Anthropic has released the Claude 4.6 series, consisting of three models covering complete functionalities such as Cowork collaboration tasks, Claude Code development system, Skills and Agents mechanisms, etc. This article organizes the most practical AI workflow guide up to March 2026 based on the logic of “different identities, different scenarios.” The source is a post by @kloss_xyz on X, edited and translated by Dongqu.
(Background: Senior engineer: I’m tired of talking about AI unless you’re actually building something valuable and cool)
(Additional background: Claude Code ultimate quick reference: hotkeys, Slash commands, Skills, Agents, MCP full operation tips)
Editor’s note: When we look back at Claude’s product evolution in 2026, a clear change emerges: the question is no longer “what can it do,” but “how should different people use it.”
This article systematically reviews Claude’s capabilities and usage methods based on Anthropic’s product updates since 2026. It’s organized around the logic of “what different users should use and how in what scenarios.” Think of it as a navigation guide: when facing specific tasks, quickly locate the relevant module and call the appropriate capabilities.
For first-time users of Claude, it’s essential to understand the model and basic abilities, including context window, model layering, and four usage modes. These factors jointly determine Claude’s capability boundaries and form the foundation for subsequent usage.
For knowledge workers, focus on the task execution system represented by Cowork. How to build workspaces, structure files, set global commands, and reconstruct interaction via AskUserQuestion determines whether you are “using AI” or “letting AI do the work.”
For developers, the core pathway unfolds through Claude Code. The key is no longer just coding but building a reusable, collaborative development system via mechanisms like CLAUDE.md, Rules, Commands, Skills, and Agents, making Claude part of the software production process.
On a more concrete application level, from data analysis and presentation in Excel and PowerPoint, to APIs, automation workflows, and visualization capabilities, Claude is gradually embedding into traditional software systems, becoming part of the underlying capabilities.
As AI shifts from “dialogue tool” to “work system,” the real difference no longer comes from the model itself but from how you use it.
Anthropic’s recent product update pace has become astonishing, making it hard even for deep users to keep up. Nearly daily releases, with major updates roughly every two weeks since January this year. New models, tools, integrations, and even entirely new product categories are continuously launched. If you get distracted or take a few weeks off, you might miss many key changes. And indeed, Claude is reshaping your work style—undeniably.
This is a “panoramic guide.” Up to March 23, 2026, all major features launched for Claude are covered here: how to set each, when to use them, and best practices for effectiveness. Distinguishing these differences is the key to telling apart “something cool” from “truly transforming your work.”
You’ll likely want to bookmark this and review repeatedly. Feel free to share with your team or friends. This is the reference manual I wished I had when I first started.
The Claude 4.6 series currently has three model tiers. Below are each model’s capability boundaries and suitable scenarios:
Claude Opus 4.6 is the current performance ceiling. Released on Feb 5, 2026, supporting a 1 million token context window (details after price adjustment). Under this 1 million token long context, MRCR v2 score is 78.3%, the highest among models at this level.
Leads comprehensively in legal, financial, and programming tasks. Anthropic reports its task execution capability can reach 14.5 hours, the longest among frontier models. API pricing is $5 per million tokens input, $25 output, with max output up to 128K tokens. Supports adaptive reasoning, with a new “max” level for extreme capabilities.
Note: MRCR v2 score measures the model’s ability to find correct information in “super-long contexts.”
· Suitable scenarios (Opus): complex large-scale context analysis, codebase refactoring, deep research, high-risk delivery, serious content production, and any “quality over cost” tasks.
· Not suitable (Opus): workflows requiring high-frequency calls. At current prices, heavy Opus use might cost $50–100 daily. Default to Sonnet first; upgrade to Opus only if Sonnet output quality is insufficient.
Claude Sonnet 4.6 released on Feb 17, just 12 days after Opus, is the default choice for most users. Also supports 1 million token context (available from Mar 13). Improvements in coding, computation, long context reasoning, Agent planning, knowledge work, and design. Early tests show about 70% of users prefer Sonnet 4.6 (vs 4.5), outperforming the previous flagship Opus 4.5 in 59% of scenarios.
Default model for free and Pro users on claude.ai. API costs $3 / $15 per million tokens, max output 64K tokens, speed improved by 30–50% over 4.5.
