This guide documents Claude AI features as they stand in July 2026, from the 2026 model lineup – Fable 5, Mythos 5, Opus 4.8, Sonnet 5, and Haiku 4.5 – to the workspaces, agentic tools, and integrations built on top of them.
For each feature you get what it does, when it launched, how it works, which tiers receive it, and the documented limits. One note up front: Claude does not natively generate image, audio, or video. It takes text and image input and returns text.
This page covers each major feature: what it is, when it launched, how it works, which tiers receive it, and the documented limits. It is a neutral Claude reference, not a product pitch. For plans and tier pricing, see the Claude pricing guide. For how Claude compares with other assistants, see Claude vs ChatGPT, Gemini, Grok, and Perplexity.
The 2026 lineup runs across five models. Fable 5 and Mythos 5 share one architecture at the top, Opus 4.8 is the flagship for coding and enterprise work, Sonnet 5 balances speed and cost, and Haiku 4.5 is the fastest. All models except Haiku 4.5 carry a 1M token context window and 128K max output.
API notes: batches and deprecations. Opus 4.8, Opus 4.7, Opus 4.6, and Sonnet 5 can return up to 300,000 tokens through the Message Batches API using the output-300k-2026-03-24 beta header. Batch processing runs asynchronously at roughly 50% of synchronous cost. Separately, the original Sonnet 4 and Opus 4 snapshots are retired from the API. Migrate pinned calls to claude-sonnet-5 or claude-opus-4-8.
Fable 5 and Mythos 5 access. Fable 5 launched in June 2026, was suspended within days after a jailbreak finding, and was redeployed in July 2026 with updated safety classifiers. When a classifier fires on Fable 5 (cyber, dangerous biology or chemistry, or model distillation), the request falls back to Opus 4.8. Mythos 5 access stays limited to Project Glasswing partners.
What Claude does not do. No native image, audio, or video generation. Claude accepts text and image input and returns text. Third-party integrations can pair Claude with image models, but that is not a native Claude capability. This is a design choice, not a roadmap gap.
Available on: Free, Pro, Max, Team, Enterprise, API.
Extended Thinking, introduced with Claude 3.7 Sonnet on 2025-02-24, forced Claude to generate a visible chain-of-thought before answering. The developer set a budget_tokens parameter to control reasoning compute. Adaptive Reasoning, introduced with the 4.6 generation in February 2026, replaced that model. Claude now evaluates problem complexity internally and decides whether and how much to reason. The developer sets an effort level (standard, high, xhigh, max) instead of a token budget.
The xhigh level, introduced with Opus 4.7, sits between high and max and gives more compute for hard tasks without committing to maximum spend. Claude Code defaults to xhigh on all plans since the Opus 4.7 release. Opus 4.8 defaults to high, which Anthropic describes as the balance point between quality and response time. Effort control in claude.ai and Cowork rolled to all plans with the Opus 4.8 launch, shown as a dropdown next to the model selector.
Adaptive Reasoning turns on interleaved thinking automatically, so the model can think, call a tool, read the result, think again, and continue. That is the structural reason it fits agentic work. On Fable 5 and Mythos 5, adaptive thinking is the only mode. The raw chain-of-thought is never returned. Set thinking.display to summarized for a readable summary or omitted for none.
Breaking change for Opus 4.7 and later. Manual Extended Thinking through budget_tokens is deprecated for Opus 4.7 and later and for Sonnet 5. Passing it returns a 400 error. Sonnet 4.6 still accepts both approaches during the transition.
Thinking summaries on claude.ai. On the claude.ai surface, thinking summaries are shown for transparency, generated by a smaller model for about 5% of long thought processes per Anthropic documentation. On Fable 5, a summary appears when thinking.display is set to summarized.
See also: Claude vs ChatGPT comparison → Suprmind Multi-Model Divergence Index →
Available on: Projects – Free (limited), Pro (unlimited), Max, Team, Enterprise.
Claude Projects create isolated workspaces where you upload reference documents and standing instructions that persist across conversations. Claude reasons over project content by retrieval, pulling relevant sections into active context rather than loading the whole project at once. Project content is cached and does not count against per-message usage limits. Per-chat file upload is capped at 20 files, 30 MB each, on every tier.
