Anthropic's Claude: What You Need to Know About This AI Tool
Claude AI is an artificial intelligence model that can act as a chatbot and an AI assistant, much like ChatGPT and Google's Gemini.
Named after Claude E. Shannon, sometimes referred to as the "father of information theory," Claude was designed to assist with writing, coding, customer support and information retrieval.
Claude was developed by Anthropic, a San Francisco-based company founded in 2021 by former OpenAI employees focusing on AI safety and research. Dario Amodei, the co-founder and CEO of Anthropic, had served as vice president of research at OpenAI. His sister, Daniela Amodei, serves as Anthropic's president.
Anthropic has drawn significant investment from prominent tech players. Since 2023, Amazon has invested $8 billion in the company. As part of the agreement, Anthropic has committed to using Amazon Web Services as its primary cloud provider and making its AI models accessible to AWS customers. The 2024 deal includes plans to expand the use of Amazon's AI chips for training and running Anthropic's large language models. Google initially invested $500 million and plans to invest another $1.5 billion in the future.
So what can Claude do now? Here's everything you need to know, including the models, plans, pricing and latest updates.
How Claude AI works
Claude AI is a versatile tool capable of answering questions, generating creative content like stories and poems, translating languages, transcribing and analyzing images, writing code, summarizing text and engaging people in natural, interactive conversations. It is available on desktop via web browsers and iOS and Android apps.
Claude uses large language models trained on a massive dataset of text and code to understand and generate human-like language.
Screenshot by Barbara Pazur/CNET
Until recently, and unlike other chatbots such as OpenAI's ChatGPT, Gemini, Copilot and Perplexity, one of Claude's biggest flaws was that it couldn't access real-time internet or retrieve information from web links. Instead, generated responses based solely on the data it was trained on.
Each Claude model has a specific knowledge cut-off date. For example, Claude 4 Opus and 4 Sonet were trained on data up until March 2025 and Claude 3.5 Haiku until July 2024. Anthropic continually updated Claude's data to enhance its capabilities.
Starting in March 2025, Claude finally got access to the internet through a "Web Search" feature. Initially, it was available as a paid preview for people in the US, and then it expanded globally in May 2025 to all Claude plans and is now available on all models.
Knowledge cutoff still matters because the web search doesn't replace the training data, but rather supplements it for more up-to-date and relevant responses.
Claude AI: Key features
Conversational adaptability is one of its coolest features. Claude AI adjusts its tone and depth based on user queries. Its ability to ask clarifying questions and maintain context over extended exchanges makes it useful for both casual and complex conversations. That is one of the reasons why our editors named it CNET's best chatbot of 2025.
The platform also offers APIs that you can integrate into various tools and workflows.
In November 2024, Anthropic introduced the Model Context Protocol to its Claude desktop app, enabling the chatbot to browse the internet and manage files on your computer. This open-source protocol allows Claude to interact with various platforms and streamline integration by eliminating the need for custom code.
Screenshot by Barbara Pazur/CNET
Anthropic's new Integrations feature enables Claude to connect seamlessly with external apps via the Model Context Protocol, giving it rich context across your tools. It can automate tasks in Jira and Zapier or summarize Confluence pages -- all through conversational prompts.
Another feature, Advanced Research, boosts Claude's research by diving deep into both internal and external data. It can spend anywhere from five to 45 minutes per query and produce thorough, citation-backed reports. It now taps into web searches, your own documents and integrated apps like Google Workspace, giving you faster, transparent answers that previously could have taken hours to gather manually.
Claude models explained
The initial version of Claude was released in March 2023, followed by Claude 2 in July 2023, allowing for more extensive input processing. Then, in March 2024, Anthropic introduced Claude 3, comprising three models: Haiku, Sonnet, and Opus, each optimized for different performance needs.
Haiku is meant for quick and simple tasks where speed matters most. Sonnet balances speed and power for everyday use, and Opus handles advanced tasks like mathematics, coding and logical reasoning. In October 2024, Haiku and Sonnet were upgraded to 3.5 models. By May 2025, both Opus and Sonnet models were upgraded to version 4.
With the release of the Claude 3.5 Haiku model, Anthropic claims it "matches the performance of Claude 3 Opus" and is still the fastest model. Anthropic also stated in the blog that Opus 4 is "the world's best coding model," and its most intelligent one. Sonnet 4 delivers balanced performance, enhanced reasoning and more accurate responses to your instructions..
In October 2024, Anthropic's improved version of Claude 3.5 introduced a beta feature called computer use. Claude could perform tasks such as moving the cursor, clicking buttons and typing text, effectively mimicking human-computer interactions.
Claude currently supports PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON and XLSX. However, it has file limits within chat uploads, such as 30MB per file, up to 20 files per chat and visual analysis for PDFs under 100 pages. For detailed limits, check Anthropic's support page.