· Suitable scenarios (Sonnet): daily work, quick drafts, routine programming, Agent workflows—balancing speed and intelligence. In many office scenarios, it already approaches or surpasses Opus (some tasks even outperform Anthropic’s OfficeQA benchmark), with about 40% lower cost.
Claude Haiku 4.5 is a low-cost, ultra-fast model designed for high concurrency scenarios, mainly used for API pipelines or sub-agent tasks like read-only processing.
But a key premise: Haiku has no prompt injection protection. If used in an Agent system with untrusted inputs, risks must be carefully evaluated and official documentation read thoroughly.
Previously, requests over 200K tokens paid a premium (up to $10 / $37.5 per million tokens). Since Mar 13, this premium is eliminated. Now, 900K tokens and 9K tokens cost the same. No multipliers, no hidden conditions, no beta headers needed.
What does this mean? About 750,000 words of context can be loaded at once: entire codebases, full legal contracts, large datasets, months of files—all stored in the same “working memory.”
Simultaneously, multimodal capabilities are enhanced: up to 600 images or PDF pages per request (previously 100, a 6x increase). Available now on Claude Platform, Microsoft Foundry, and Google Cloud Vertex AI.
For teams, this change is straightforward: previously, large content had to be chunked, summarized, and managed with rolling contexts; now, it can be loaded directly. Some companies report that increasing context from 200K to 500K tokens actually reduces total token consumption, as the model no longer repeatedly reads and reprocesses historical info.
Claude offers four modes, but most users only use one:
Chat
Familiar browser/mobile interface. Suitable for questions, brainstorming, drafts.
Each conversation starts from scratch; you always lead.
Cowork
Desktop Agent. Can read and modify local files, automate multi-step tasks, output results to folders.
Ideal for “delegating tasks,” not back-and-forth dialogue.
Code
Developer mode, runs in terminal. Can access code repositories, write code, execute commands, manage Git.
If coding, this is the most leverage.
Projects
Persistent workspace. Upload files and commands once; each new conversation carries full context.
Suitable for repetitive tasks like weekly reports, newsletters, client deliverables.
Simple rule:
Ask questions quickly in Chat, delegate work in Cowork, develop in Code, repeat work in Projects.
As of Mar 2, 2026, Claude has enabled conversation-based memory for all users (including free). Claude extracts relevant context from dialogues and generates a cross-session memory summary. View, edit, or delete in Settings > Capabilities. Supports import/export of full memory data—useful for backups or account migration. Incognito mode content is not saved.
Key operation: check “what Claude remembers” in Settings > Memory. Correct inaccuracies, add background info. The more accurate, the less you need to re-explain in future sessions.
Note: Cowork sessions do not inherit memory across conversations. To maintain continuity, use “file structures” (see Limitations below).
Cowork fundamentally changes the game:
Launched on Jan 12 in macOS (for Claude Max users), then expanded to Pro on Jan 16, Team and Enterprise on Jan 23, and Windows later. Market response was immediate—investors quickly realized what this meant: SaaS market cap evaporated by billions in days, Wall Street understood this path.
But the key: don’t treat it as just a chat interface.
Cowork is task delegation.
Describe “what the result should look like,” and Claude will plan, break down sub-tasks, autonomously execute in your real environment, and deliver final files to your folder. You can leave; when you return, work is done.
Anthropic built Cowork with only about 10 days of using Claude Code.
Those who misuse Cowork often stick to old habits: writing lengthy, detailed prompts for each task, resulting in instability.
The truly skilled do this: spend an afternoon setting up the “context environment” (files, global commands, folder structure), then use just 10 words to produce deliverables directly client-ready.
Underlying logic:
ChatGPT trains you to write better prompts
Cowork rewards you for building better “file systems”
The former is a skill that depreciates as models evolve; the latter is a compound, continuously improving ability.
Step 1: Build your workspace folder
Create a dedicated folder for Cowork on your computer.
Don’t point it directly to the entire Documents directory. If issues arise (possible), keep scope minimal. Cowork needs real read/write permissions for the authorized folder.