As a project’s files approach the model’s context limit (200K or 1M tokens), Projects switches to retrieval mode automatically. Claude then searches the project and pulls only the passages it needs, which raises effective capacity by up to about 10x without any manual setup. Team and Enterprise users can share projects with view or edit permissions.
Claude Artifacts render code, documents, diagrams, and interactive content in a side panel. When Claude produces standalone content (code, HTML, SVG, Mermaid diagrams, React components, or formatted Markdown), a live preview opens next to the chat. As of June 2026 you can edit an artifact in place: highlight the part you want changed, type the instruction, and Claude edits inline.
Each artifact supports up to 20 MB of persistent storage. State can be personal, isolated to the viewer, or shared, where every viewer acts on the same state, which suits forms, trackers, and small tools. When you publish an artifact, collaborators can fork it and change their own copy without touching the original.
Known limitation: context saturation. Project retrieval pulls only the most relevant content per query, so full retrieval is not guaranteed within a single response. As a conversation grows, the window can fill and leave little room for the exchange itself. One mitigation is to summarize the session state into a single anchor message before the window fills.
Available on: GA on Pro, Max, Team, Enterprise.
Claude Cowork packages the autonomous work of Claude Code into a visual interface for knowledge workers. Rather than printing text to copy and paste, Cowork manipulates files directly: it structures folders, builds presentations, and writes spreadsheets with working formulas. It grants Claude access to a user-specified folder, and it supports multi-step tasks and sub-agent coordination for parallel work. The research preview launched in January 2026 on macOS for Max users. General availability on macOS and Windows landed in April 2026.
As of July 2026, Cowork runs on web and mobile in addition to desktop. Sessions can run remotely in beta, with files and session state saved to your Claude account and synced across devices. Work continues when you close your laptop, and scheduled tasks run server-side with no device online. Chat and Cowork now share one home for Projects and Artifacts across both surfaces.
Cowork runs in two modes. Remote execution is the default: the agent loop and code run inside isolated cloud sandboxes on Anthropic infrastructure, with no outside network access unless an admin allowlists it, and file access limited to folders you authorize. Local execution runs shell commands and Python inside a Linux virtual machine on your own hardware, which needs hardware virtualization and can use up to 25 GB of disk and 8 GB of RAM.
File modification risk. Users have reported Claude changing files without prior review. Back up any folder that holds important files before you grant Cowork access to it.
Token cost. A single Cowork task involves multi-step reasoning, file operations, and sub-agent coordination, so it can use the token equivalent of many regular chats. Auto mode runs parallel safety classifiers and uses more tokens than step-by-step approval. Users report that heavy daily Cowork use draws down Pro allocations quickly.
Available on: Max, Team, Enterprise, SDK.
Claude Code is Anthropic’s terminal-first agentic coding tool, generally available since 2025-05-22 (research preview 2025-02-24). It runs Claude as a coding agent that searches a codebase, edits files, runs tests, and manages git. It maps structural dependencies with semantic search, so you do not hand-pick context files. Native integrations include VS Code and JetBrains extensions with inline edits, GitHub and GitLab pull requests, and the Claude Code SDK. Claude Sonnet 5 became the default model in Claude Code on July 1, 2026.
Install with a native script (curl -fsSL https://claude.ai/install.sh | bash) on macOS and Linux, PowerShell on Windows, or a package manager like Homebrew or WinGet. The older NPM distribution is deprecated. A CLAUDE.md file in the project root sets standing instructions (architecture decisions, coding standards, preferred libraries) read at the start of every session. Lifecycle hooks run linters or formatters after an edit, and an auto-memory keeps build commands and fixes across sessions without manual setup.
For larger tasks, a lead agent breaks a feature request into parts, dispatches them to parallel background agents, and merges the results. You watch that parallel state with the claude agents command. Dynamic Workflows (research preview with Opus 4.8, May 28, 2026) extends this to tens or hundreds of parallel sub-agents in one session, which enables codebase-scale migrations across hundreds of thousands of lines. Available on Max, Team, and Enterprise.
Reported at the Opus 4.7 release (April 2026). Numbers stay attributed to the version that achieved them. Opus 4.8 is the current flagship for coding.
