Claude, reviewed

Claude is Anthropic's family of models and the apps built around them. In a few weeks this autumn, Anthropic renewed its Opus, Sonnet and Haiku models, folded its agent workspace into the main app and lowered its prices. It is a good moment to look at Claude the way a company would: as a tool people use for real work every day.

We build on Claude and on its rivals, so this is a review from the bench, not from the press release. Where we give our own view, we say so.

What it is now

Since 16 September, Claude has two modes: chat and Claude Code. Cowork, the separate place Anthropic had built for bigger tasks, was merged into chat for Pro and Max subscribers first, with Team and Free to follow. Claude now decides how much work a request needs, and in Anthropic's words a bigger job goes on “even after you’ve closed your laptop.” Its explanation for the change was short: “So we stopped making you choose.” Claude Docs and Claude Slides arrived on the same day.

Claude Code is the agent for software. It reads a codebase, edits files, runs commands and opens pull requests, in the terminal, in VS Code and JetBrains, in the desktop app and in the browser. It can split a task across several agents and run routines on a schedule in the cloud.

Underneath sit four current models. Fable 5.1 is the one Anthropic recommends for the most demanding reasoning and long agentic work, at 10 dollars per million input tokens and 50 dollars per million output tokens. Opus 5.5, released on 22 September, is the one Anthropic tells developers to start with. It says Opus 5.5 performs at the level of Fable 5.1 on most work and costs 40 percent less to run than Opus 5, at 4 and 20 dollars. Sonnet 5.5 and Haiku 5.5 followed as faster and cheaper options. All four read up to one million tokens at once.

Connectors, skills and plugins

What turns Claude from a clever chat into a work tool is what it can reach. Connectors link it to a service through the Model Context Protocol, MCP, so it can read your files, calendar or issue tracker and act in them. Skills are written instructions, with scripts or reference files if needed, that Claude loads when a task calls for them: your offer format, your contract checklist, your weekly report. Plugins package both into one install that a team can share.

The directory now lists more than 950 MCP servers. For larger companies the more important change is quieter: admins can provision connectors for the whole organisation through their identity provider. In Anthropic's words, “Admins authorize a connector once, users inherit access through their existing IdP groups.” A research preview connects Claude to servers inside a private network without exposing them to the internet.

Where it is strong

Companies voted with their budgets early. Menlo Ventures estimated in December 2025 that Anthropic earned 40 percent of enterprise spending on large language models, up from 12 percent in 2023, and 54 percent of spending on coding.

Our view from daily use matches that. Claude writes calmly, keeps to long and detailed instructions, and holds a voice across a long document. Claude Code is at its best on real codebases, where it plans, changes many files and checks its own work. And because skills are plain text, the know how a team writes down once keeps working when the model underneath changes.

Where it falls short

In our experience, the limits are the sorest point. Paid plans run on a rolling five hour window plus weekly limits, and the plans are sold as multiples of Pro. A class action filed on 8 September argues that the Max plans do not make clear enough how those limits work, following a separate suit by a single subscriber in June. These are allegations, and Anthropic had not answered Engadget's request for comment when it reported the case. Opus 5.5 came with higher five hour limits on the Pro, Max, Team and seat based Enterprise plans.

Quality can also move under your feet. In April, Anthropic traced weeks of complaints about Claude Code to three product changes: a lower default reasoning effort, a bug that cleared Claude's earlier reasoning on every turn, and a prompt that capped how much it wrote. It fixed all three, reset usage limits for every subscriber and wrote: “We never intentionally degrade our models.” The candour was welcome. The lesson for a company is to test its own tasks after every change, not only after a new model.

Cost rises quickly at the top. A Team premium seat costs 125 dollars a month on monthly billing, and even there Fable is limited to half of the weekly allowance. Enterprise starts at 20 dollars a seat plus usage at API rates, so the bill grows with every model and task people run.

Lock in and cost

For one person, Pro costs 20 dollars a month, or 17 on annual billing, and includes Claude Code. Max costs 100 or 200 dollars for five or twenty times the usage of Pro. Team runs from 20 dollars a seat on annual billing, for 2 to 150 people, and training on company content is off by default.

In our view the lock in is lower than with most assistants. MCP is an open protocol and skills are text files, so the work a company puts into them can move. What does not move as easily are chat histories, projects and the habits a team builds around one app.

Our verdict

For writing, coding and agents that work through a task on their own, Claude is among the best tools a company can buy today. Plan for its limits, keep an eye on the bill at the higher tiers, and test after every change Anthropic ships.

Most of all, keep the knowledge outside the model. That is how we built Twin, a company's AI brain: its knowledge, rules, skills and integrations, delivered as source code and working in Claude, ChatGPT and Codex, or in a private model. If Claude stays ahead, the brain works in Claude. If something better comes, the brain goes with you.