This maintains clarity and limits Claude’s access. Most experienced users’ setups converge on similar structures. Folder names don’t matter; key is proper layering and isolation.
Step 2: Build your file structure
This is key to solving “AI output homogenization.” In your CONTEXT folder, create three Markdown files:
about-me.md
Defines your role and current focus—not a resume, but your actual work, target audience, priorities, and high-value items. Can include 1-2 achievements as standards.
brand-voice.md
Solidifies your expression style: tone, common/disallowed words, formatting preferences, 2-3 writing samples. Differentiates “generic AI content” from “personal style output.”
working-preferences.md
Specifies Claude’s execution norms: ask clarifying questions first, output task breakdowns, avoid deletions without confirmation, default formats, quality standards, and behaviors to avoid.
These files solve the “cold start” problem: without context, each task begins from zero. After setup, Claude starts each session with full knowledge of your style, standards, and preferences.
These files have a “compound interest” effect. Weekly iteration recommended. When Claude’s output is off, first check: prompt issue or context issue? Usually context. Fix by adding rules in the relevant file for long-term correction.
In practice, this setup is low-cost: I spent about 45 minutes creating the initial context folder—three .md files defining “who I am,” “what I do,” and “Claude’s execution style.” With this, a single 10-word project prompt can produce expected results on first try, instead of explaining everything each time.
User comment: “Claude Cowork is also very useful for file handling and editing. Just describe what file you want (e.g., ‘a video with a squirrel’), give simple commands, and Claude can call ffmpeg to process. Even without experience in file editing or format conversion, you can do it smoothly.”
Step 3: Set global instructions
Go to Settings > Cowork > Edit Global Instructions.
Global instructions load before everything else—before your files, prompts, even before Claude reads your folder. They define the “bottom-layer behavior” for every session.
Sample template:
{
"instructions": "Always prioritize reading the files in the context folder. Follow the rules in CLAUDE.md. Clarify ambiguous tasks before proceeding. Do not perform destructive actions without confirmation. Maintain a professional tone. Use markdown formatting for outputs."
}
This ensures even casual prompts produce calibrated results. Claude always knows who you are, prioritizes correct files, confirms before judgment, and only the prompt handles specific tasks.
Step 4: Learn to use AskUserQuestion
This feature fundamentally changes interaction: no longer you craft “perfect prompts,” but Claude asks “perfect questions.” Add “Start by using AskUserQuestion” in your prompt, and Cowork will generate an interactive form: multiple-choice, options, clear paths, structured questions to clarify needs before execution.
Result: no need to craft lengthy prompts upfront; Claude actively determines what info it needs. If the first round still misses the mark, you can point out the issue, and it will generate a new round of questions, iterating.
A universal prompt template:
Start by asking me questions to clarify the task. Use multiple-choice and structured questions. Confirm understanding before proceeding.
Simple. With your file structure, this covers about 80% of scenarios. Workflow remains consistent; only the prompt text varies.
Connectors
Launch date: Feb 24.
Claude Cowork now supports connecting to Google Drive, Gmail, DocuSign, FactSet, Google Calendar, Slack, and more, with enterprise updates.
These are not shallow integrations. Claude can autonomously:
· Search and browse files in your Drive
· Extract and consolidate data from multiple sources
· Automatically draft emails based on retrieved info
· Scan contracts and flag risks
Once connected, Claude can access real-time data from these tools in each session—no copy-paste, screenshots, or manual downloads.
Setup: Settings > Connectors, browse 50+ integrations, click “Add,” authorize.
This only needs to be done once. Connectors are free for all users (including free tier from Feb 24). Still underrated.
Typical uses:
· Slack: “Retrieve my messages from last 7 days, summarize follow-ups, prioritize by urgency.”
· Google Drive: “Find latest files related to project X, read and highlight 3 key points.”
· Google Calendar: “Review this week’s schedule, identify conflicts, draft reschedule emails for lowest priority meeting.”
Plugins and Marketplace
Launch date: Feb 24.
Plugins are prebuilt modules for specific roles, packaging skills, commands, and connectors as “role-based toolkits.” Anthropic released official plugins covering sales, marketing, legal, finance, data analysis, product management, support, enterprise search, engineering, HR, operations, design, branding, life sciences, etc.
