Claude capabilities

Adaptive thinking lets Claude dynamically decide, per request, whether and how much to use extended reasoning based on the complexity of the task, instead of the developer presetting a fixed thinking-token budget. It is the recommended replacement for extended thinking and is core to the newer Claude models. A developer can still steer how hard the model thinks via an effort level (low, medium, high, xhigh, max), but the model makes the in-the-moment call on reasoning depth. It is enabled by setting the thinking type to 'adaptive' on the Messages API.

Interactive, standalone content windows for code, documents, and applications generated alongside conversations.

Anthropic's flagship Opus-tier model, released 2026-05-28, for complex reasoning and long-horizon agentic coding. It is no longer the top of the lineup: since 2026-06-09 Claude Fable 5 sits above it, and claude.com/pricing now also lists a higher 'Mythos' tier above Fable (capabilities not yet publicly documented). It adds two user-facing controls: an effort setting that lets you trade tokens/latency for answer quality (default high; raise to extra/max for harder work), and a fast mode that returns up to about 2.5x more output tokens per second at premium pricing - and that premium fast pricing is roughly 3x cheaper than it was on previous Opus models. Anthropic reports strong agentic and coding benchmark results (e.g. Terminal-Bench 2.1, CursorBench, and 84% on Online-Mind2Web browser-agent evaluation).

Claude Science, launched in beta on 2026-06-30, is a dedicated desktop application that pairs Claude with a local analysis environment for researchers. It consolidates literature review, code execution, domain-specific specialist agents (genomics, single-cell, proteomics, structural biology, cheminformatics), 60+ scientific database connectors, and HPC/compute management into one workbench, producing versioned, fully-provenanced artifacts (figures, notebooks, manuscripts) with a background reviewer agent that flags citation errors and code-figure mismatches.

Browser-resident Claude agent (extension) that can read the current page, fill forms, click through workflows, and execute multi-step tasks on the user's behalf — Anthropic's productized Computer-Use surface for consumers.

Sandboxed Python and Bash environment (originally launched as the 'analysis tool') that lets Claude run code, analyze uploaded files, perform complex math, generate charts, and produce downloadable files (.xlsx, .csv, .pdf) directly inside a chat.

First-party and custom integrations that let Claude read from and act on external SaaS via the Model Context Protocol (MCP). Pre-built connectors cover Google Drive, Gmail, Calendar, SharePoint, Outlook, Slack, GitHub, Notion, Asana, Linear, and more; custom remote MCP servers can be added per user or org.

Generate a public link to a snapshot of a Claude chat so anyone with the URL can view the messages and any Artifacts up to the moment of sharing.

Claude has no single 'GPTs/Gems' SKU. The functional equivalent is a composite of Projects (instructions + knowledge base + isolated chat history), Skills (reusable instruction packs Claude auto-invokes), and Styles (output-format presets) — combined they let a user package a persistent specialized assistant.

Multi-agent research mode where a lead Claude agent decomposes a question, spawns parallel sub-agents to search the web and connected sources, and synthesizes a cited report. Anthropic markets this product simply as 'Research' inside Claude.ai.

Attach files to a chat or load them into a Project's knowledge base. Claude can read text, parse PDFs (including images and charts inside them), analyze spreadsheets, and use uploaded files as input to the code-execution sandbox.

Persistent memory across conversations and searchable chat history.

The Message Batches API lets developers submit large volumes of Messages requests as a single asynchronous batch instead of one synchronous call at a time, and pay 50% less for it. You upload many independent requests at once, the batch is processed in the background (most finish in under an hour), and you poll for status and download the results when processing has ended. It is built for non-interactive bulk work where an immediate response is not needed, such as large-scale evaluations, classification, content moderation, and dataset analysis. The 50% batch discount stacks with prompt caching, so combining both on requests that share a long common prefix can cut costs further.

Native iOS and Android apps plus a Windows/macOS desktop app and embeddings into Slack, Microsoft 365, and Outlook — Claude is a multi-surface product, not just claude.ai.

Dropdown in the chat composer to switch between the Claude tiers — Haiku (fast/cheap), Sonnet (daily-driver balanced), Opus (deep reasoning), Fable (long-horizon autonomous work), and Mythos (invitation-only top tier, available through Project Glasswing) — at their current generation.

Six-tier table (Free, Pro, Max 5x, Max 20x, Team, Enterprise) governing model access, usage budget, and admin features. Verified directly from claude.com/pricing on 2026-07-04.

Workspaces that group conversations with shared instructions, knowledge files, and project-scoped memory.

Lets a Claude API request reuse a large, unchanging prompt prefix (tool definitions, system instructions, and earlier messages) instead of reprocessing it on every call. Reading from the cache costs about 10 percent of the normal input price and is faster, so for workloads that repeat the same long context, prompt caching can cut input cost and latency by up to roughly 90 percent. Caches are isolated per organization, and on the Claude API they are also isolated per workspace, so cached content is never shared across tenants.

Packaged instructions, scripts, and resources that extend Claude into a specialized agent for specific tasks.

Persistent profile information that Claude uses to personalize responses across all conversations.

Image input — Claude can read and reason over images (charts, screenshots, handwritten notes, photos, documents) and discuss them in chat or visualize derived data inside Artifacts. Claude does NOT generate raster images natively; it produces SVG, HTML/canvas, or chart code instead.

Real-time spoken conversation with Claude — speak prompts and hear synthesized voice responses, with seamless switching back to text in the same thread.

Search the web for current information and conduct multi-step research within conversations.