Reported telemetry: automated Code Review. Anthropic’s Code Review is billed on token use. Reported figures put it at $15 to $25 per pull request, and on pull requests over 1,000 lines it surfaces about 7.5 issues at a roughly 84% finding rate. These are reported figures, not guaranteed outcomes. Running review in a fresh session, separate from the one that wrote the code, reduces the chance the model validates its own flawed logic.
The Claude got dumber postmortem (April 23, 2026). Anthropic confirmed three separate causes across March and April 2026. Default reasoning effort changed from high to medium on 2026-03-04 and was reverted 2026-04-07. A caching bug cleared thinking history on stale sessions and was fixed 2026-04-10. A system prompt verbosity constraint on 2026-04-16 caused a 3% eval drop and was reverted 2026-04-20. The intentional degradation claim was unsubstantiated. A viral BridgeMind benchmark claiming a 15-point drop was based on n=6 tasks. An independent retest at n=30 showed negligible movement, from 87.6% to 85.4%.
Billing note and Pro inclusion. As of June 15, 2026, non-interactive use (GitHub Actions, headless claude -p, third-party agents via the Agent SDK) draws from a dedicated monthly credit sized to each plan and does not roll over. Interactive terminal sessions and claude.ai chat are unaffected. Pro-tier inclusion of Claude Code is contested: the pricing page has listed it under Pro, while an independent changelog reported it removed from Pro in April 2026. Max, Team, and Enterprise are confirmed, and SDK access is uniform.
See also: Claude Code pricing details →
Available on: API, Enterprise.
Claude Managed Agents is hosted infrastructure for stateful, long-running agents, so a team does not have to build the execution loop, tool error handling, and session state itself. It sits alongside the stateless Messages API, which needs a custom loop. Managed Agents runs on Anthropic’s managed cloud or on self-hosted sandboxes. This is a compact overview, not a full API tutorial.
Cloud sandboxes are isolated Ubuntu 22.04 containers with up to 8 GB of RAM and 10 GB of disk, pre-loaded with Python 3.12, Node 20, Git, and common database clients. Network access is off by default and can be opened to an explicit allowlist. Self-hosted sandboxes run the same work queue on your own Linux hosts for data-residency needs, so sensitive data does not leave your infrastructure.
Managed agents can hold memory across sessions through the agent-memory beta. Each store keeps up to 2,000 memories, with individual files capped at 100 kB. A store can be attached read-only so untrusted external data cannot write into an agent’s long-term memory. A consolidation pass, which Anthropic calls dreaming, can dedupe and reorganize a store into a cleaner one.
Where dreaming applies. The dreaming and consolidation idea belongs to Managed Agents memory stores. It is separate from chat memory and the file-system /memory folder covered in the Memory section below.
Available on: Pro (research preview), Max, API (Messages API).
Computer Use lets Claude operate a graphical interface. It reads screenshots, moves the mouse, types, and inspects elements with a zoom action. It shipped as beta with Claude 3.5 Sonnet on 2024-10-22 and reached general availability on claude.ai in March 2026. Developers pass the computer-use tools through the Messages API. Claude returns tool-use requests (a stop reason of tool_use), the client runs them in a sandboxed VM with an X11 display, and results come back as tool_result blocks. The loop runs until the task finishes or hits an iteration limit (default 10, adjustable).
Sonnet 5 and Opus 4.8 accept screenshots up to 2,576 pixels on the long edge. Older models cap at 1,568 pixels, about 1.15 megapixels. Because the API downscales oversized images, the client has to rescale the coordinates Claude returns back to real screen dimensions, which matters most on high-density displays.
Computer Use has no application-level sandbox, so classifiers scan each screenshot for prompt-injection. If one is detected, the model is steered to stop and ask for explicit confirmation before it continues.
Reported benchmarks (Opus 4.7, April 2026). Opus 4.7 raised Computer Use reliability with high-resolution vision, scoring 98.5% on XBOW’s visual-acuity benchmark, up from 54.5% on Opus 4.6, and 78% on OSWorld, with GPT-5.5 at 78.7%. These figures stay attributed to Opus 4.7.
Setup. API-level Computer Use needs a sandboxed VM with a lightweight desktop (Mutter and Tint2 are recommended). It is embedded in the Messages API, not a standalone endpoint. Cowork computer use needs no setup, since Claude uses your computer directly through the desktop app or, as of July 2026, through a remote web or mobile session.