Installation: Customize > Browse plugins, click install; type “/” in chat to see commands.
Recommended plugins:
· Productivity: manage tasks, schedules, daily workflows. Type /productivity:start, Claude organizes your day.
· Data Analysis: upload CSV, /data:explore, Claude analyzes columns, detects anomalies, suggests insights, generates SQL.
Choose role plugins matching your work:
/marketing:draft-content — generate content in brand tone
/sales:call-prep — research clients, generate talking points
/legal:review — review contracts, flag risks
Team users can build private plugin markets, distribute custom plugins internally, with admin controls (Team/Enterprise). One build, scalable deployment.
Anthropic also launched public plugin marketplace and Ambassador program, supporting community development, ecosystem expanding fast.
Plugins can be personalized: after installation, tell Claude “Customize this plugin for my company,” and it will ask about workflows, terminology, preferences, embedding these as long-term context.
This means a general sales plugin can evolve into a tailored tool understanding your ICP, pricing, and communication style.
Scheduled Tasks
Launch date: Feb 25.
Set once, Claude can run tasks periodically, e.g.:
· Daily morning email summaries
· Weekly data reports
· Regular competitive analysis
Assumes your computer is on and Claude Desktop is running.
A verified real use case:
Wake up Monday morning to a ready-made presentation. With connectors, scheduled tasks can truly “auto-execute”: e.g., “Every Monday, fetch unread Slack messages from #product-feedback, categorize by topic, generate summary in Google Drive.” Scheduled trigger, connector pulls real-time data, Claude processes, results saved automatically.
I run 3–4 scheduled tasks daily: morning AI news briefing, noon competitor updates, afternoon social media summaries, evening performance review.
Each saves 20–30 minutes manual work, adding up to nearly 2 hours daily, with minimal management.
This feature now integrates with Claude Desktop’s new Customize module, unifying skills, plugins, and connectors.
Dispatch
Launch date: Mar 17.
A bridge for remote management of Cowork tasks via mobile, now for Pro and Max users. Use Claude Desktop or iOS/Android app to remotely control tasks.
Setup: In Desktop, go to Cowork > Dispatch, enable “Keep awake” (to prevent sleep interruption). On mobile, open Claude app, tap Dispatch.
Core experience: continuous cross-device conversation thread. On commute, command Claude via phone to handle desktop tasks—organize spreadsheets, generate reports; when arriving, work is done. Multiple tasks can be stacked in one Dispatch command, Claude executes sequentially during your absence.
A subtle point (from Product Compass): Dispatch schedule does not read CLAUDE.md; it generates prompts based on default assumptions. Subtasks read it, but initial instructions may be off.
Solutions:
· Explicitly add “read CLAUDE.md” in Dispatch commands.
Limitations:
· Cannot add connectors on mobile yet—must set up Gmail, Slack, Notion on desktop first; Dispatch inherits.
· Cannot upload files on mobile—send files to email, then use Gmail connector.
Overall, Dispatch extends local capabilities to any time and space. Not just remote control, but redefines task execution boundaries.
Projects
Launch date: Mar 20.
Organize related tasks into persistent workspaces, each with files, links, commands, and memory. Import existing folders or start fresh. Enables managing multiple projects: “Q1 Finance” and “Product Launch,” with separate contexts.
Purpose: upgrade Cowork from one-off Agent sessions to evolving workspaces. Critical for research-heavy tasks, avoiding context loss and repeated goal explanations.
Computer Use
Launch date: Mar 23.
In research preview, macOS only, for Pro/Max, available in Cowork and Claude Code.
Claude can now directly operate your computer: click, input, navigate, open apps, browse, fill forms, control local tools.
When official connectors (e.g., Slack, Google Calendar) exist, Claude prefers interface calls; otherwise, uses mouse+keyboard.
Operation & risks:
Claude requests permission before key actions. Anthropic advises against handling sensitive info here.
Main risk: prompt injection via screen content. Visiting untrusted sites can inject content into context, affecting behavior.
Recommendation: use only trusted apps/websites.
With Dispatch, capabilities expand: command Claude on mobile to perform desktop or browser tasks not yet connected via plugins.
This bridges a key boundary: from “calling tools” to “direct OS operations.”