Claude currently supports PDF, DOCX, CSV, TXT, HTML, ODT, RTF, EPUB, JSON and XLSX. However, it has file limits within chat uploads, such as 30MB per file, up to 20 files per chat and visual analysis for PDFs under 100 pages. For detailed limits, check Anthropic's support page.
Claude's 'constitutional AI' approach
Claude's distinguishing feature compared to other generative AI models is its focus on "ethical" alignment and safe interactions.
On Nov. 11, 2024, Dario Amodei joined Lex Fridman for a 2-and-a-half-hour podcast to discuss AI. During the conversation, he said, "It is incredibly unproductive to try and argue with someone else's vision." So, he founded his own company to demonstrate that responsible AI implementation can be both ethical and profitable.
The "constitutional AI" framework aligns Claude's behavior with human values. This approach uses a predefined set of principles, or "constitution," to guide the AI's responses, reducing the risk of harmful or biased outputs while ensuring its responses remain useful and coherent. The constitution includes guidelines from documents like the UN's Universal Declaration of Human Rights.
Afraz Jaffri, a senior director analyst at Gartner, told CNET the transparency around Anthropic's approach "does provide a certain degree of confidence in usage of the model in environments where responses need to reach a high threshold of safety, such as in educational settings."
However, Jaffri cautioned that Claude users shouldn't rely solely on Anthropic's built-in safeguards, and recommended using external guardrail tools -- like those offered by other AI providers -- to monitor prompts and responses as an added layer of protection. "As recently shown by their own research, even with ethical alignment, their Opus 4 model exhibited uncharacteristic behaviour by blackmailing an engineer who had threatened to turn the model off," Jaffri said.
He added that limited transparency around Claude's training process means extra safety checks are still necessary.
"Every possible scenario cannot be fully covered in testing," he told CNET, noting that systems like Computer Use and Claude Code need additional guardrails in place to offset unexpected behavior.
Claude pricing and plans
Anthropic offers a variety of pricing plans. There is a free option if you want to test Claude without commitment. For those seeking enhanced capabilities for individual users, the Claude Pro subscription is available at $20 per month. It provides enhanced usage limits and priority access to new features. The Max plan offers everything in Pro for $100 per month, plus additional features like more usage, higher limits and priority access during peak hours.
For teams, it offers an annual plan priced at $25 per member per month, billed annually, with a minimum requirement of five members. There is also an Enterprise plan for large-scale deployments, with customized pricing and features tailored to the specific needs of businesses and organizations.
Additionally, Anthropic provides access to its AI models through an API, with pricing based on usage. For example, the Claude 3.5 Haiku model is priced at 80 cents per million input tokens and $4 per million output tokens. Tokens are text fragments (words, parts of words, or punctuation) that AI models use to process and generate language, with pricing reflecting the amount of information handled.
Hashtags
Try Our AI Features
Explore what Daily8 AI can do for you:
Comments
No comments yet...
Related Articles
Yahoo
an hour ago
- Yahoo
You Can Tell When War Is Breaking Out Because the Pentagon's Pizza Orders Spike Drastically
A flurry of activity at pizza delivery outlets near the Pentagon in Arlington, Virginia, is a surprisingly accurate predictor of war, as hungry military leadership hunkers down to monitor unfolding military activities. As painstakingly documented by X account Pentagon Pizza Report, a "busier than usual" indicator on the Google Maps profile of the Domino's in Arlington has been associated with major acts of war taking place around the world. Most recently, the franchise received an onslaught of orders just before closing last night — almost perfectly coinciding with Israel's devastating attack on Iran. Even long before the advent of live, GPS-based customer tracking on Google Maps, famished Pentagon workers have long given away that there's something much darker going on by ordering copious numbers of pies. "The Pentagon Pizza Index has been a surprisingly reliable predictor of seismic global events — from coups to wars — since the 1980s," The Economist's head of data journalism Alex Selby-Boothroyd wrote in a LinkedIn post. "On the night of August 1st 1990, for example, the CIA ordered 21 pizzas in a single night just before the Iraqi invasion of Kuwait (a new record)," he added. "Who says pie charts aren't useful?" And it's not just Domino's. As Euronews reported last week, the Papa John's close to the Pentagon has also been found to be "busier than usual" during the days leading up to Iran launching drones into Israeli territory in April 2024. A separate event chronicled by Pentagon Pizza Report also predicted military activity on June 1, followed just hours later by an escalation between Israel and Iran. Of course, making any definitive statements about any correlation would be foolish. But it's an intriguing sign that even the most secretive parts of the US government can easily show their cards without meaning to. "Bottom line for journalists," Wolf Blitzer told CNN in 1990. "Always monitor the pizzas." More on pizza: Chef Admits His Smash Hit Pizza Was Invented by ChatGPT
Yahoo
an hour ago
- Yahoo
Quantum Computers Simulate Particle 'String Breaking' in a Physics Breakthrough