Available on: Free, Pro, Max, Team, Enterprise, and the /v1/skills API endpoint.
Skills are file-system folders that hold a required SKILL.md plus optional scripts and resources. Claude scans available skills at session start, loads only minimal metadata first, then loads more files only if the skill is relevant to the task. That progressive disclosure keeps context small. Skills are composable, so Claude coordinates several of them on its own, and they run across the Claude app, Claude Code, and the API through the /v1/skills endpoint.
The initial release was October 15, 2025. Skills 2.0, full workflow packages with executable scripts, shipped in Q1 2026. Anthropic ships pre-built Skills for Excel, PowerPoint, Word, and PDF work. Skill Creator, updated March 2026, added a test-measure-refine loop: create a skill, run a test suite, measure it against benchmarks, and iterate. It treats skills as versioned assets with evaluations rather than prompt tinkering.
Agent Skills open standard (December 2025). Skills are an open standard that works across AI platforms, not only Claude. A platform that implements the Agent Skills spec can use Claude-authored skills, which allows cross-platform reuse. Team and Enterprise organizations can deploy Skills org-wide.
Available on: Chat memory – Free (since March 2026), Pro, Max, Team, Enterprise.
Memory runs in two modes. Chat memory derives summaries of past conversations and carries them across sessions, viewable and editable at Settings → Capabilities → Memory. File-system memory for agentic use writes to a /memory folder, read at session start, with an optional auto-memory mode that lets Claude decide what to store. Opus 4.7 improved file-system memory reliability for long multi-session work.
Chat memory reached Team and Enterprise in September 2025, Pro and Max in October 2025, and Free in March 2026. A separate August 2025 data policy change extended conversation data retention to five years for users not opted out of training, which is distinct from active memory. Memory can be turned off at Settings → Capabilities, and training data can be opted out at Settings → Privacy → Data Usage.
Context window compaction (November 2025) summarizes earlier messages when a chat nears its limit, which enables very long conversations. Claude Code on Opus 4.8 handles compaction automatically in agentic sessions, and API users have a beta compaction feature. A Monthly Recap at Settings → Reflect (July 2026 beta) shows the topics you spent time on and how you work, and it needs memory turned on.
Available on: Remote connectors – Pro, Max, Team, Enterprise. Local MCP on Claude Desktop – any plan with the Desktop app.
MCP is an open standard Anthropic designed so Claude can connect to external tools, data sources, and services through one interface. It launched November 18, 2024. One-click local installation on Claude Desktop landed in June 2025, and remote MCP connectors landed in January 2026. MCP servers expose tools Claude can call (file access, database queries, API calls) with per-action approval in desktop mode. Third-party servers exist for Notion, Zapier, GitHub, and major IDE tools.
Opus 4.7 scored 77.3% on MCP-Atlas at its April 2026 release, ahead of Gemini 3.1 Pro at 73.9% and GPT-5.4 at 68.1%, the top score reported on that tool-orchestration benchmark at the time. Research Mode extended MCP support in February 2026 so it can connect to any MCP server for enterprise data without custom API plumbing.
No published hard limit. Anthropic has not published hard limits on MCP server count or tool calls per session. Setup complexity for local servers, which involves editing a config file, is the documented friction point.
Files attach directly to chat messages for reference within the context window, while project knowledge gives persistent cross-session access through retrieval. Accepted formats include PDF, text files (.txt, .md), and code files, plus images (PNG, JPEG, GIF, WEBP) for vision-enabled models, Office formats through Skills, and CSV or structured data through the code execution tool.
Every tier caps a chat at 20 files and 30 MB per file. With Opus 4.6 and later on the API, a single request supports up to 600 images or PDF pages. Enterprise plans give a 500K context window in chat, other plans give 200K in chat, and Opus and Sonnet reach 1M on the API. Claude 3.5 and later read PDFs including embedded images.
The Code Execution tool runs in a gVisor-isolated sandbox with a Python read-eval-print loop and bash. State persists across requests in the same container, so a later step can build on an earlier one. The sandbox has no outside network access by default. When Code Execution runs together with the native Web Search or Web Fetch tools, Anthropic does not charge for the code compute, since it trims web data before it fills the context window.