Chrome with Claude
Chrome extension enables Claude to interact with your browser: read pages, click, fill forms, navigate.
Most overlook: you can demo workflows, teach Claude to reproduce them. Any repetitive weekly task can be recorded as a workflow.
Integration with Claude Code further streamlines dev: write code in terminal, test in browser. The extension can read console errors, network requests, DOM state—often pinpointing issues before you ask.
You can also control browser from Claude Desktop, no need to switch windows. Admins can manage allowed sites via whitelist/blacklist for teams.
Typical use: record “weekly competitor price check” as a workflow. Claude visits sites, scrapes prices, compares, saves in Cowork. Previously 45 min manual clicks, now one click.
Caution: restrict site permissions; webpage content is a major prompt injection vector. Limit to trusted sites.
Organize past months’ files:
Point Cowork to a folder with six months of scattered files—receipts, contracts, notes, screenshots.
Claude reads, classifies, renames by date, structures files, logs actions. 2 hours manual work now 10 minutes.
User example: organize 317 Disney World videos—Claude extracts GPS from metadata, determines park zones, auto-sorts into folders.
Lenny has it traverse all his podcasts, extract key insights like “most important product lessons” and “counterintuitive insights.” The process takes minutes, previously days or weeks. Read more.
Generate client deliverables from raw materials: meeting notes, transcripts, research links—compile into a structured, ready-to-submit report.
Claude reads all raw materials, consolidates into a formatted report, saves as deliverable. 90 min work now 15 min.
Weekly research briefs: set scheduled task, every Monday 7 am, Claude researches competitors, scans industry publications, generates formatted report. Review at your convenience. With connectors, fetch real-time data from Slack, Gmail, Drive.
Financial modeling: a creator asked Cowork to build a valuation model for social media exit. Claude devises plan, finds errors, delivers a “Wall Street style” Excel with 4 valuation methods, 129 formulas, covering revenue multiples, EBITDA, user/subscription value, 5-year DCF. Quite impressive.
Cowork consumes quickly: complex tasks may use 10x normal quota. Under $20/month Pro, daily use hits limits in days. Heavy users report hitting rate caps in 3–4 days, impacting critical tasks.
Multi-step tasks (file reading, generation, parallel subtasks) are compute-intensive. For main workflows, Max ($100/month, ~5x quota; or $200/month, ~20x) is more feasible. Use Settings > Usage to monitor.
Context compression in long sessions is also key: near limit, Claude auto-summarizes early content to free space. Maintains conversation but reduces info fidelity—values simplified, early decisions compressed.
If Claude begins giving “patterned” rather than specific file answers, compression has occurred. Solution: at key points, write important info into files for traceability.
Current stage: research preview. Anthropic notes models may misread files or take unnecessary detours on simple questions. About 10% of complex multi-step tasks deviate, with some inconsistency. Human review needed before external output.
Cross-session memory is absent: each new Cowork session is independent—no memory of who you are or prior discussions. Major friction point.
But with file structures, this can be mitigated:
· Write preferences into files
· Save project plans
· Write standard instructions
For continuity, embed it into files. The upside: structured workflows are portable, shareable, version-controlled.
Tasks depend on client-side execution. Cowork runs in Claude Desktop; closing window stops tasks. Better to let computer sleep than close app, so sessions persist.
Only desktop supported. No mobile Cowork or browser version; no cross-device sync (Dispatch helps partially). Store files in cloud sync folders (iCloud, Dropbox, OneDrive).
If Cowork is for knowledge workers, Claude Code targets developers.
Claude Code debuted Feb 2025 as CLI coding tool, now a platform for scheduling AI Agents across dev workflows, with annual revenue of $2.5 billion.
Setup: npm install -g @anthropic-ai/claude-code, go to project, run claude to start an agent with access to codebase.
Operations include: read/write files, run commands, web search, testing, commit.
Web version on claude.ai upgraded Feb, adding multi-repo sessions, better diffs, Git visualization, slash commands. Deepest capabilities still in terminal.
The real difference: not just coding, but extensibility—turning Claude Code into a configurable dev platform.
