plus usage billed at API rates, adding SCIM provisioning, audit logs, a compliance API and HIPAA-ready configurations.
The jump from Team to Enterprise is less about raw model access and more about governance: custom data retention windows, role-based access control, and the audit trail a compliance or security team will actually ask for before approving a company-wide rollout. Budget for usage costs separately from seat costs, since a small group of heavy Claude Code users can burn through more in API-rate usage than the seat fees themselves.
How Can Training and L&D Teams Put Claude to Work?
Learning and development teams use Claude to draft learning objectives and assessment items against Bloom’s taxonomy, summarize SME interviews into structured content outlines, and turn a policy document into a first-pass compliance training script, all inside one conversation instead of stitching several tools together.
It is not a full authoring tool and will not publish SCORM or xAPI packages by itself, so the realistic workflow is Claude for drafting and iteration, then your LMS or authoring tool for packaging and delivery. Teams building out a broader AI stack alongside Claude can compare notes against our roundup of AI tools for training and L&D use cases more broadly, or the narrower list of AI tools for developers if your team is evaluating Claude Code alongside other coding assistants.
Conclusion
Claude in September 2026 is a genuinely capable, actively evolving assistant: five model tiers, a merged chat-plus-agent product, and pricing that scales from free to enterprise. It earns its reputation on long-document reasoning and coding, and it is honest, in Anthropic’s own materials and in independent testing, about where it still falls short on recall, agentic reliability and image generation.
The practical next step is not to take any vendor’s word for current limits. Start on the free tier or a Pro trial with a real task from your own backlog, a contract to summarize, a codebase to refactor, a course outline to draft, and judge the output against what you would have produced without it before committing budget to Max, Team or Enterprise.
FAQ
Q1. What is Claude used for?
Claude is used for drafting and editing writing, summarizing long documents and research, generating and reviewing code, and running multi-step tasks such as compiling a report from several sources. Businesses also use it through Anthropic’s API to power features inside their own products, so many people use Claude without ever seeing its name.
Q2. How does Claude actually work?
Claude predicts likely next text based on patterns learned from training data, then applies Constitutional AI training so it follows instructions and avoids harmful outputs. Current models add long context windows up to 1 million tokens, optional “extended thinking” for harder problems, and direct tool use for web search, code execution and file access.
Q3. Who is Claude built for?
Claude is built for knowledge workers doing document-heavy reasoning, such as writers, researchers, lawyers and consultants, plus developers using Claude Code for software engineering. Anthropic also sells Team and Enterprise tiers for organizations that need SSO, audit logs and compliance controls layered on top of the same models.
Q4. What can Claude not do?
Claude cannot generate photorealistic images, does not perfectly recall everything inside very long contexts, and grows less reliable the longer an autonomous task runs unsupervised. It still hallucinates on open-ended questions more than on grounded ones, and its headline benchmark scores tend to overstate real-world performance on unfamiliar work.
Q5. Is Claude free to use?
Yes, Claude has a free tier covering chat, web search, memory and file creation, though message limits reset frequently and heavy users hit them fast. Paid plans start around $17 to $20 a month for an individual Pro subscription and scale up through Max, Team and Enterprise tiers for higher usage caps and admin features.
Q6. How is Claude different from the alternatives?
Claude distinguishes itself with long context windows, a strong coding reputation through Claude Code, and heavier emphasis on safety-focused training, while it lags ChatGPT and Gemini on native image generation and deep single-ecosystem integration. No one assistant wins every task, so the right choice depends on whether your priority is writing and code, images, or workspace integration.
Q7. What does the name Claude mean?
Claude is not an acronym. Anthropic named its models after Claude Shannon, the mathematician and engineer regarded as the founder of information theory, whose work underpins the statistical foundations modern language models are built on. The name signals a connection to rigorous, foundational computer science rather than any marketing wordplay.