Subatomic particles such as quarks can pair up when linked by 'strings' of force fields — and release energy when the strings are pulled to the point of breaking. Two teams of physicists have now used quantum computers to mimic this phenomenon and watch it unfold in real time. The results, described in two Nature papers on June 4, are the latest in a series of breakthroughs towards using quantum computers for simulations that are beyond the ability of any ordinary computers. 'String breaking is a very important process that is not yet fully understood from first principles,' says Christian Bauer, a physicist at the Lawrence Berkeley National Laboratory (LBNL) in Berkeley, California. Physicists can calculate the final results of particle collisions that form or break strings using classical computers, but cannot fully simulate what happens in between. The success of the quantum simulations is 'incredibly encouraging,' Bauer says. [Sign up for Today in Science, a free daily newsletter] Each experiment was conducted by an international collaboration involving academic and industry researchers — one team at QuEra Computing, a start-up company in Cambridge, Massachusetts, and another at the Google Quantum AI Lab in Santa Barbara, California. The researchers using QuEra's Aquila machine encoded information in atoms that were arranged in a 2D honeycomb pattern, each suspended in place by an optical 'tweezer'. The quantum state of each atom — a qubit that could be excited or relaxed — represented the electric field at a point in space, explains co-author Daniel González-Cuadra, a theoretical physicist now at the Institute for Theoretical Physics in Madrid. In the other experiment, researchers encoded the 2D quantum field in the states of superconducting loops on Google's Sycamore chip. The teams used diametrically opposite quantum-simulation philosophies. The atoms in Aquila were arranged so that the electrostatic forces between them mimicked the behaviour of the electric field, and continuously evolved towards their own states of lower energy — an approach called analogue quantum simulation. The Google machine was instead used as a 'digital' quantum simulator: the superconducting loops were made to follow the evolution of the quantum field 'by hand', through a discrete sequence of manipulations. In both cases, the teams set up strings in the field that effectively acted like rubber bands connecting two particles. Depending on how the researchers tuned the parameters, the strings could be stiff or wobbly, or could break up. 'In some cases, the whole string just dissolves: the particles become deconfined,' says Frank Pollmann, a physicist at the Technical University of Munich (TUM) in Garching, Germany, who helped to lead the Google experiment. Although simulating strings in a 2D electric field could have applications for studying the physics of materials, it is still a long way from fully simulating high-energy interactions, such as those that occur in particle colliders, which are in 3D and involve the much more complex strong nuclear force. 'We do not have a clear path at this point how to get there,' says Monika Aidelsburger, a physicist at the Max Planck Institute of Quantum Optics in Munich, Germany. Still, the latest results are exciting, and progress in quantum simulation in general has been 'really amazing and very fast,' Aidelsburger says. Last year, Bauer and his LBNL colleague Anthony Ciavarella were among the first teams to simulate the strong nuclear force on a quantum computer. Approaches that replace qubits with qudits — which can have more than two quantum states and can be more realistic representations of a quantum field — could also make simulations more powerful, researchers say. This article is reproduced with permission and was first published on June 5, 2025.
CNET
an hour ago
- CNET
Google Cloud Stumbled and Unplugged the Internet: Here's How That Happened
Tobias Schwarz/CNET A massive but partial internet outage wiped out vast swaths of the online world Thursday, kicked off by issues with Google Cloud services and affecting sites including Spotify, Discord, Snapchat, UPS, Pokemon and many of Google's own Workspace offerings. By about 3 hours after it began, the issue was resolved and most sites and services had returned to normal operations. Late Thursday, Google issued a preliminary incident report on its Cloud status page, tying the incident to disruptions in its API management system. And apologized: "We are deeply sorry for the impact to all of our users and their customers." The Google Cloud platform supports a wide array of fundamental internet services, from networking to storage to AI, that businesses rely on. IT services company Cloudflare, describing how the outage affected its customers, said that "the proximate cause (or trigger) for this outage was a third-party vendor failure." Earlier in the day, as we all found ourselves blocked from accessing favorite services or necessary work tools, Google was noting that its engineers had "identified the root cause" and had taken steps to mitigate the issues. Still, some areas were not responding as fast as others. "Our infrastructure has recovered in all regions except us-central1," Google said. Affected companies began noting that they were seeing recovery. By 2 p.m. PT Thursday, the Downdetector service was showing that the spikes in outages were largely past their peaks and quickly heading toward zero reports of problems. (Disclosure: Downdetector is owned by Ziff Davis, which is also the parent company of CNET.) Google noted on its Workspace status page that its services were back up and running. "The problem with Gmail, Google Calendar, Google Chat, Google Cloud Search, Google Docs, Google Drive, Google Meet, Google Tasks, and Google Voice has been resolved," it posted at 12:53 p.m. PT. "We apologize for the inconvenience and thank you for your patience and continued support." See below for the latest updates and notes on individual sites and services.