1. CLAUDE.md: Project-level instruction manual
On each session start, Claude reads CLAUDE.md, directly loaded into system prompt, always active. Content here guides behavior. Many ignore or overload it, reducing output quality. Too little or too much info harms performance—find the “threshold.”
What to write:
Focus on impactful content:
· Key commands: build, test, lint (bash commands)
· Core architecture decisions (e.g., “use Turborepo monorepo”)
· Non-obvious constraints (“TypeScript strict mode enabled, no implicit any”)
· Import standards, naming, error handling
· Module structure
Avoid:
· Configs better suited for linters/formatters
· Complete files available via links
· Lengthy theoretical explanations
Keep under 200 lines; longer reduces Claude’s ability to follow instructions due to attention competition.
Write “what” and “why”:
“Use TypeScript strict mode” is basic; “Use strict because we had bugs from implicit any” is better. Explains reasoning, helps Claude make better decisions in unhandled cases.
Update continuously:
Press “#” during work; Claude appends rules to CLAUDE.md. Repeated corrections signal rules should be written into the file. Over time, it becomes a “living document” reflecting actual codebase.
Good vs bad:
Bad CLAUDE.md is like a new hire’s onboarding file; good one is a work memo before memory loss.
2. CLAUDE.md layering
Most overlook: CLAUDE.md isn’t a single file but a layered structure merged at session start.
Managed Policy (organization-wide)
IT deployment, cannot override, applies company rules.
~/.claude/CLAUDE.md (global)
Personal preferences, cross-project, not versioned.
./CLAUDE.md (project)
Team-shared, version-controlled, applies to all members.
CLAUDE.local.md (local override)
Per-project personal tweaks, ignored in commits.
Conflicts resolved by hierarchy: local > project > global > organization.
Common team issue:
Developers put key rules only in local (~/.claude/CLAUDE.md), causing inconsistency when new members clone repo. Misdiagnosed as model randomness. Solution: all rules in project CLAUDE.md.
3. Rules Directory: modular rules
When CLAUDE.md grows large, split into .claude/rules/ folder.
Each Markdown file loads at session start with CLAUDE.md, enabling modular extension and cleaner main file.
Focus each file: API conventions, testing, etc. Clear responsibilities.
Path scope rules: add YAML glob patterns to limit rules to certain paths:
# in rules/api-conventions.md
glob: "src/api/**/*.md"
Applies only to matching files, reducing token use. Compared to directory-level CLAUDE.md, path-based rules are more flexible and efficient.
Different extension mechanisms:
-
Commands (.claude/commands/) are manual slash commands, e.g., review.md → /project:review. Can embed shell commands with !.
-
Skills (.claude/skills/) are automatic triggers based on task matching, no input needed. Defined with YAML headers, e.g.:
name: careful
effort: high
trigger: "if task involves sensitive data"
-
Agents (.claude/agents/) are sub-roles with independent system prompts, tool permissions, and model choices. Use tools like tools: [Read, Grep], restrict capabilities.
-
Plugins: packaged skills, commands, connectors, for specific roles, distributed via marketplace or private repo.
4. Headless & CI/CD
Claude Code supports non-interactive mode: -p parameter, for automation.
Output structured JSON (--output-format json) for parsing and PR comments.
Sample GitHub Actions:
on: pull_request
jobs:
review:
runs-on: ubuntu-latest
steps:
- name: Run Claude review
run: claude -p "review this diff" --output-format json
- name: Post comment
uses: actions/github-script@v4
with:
script: |
// parse JSON and post comment
Cost: ~15 min setup, pre-PR detection.
Use separate Claude instances for code review to avoid bias.
Claude Code security
Supports full codebase security audit. Unlike rule-based scanners, Claude uses semantic analysis, detecting complex vulnerabilities. Anthropic reports <5% false positives. In tests, found 500+ vulnerabilities in open-source projects, some years old. Red team mechanisms further reduce false alarms.
Voice Mode
Supports voice input for hands-free coding. Say “refactor this” or describe logic verbally, enabled via /voice.
Automated code review & PR workflows
Claude can analyze diffs, assess quality, flag issues, comment on PRs. Integrate with CI/CD for auto-generation, testing, change summaries, pre-merge checks.