Available on: Web Search – Free, Pro, Max, Team, Enterprise, and the Web Search API at $10 per 1,000 searches. Research Mode – Pro, Max, Team, Enterprise.
Web Search has been a toggle in claude.ai across all tiers since May 2025 and is available through the Web Search API at $10 per 1,000 searches. With it on, Claude queries the web in real time and cites URLs inline. With it off, responses draw from parametric knowledge with a training cutoff around January 2026 for current models.
Research Mode is an agentic research feature that combines web search, Google Workspace access, and connected integrations into multi-source reports. It launched in April 2025 with Google Workspace, and mobile plus advanced mode followed in May 2025. As of February 2026 it can connect to any MCP server for enterprise data.
Web search cannot read paywalled content, private accounts, deleted content, content blocked from Claude-SearchBot in robots.txt, or content from sanctioned jurisdictions. One documented gap: within a single response, the interface does not mark which claims came from web search versus parametric knowledge.
Available on: Pro, Max, Team, Enterprise.
Claude for Excel launched in beta in October 2025 and was upgraded in February 2026 to run native Excel operations like pivot-table edits and conditional formatting. Claude for PowerPoint launched in February 2026. A March 2026 update let the Excel and PowerPoint add-ins share full conversation context, so an action in one application is informed by what happened in the other, and Skills work inside both add-ins.
As of July 2026 the Microsoft 365 connector gained write tools. Claude can draft, send, and organize email, manage calendar events, update mailbox settings, and create or update files in OneDrive and SharePoint. Read and search tools work as before, and Teams stays read-only. A Microsoft Entra administrator has to consent to the updated permissions before write tools turn on for an organization.
Available on: Pro, Max, Team, Enterprise.
Claude Design turns mockups and live sites into editable vector graphics and front-end code. It reads an uploaded design or a URL through Claude’s vision models and generates matching components you can edit, which shortens the handoff between design and engineering.
Before it shows a component, Design checks the generated code against imported design-system rules (typography, spacing tokens, and brand colors) and corrects deviations in the background, similar to a review loop before output. It launched from Anthropic Labs in April 2026.
Documented critique. Two caveats are on record. Early outputs tended toward a uniform Claude aesthetic (teal gradients, serif type, and stacked pills and cards) unless steered with custom tokens. And Design usage now counts against the global account quota, so heavy design work draws down the same limit as chat and coding.
Available on: Pro, Max, Team, Enterprise.
Claude Science is an AI workbench for life-sciences research. It brings notebooks, terminal access, literature search, and high-performance computing into one artifact-first interface, and it launched July 1, 2026.
It ships with more than 60 pre-built skills and connectors for genomics, proteomics, structural biology, and cheminformatics, and it connects to NVIDIA’s BioNeMo toolkit for models like Evo 2, Boltz-2, and OpenFold3. It renders 3D protein structures and genome-browser tracks directly in the interface.
Every figure, sequence, or draft carries an immutable run history: the exact script, the environment, and the conversation that produced it. For scale, it schedules SLURM jobs across remote GPU clusters over SSH, so researchers spend time on hypotheses rather than server upkeep.
Available on: Team, Enterprise.
The claude.ai consumer interface does not show a real-time, per-message indicator of which underlying snapshot handled a given query. The model selector shows the choice, and system-prompt probing reveals the dated snapshot ID, but the persistent interface does not. Default model transitions, for example Sonnet 4.6 to Sonnet 5 as the Free and Pro default on July 1, 2026, are announced through the Anthropic newsroom rather than an in-product notice for existing users.
Developers who call API model IDs like claude-opus-4-8 or claude-sonnet-5 receive the pinned snapshot tied to that ID at call time. The 4.6 generation introduced dateless API IDs that look like aliases but are pinned snapshots, not evergreen pointers. On Fable 5, when a safety classifier fires, the request is served by Opus 4.8 and the response carries a fallback indicator, so users are told when a fallback happens.
See also: Claude vs ChatGPT comparison → Claude pricing details →
FAQ
It depends on the job. Fable 5 sits at the top for long-running agent work and holds context across long sessions. Opus 4.8 is the flagship for complex agentic coding and enterprise work. Sonnet 5 balances speed and cost and is the default for Free and Pro. Haiku 4.5 is the fastest for high-volume, latency-sensitive tasks. Mythos 5 is a restricted, invite-only variant for defensive security.
