Anthropic has turned Claude 4.6 into a broader product stack rather than a single chatbot. The lineup now spans Opus 4.6, Sonnet 4.6, and Haiku 4.5, along with four usage modes—Chat, Cowork, Code, and Projects—and a set of workflow layers such as Skills, Agents, Connectors, and Plugins.
Model tiers and pricing trade-offs
Claude Opus 4.6 launched on February 5, 2026 with a 1 million token context window. Anthropic said it scored 78.3% on MRCR v2 under long-context evaluation and can sustain tasks for up to 14.5 hours. API pricing was listed at $5 per million input tokens and $25 per million output tokens, with a maximum output of 128K tokens. The model is positioned for large-scale context analysis, codebase refactoring, deep research, and high-stakes deliverables.
Sonnet 4.6 arrived on February 17 and serves as the default choice for most users. It also supports a 1 million token context window. According to Anthropic’s early testing, about 70% of users preferred Sonnet 4.6, and in 59% of scenarios it outperformed the earlier flagship Opus 4.5. API pricing was set at $3 input and $15 output per million tokens, with a maximum output of 64K tokens. Anthropic said it runs about 30% to 50% faster than 4.5.
Haiku 4.5 is aimed at high-concurrency and lower-cost use cases such as API pipelines or subagent work. The source also makes a clear warning: Haiku has no prompt injection protection, which raises risk when it handles untrusted input.
Long context and multimodal expansion
A major pricing change took effect on March 13. Requests above 200K tokens had previously carried a premium, but that surcharge was removed. After the update, 900K-token and 9K-token requests were billed at the same rate, and no beta header was required. The article frames this as a shift that lets users load an entire codebase, full legal contracts, large datasets, or months of records into a single working context.
Multimodal limits were also raised to 600 images or PDF pages per request, up from 100. The capability is available through Claude Platform, Microsoft Foundry, and Google Cloud Vertex AI.
Four modes for four kinds of work
Claude now separates usage into Chat, Cowork, Code, and Projects. Chat is built for quick questions, brainstorming, and draft writing. Cowork is the desktop agent layer that can read and edit local files while carrying out multi-step tasks. Code runs in the terminal and is focused on repositories, commands, and Git workflows. Projects is designed for persistent context, which makes it useful for recurring tasks such as weekly reports, newsletters, and client deliverables.
On memory, Claude opened chat-history-based Memory to all users, including free accounts, as of March 2, 2026. Users can review, edit, or delete saved memory in Settings > Capabilities, and also import or export the full memory set. Incognito chats are excluded from memory writes.
Cowork as the center of the desktop agent workflow
Cowork entered research preview on macOS on January 12 for Claude Max users, expanded to Pro on January 16, then to Team and Enterprise on January 23, and later added Windows support. The article describes Cowork as a task delegation layer: users describe the outcome they want, while Claude breaks the work into steps and executes it inside a real desktop environment.
The piece places heavy emphasis on setup rather than long prompts. It recommends creating a dedicated workspace folder instead of granting access to an entire Documents directory, then building a CONTEXT folder with three Markdown files—about-me.md, brand-voice.md, and working-preferences.md—to define role, writing style, and execution rules.
Another recurring tactic is to begin prompts with “Start by using AskUserQuestion”. That shifts the interaction from manually crafting the perfect prompt to letting Claude generate a structured question flow before execution starts.
Connectors, plugins, and scheduled tasks
On February 24, Claude Cowork added Connectors for Google Drive, Gmail, DocuSign, FactSet, Google Calendar, Slack, and other tools. The article says there are now 50+ integrations, accessible through Settings > Connectors, and that the feature is free for all users. Once connected, Claude can search files, combine information from multiple sources, draft emails, and review contracts for risk markers.
Plugins and the Marketplace launched the same day. Plugins package skills, slash commands, and connectors into role-specific toolsets for functions such as sales, marketing, legal, finance, data analysis, product, customer support, engineering, HR, operations, design, branding, and life sciences research.
Scheduled Tasks went live on February 25. Users can set recurring jobs once and let Claude run them on a schedule, including email digests, weekly metrics summaries, and competitive research, as long as the computer stays on and Claude Desktop is running.
Dispatch, Projects, and Computer Use
Dispatch, released on March 17 for Pro and Max users, links mobile and desktop task management for Cowork. One operational detail in the source stands out: the Dispatch scheduling layer does not read CLAUDE.md by default, so users are advised to add “read CLAUDE.md” directly in the instruction.
Projects followed on March 20, giving users persistent workspaces with separate files, links, instructions, and memory per project. Then on March 23, Anthropic introduced Computer Use in research preview for macOS, available in both Cowork and Claude Code. Claude can click, type, move through interfaces, open apps, fill forms, and operate local tools. Anthropic also warned against using the mode for sensitive information and flagged screen-content-based prompt injection as a key risk.
Claude’s Chrome extension adds another layer. It can read webpages, click elements, fill forms, and support browser-side workflow replay and frontend debugging alongside Claude Code. For Team and Enterprise accounts, admins can manage allowed and blocked sites at the organization level.