Excel & PowerPoint integration
Now, with context sharing, analysis in Excel flows into PowerPoint for presentations, no info lost. Skills can run in plugins; enterprise users via cloud gateways.
Visuals in chat
Launched Mar 12 (Beta). Generate interactive charts, flowcharts, visuals in chat. Based on HTML/SVG, support hover/click, update dynamically.
-
Inline visuals: temporary, in conversation
-
Artifacts: persistent, shareable files
Use inline visuals for exploration, artifacts for deliverables.
Example: “Draw a certification flowchart,” Claude generates instantly, continue discussion seamlessly.
Inference & reasoning updates
Effort-based reasoning replaces previous token budget. Sonnet 4.6: set to “medium” for cost savings. Opus 4.6: new “max” mode for max performance, higher token use.
Token billing: $25/M for Opus, effort parameter controls reasoning cost.
Tools & structured outputs
Flowed tool calls now GA. Structured outputs also GA. Data residency options available (~1.1x price).
Web capabilities
Code execution free with web search/fetch. Dynamic filtering supported. Search/fetch GA, no beta needed. Critical for dev workflows.
API Skills
New, not widely used yet. Prebuilt skills for PowerPoint, Excel, Word, PDF. Custom skills via /v1/skills. Require code execution enabled. Replaces many custom tools.
Context compaction
When near context limit, system auto-summarizes history, preserving key info. With 1M window, trigger frequency drops.
Funding & valuation
On Feb 12, 2026, Anthropic raised $30B at $380B valuation, second largest VC deal after OpenAI’s $40B. Investors include GIC, Coatue, Microsoft, NVIDIA.
Annual revenue: $14B, 10x growth over 3 years. Claude Code alone: $2.5B/year, doubled since early year. Enterprise subscriptions quadrupled.
Top clients:
-
8 of top 10 Fortune 500 companies
-
500+ clients spending over $1M/year (vs 10 two years ago)
-
7x growth in clients over $100K/year last year
Enterprise contributes ~80% revenue, with direct online purchase options.
Infrastructure & certifications
Jan: HIPAA-compliant enterprise plan for healthcare data.
Feb 13: Enterprise Analytics API for organizational metrics.
Anthropic launched Claude Partner Network, investing $100M in training, co-marketing, tech support.
March 12: Claude Certified Architect (Foundations)—exam on Agent design, MCP, Claude Code, reliability.
Accenture plans to train 30,000 staff on certification. Anthropic Academy launched March 2, with 13 free courses, now 15, more coming.
Internal use & product iteration
~60% of Anthropic engineers rely on Claude (vs 28% a year ago).
Daily releases: 60–100 internal versions.
Cowork built in 10 days, entirely on Claude Code.
This feedback loop accelerates product iteration from monthly to weekly, then daily.
Platform commoditization
Six months ago, features required custom frameworks; now native. The moat is not infrastructure but taste, distribution, and what you build on top.
Leverage for builders
Extensible via Skills, sub-agents, Agent Teams, Hooks, Channels, MCP, plugins.
A finely tuned Claude Code with custom skills and scope agents is a different tool from a simple prompt in chat.
Understanding and configuring these layers as per your workflow yields continuous compound benefits.
For knowledge workers
Starting now, build file structures, set global commands, install plugins, prioritize AskUserQuestion, schedule tasks, connect via Dispatch, and expand with Computer Use.
For managers
Plugins and enterprise capabilities enable standardization across organizations—turning Claude into an operating system.
Pace will only accelerate.
Anthropic is building the next generation of tools with its own tools. Each model iteration accelerates the next, transforming the entire industry’s compute paradigm.
Understand this platform now, not next quarter or next month—right now.
If you’ve read this far, you’re ahead of 99.9%. Those who bookmark but rarely revisit will still treat Claude as just a chat tool; you won’t.
I’m not an engineer, just self-taught. I don’t claim to have all answers or perfect configurations. If someone says so, they’re probably misleading you. All content here is from daily practice—trial, error, and recording what works—so others don’t have to start from zero.
The key: get hands-on, experiment, even “tinker.” That’s the only way to learn.
If you find inaccuracies, omissions, or outdated info, point them out—I prefer to correct than propagate errors.
Thanks for reading.
[Original link]