Extended Thinking (Claude 3.7 Sonnet, 2025-02-24) allocated a visible pre-response reasoning budget set with budget_tokens. Adaptive Reasoning (4.6 generation, February 2026) replaced it: the developer sets an effort level (standard, high, xhigh, max) and Claude allocates compute internally. Manual budget_tokens is disabled for Opus 4.7 and later and for Sonnet 5, and returns a 400 error. On Fable 5 and Mythos 5, adaptive thinking is always on and the raw chain-of-thought is never returned, only an optional summary.
Claude Code is a terminal-first tool for developers. It runs in your terminal, reads your codebase, writes code, runs tests, and manages git. Claude Cowork is the knowledge-worker agent. It grants Claude access to a folder (or a remote session as of July 2026), reads and writes files, and runs multi-step tasks across applications like documents, spreadsheets, presentations, and email. As of July 2026, Cowork sessions can run remotely, continuing server-side when you close your laptop and reachable from phone, web, or desktop.
Fable 5 and Mythos 5 are the same underlying model. They share capabilities, a 1M token context window, a 128K output limit, and $10/$50 per Mtok pricing. The difference is the safeguards. Fable 5 runs safety classifiers that intercept requests around cybersecurity offense, dangerous biology or chemistry, and model distillation, and it falls back to Opus 4.8 for those. Mythos 5 runs without the cyber classifiers for defensive security work and is available only to approved partners through Project Glasswing.
Projects group related conversations, uploaded files, and custom instructions under a persistent context that all chats in the project can reach. File uploads are capped at 20 files per chat at 30 MB each, and project content is cached and does not count against per-message usage limits. Free has Projects with limits, and Pro, Max, Team, and Enterprise have unlimited Projects. As a project’s files approach the context limit, retrieval mode raises effective capacity by up to about 10x.
MCP is an open standard Anthropic designed so Claude can connect to external tools, data sources, and services through one interface. Third-party servers exist for Notion, Zapier, GitHub, and major IDE tools. Remote connectors are on Pro, Max, Team, and Enterprise, and local MCP through Claude Desktop works on any plan with the desktop app. Opus 4.7 scored 77.3% on MCP-Atlas at its April 2026 release. As of February 2026, Research Mode can connect to any MCP server.
Yes. Web search is a toggle in claude.ai across all tiers since May 2025 and is available through the Web Search API at $10 per 1,000 searches. With it on, Claude queries the web in real time and cites URLs inline. With it off, responses draw from parametric knowledge with a training cutoff around January 2026 for current models. Research Mode (Pro and above) is the full agentic version that combines web search, Google Workspace, and MCP connectors into multi-source reports.
Memory runs in two modes. Chat memory derives summaries of past conversations and carries them across sessions, viewable and editable at Settings → Capabilities → Memory, and it has been available to all users including Free since March 2026. File-system memory for agentic use (Claude Code, Cowork) writes notes to a /memory folder read at session start. Memory can be turned off in Settings, and training-data use can be opted out at Settings → Privacy → Data Usage.
No. The full 2026 lineup (Fable 5, Mythos 5, Opus 4.8, Sonnet 5, and Haiku 4.5) does not generate images, audio, or video. Accepted inputs are text and image. Third-party integrations can pair Claude with image models, but that is not a native Claude capability, and it is a design choice. Claude Design creates visual outputs like layouts and slides, but it does so by generating code and HTML, not by generating images directly.
As a conversation nears the context-window limit, the oldest content is gradually displaced. The symptoms (forgotten formatting rules, re-asked questions, contradictory answers from partial recall) are mechanical context overflow, not forgetting. Context compaction (November 2025) summarizes earlier messages near the limit, and Claude Code on Opus 4.8 handles this automatically in agentic sessions. For manual sessions, summarize the state into a single anchor message before the window fills.
Sources
Last verified July 2026.
Suprmind runs Claude in the same conversation as Grok, GPT, and Gemini, and when one of them makes something up, the others catch it.
7-day free trial. Four AI models. No credit card required.
