Claude Best for writing & code
Anthropic's AI assistant. Excels at long-form writing, analysis, coding, and nuanced conversations.
Where Claude came from
Anthropic was founded in 2021 by Dario Amodei, Daniela Amodei, and several other senior researchers who left OpenAI. The split was philosophical: Dario and his team believed AI safety should be the central organizing principle of building large models, not a downstream consideration. OpenAI's increasing commercial focus didn't fit that vision. Anthropic was structured as a public benefit corporation from day one, with explicit governance to balance research mission against revenue pressure.
The funding scale tells the story: a $124M Series A in 2021, a $580M Series B (2022, led by Sam Bankman-Fried's FTX before that imploded), Google's $300M and then $2B investments, and Amazon's $4B (and later expanded to $8B) backing through 2024. By late 2025 Anthropic had raised more than $15B across all rounds. The company employs roughly 1,000 people — small compared to OpenAI's headcount but with an unusually high ratio of researchers to product engineers.
Claude as a product launched in March 2023, first as a waitlist-limited preview, then a public Pro plan later that year. Claude 2 in mid-2023 pushed context windows past competitors. Claude 3 in March 2024 (the Haiku/Sonnet/Opus tier-naming convention) made Claude broadly competitive on benchmarks. Claude 3.5 Sonnet in mid-2024 was the moment Claude pulled ahead on writing and code in head-to-head tests with GPT-4. The current Claude 4.7 Sonnet (released spring 2026) extends that lead.
The strategic positioning is clear: Anthropic isn't trying to be a platform. There's no Claude App Store, no plugin ecosystem, no built-in image generation. The product surface is intentionally narrow — chat, Artifacts, Projects, Computer Use (beta), and the API. The thesis is that being the best raw model wins more than being the broadest product.
What Claude actually is
Claude is a conversational AI assistant from Anthropic. The user-facing product is claude.ai (web), iOS and Android apps, and a Mac desktop app. The same models also power the API that thousands of products use under the hood — including Cursor, GitHub Copilot fallback options, Notion AI, and many others.
Major capabilities, as of mid-2026:
- Text generation — writing, analysis, coding, conversation. Multiple model tiers from fast (Haiku) to powerful (Opus).
- Vision — upload images, charts, documents; ask questions; extract data.
- Artifacts — side-by-side document panel for collaborative writing, code, diagrams.
- Projects — persistent knowledge files and instructions tied to a conversation thread.
- Computer Use (beta) — Claude operates your screen visually, clicks, types. Pro/Max only.
- 200K-token context — standard. 1M-token context on Enterprise plans.
What Claude deliberately doesn't have: image generation, voice mode (text-to-speech exists but no real-time voice yet), browser integration, plugin marketplace, web search built into responses. Claude's bet is depth over breadth. For users who need those features, ChatGPT or Gemini fills the gap.
First five minutes in action
Open claude.ai. Sign in with Google or email. The interface is intentionally clean — single chat panel, model picker top right, "Projects" in the sidebar. Type a question. Get a response. The first thing most users notice: Claude's responses are longer, more nuanced, and feel less like search results than ChatGPT's. It will often ask a clarifying question before answering, especially on ambiguous prompts. This trait is the most reliable predictor of long-term satisfaction with Claude.
Things to try in your first session:
- Paste a long email you're stuck on — Claude will draft a response with the right register
- Upload a chart or screenshot — Claude reads it accurately, including handwriting
- Ask "explain this code" with any source file pasted in
- Use Artifacts: ask Claude to write a document, then iterate. The side panel makes editing painless.
The thing that turns first-time users into power users: Claude's willingness to push back. Ask "should I use Redux for state management here?" and ChatGPT typically describes Redux. Claude asks what kind of state, who'll maintain the codebase, and whether your team knows the pattern. That's the writer's instinct — clarify before answering.
Artifacts and Projects — the killer features
Artifacts is Claude's standout product feature. When you ask Claude to write a document, code, or diagram of meaningful length, it opens a side panel. The document lives there, persistent. You can highlight specific paragraphs and ask "make this more concise" or "rewrite in active voice" — Claude edits just that part, without regenerating the whole thing. For long-form work, this changes everything.
Use Artifacts for:
- Documents over 500 words — essays, reports, README files
- Code projects — Claude generates runnable HTML/JS/React directly in the panel; you can iterate in place
- Interactive demos — Claude can write a working component you can click and test
- Charts and SVGs — Claude renders them live in the panel
Projects is the second feature most reviews underrate. Create a project, upload up to 5 knowledge files (PDFs, docs, code), add persistent instructions ("Always write in our brand voice — see voice-guide.pdf attached"). Every conversation in that project gets the context loaded automatically. Use Projects for:
- Long-running work where the context matters — book drafts, codebases, multi-week consulting engagements
- Team standards — upload your style guide once, every conversation gets it
- Customer support scripts where Claude needs to know your product
The model family
Claude Haiku — the fast and cheap one
Smallest model, fastest responses (sub-second typical), available on Free tier. Good for high-volume work, quick lookups, classification. Not for creative work — Haiku's writing feels notably more generic than Sonnet's. Use Haiku when you'd otherwise use a Google search.
Claude Sonnet — the daily driver
Best-in-class balance of quality and speed. Available on Free (with low limits), Pro, and higher. Claude 4.7 Sonnet (current) is what most Claude users interact with most of the time. Strong at code, writing, analysis, and reasoning. About 2-3 seconds per typical response. The model behind most "this is good" Claude moments.
Claude Opus — the heavy thinker
Largest model. Available on Pro tier with usage limits, unlimited on Max. Use for complex reasoning, hard code problems, nuanced writing where stakes matter. 5-10 seconds per response. Plus, Opus has stronger "I'm uncertain" calibration — it more reliably flags when it doesn't know.
The skill is matching task to model. Quick lookups and casual chat — Haiku or Sonnet. Daily writing/coding work — Sonnet. Hard problems where being right matters — Opus. Switching costs nothing. Power users routinely use all three within a single work session.
How it actually feels
The dominant feeling using Claude is that you're talking to someone competent. The model doesn't perform expertise the way GPT often does (lots of "Great question! Here are 7 things to consider..." enthusiasm). It just answers, clearly, with appropriate hedging where hedging is right. When it doesn't know, it says so. When the question is malformed, it asks a clarifying question instead of guessing.
For writing, Claude is the most consistently good model in our blind tests. Even when GPT-5 produces technically correct output, Claude's output tends to have voice — a discernible point of view, sentence rhythm, deliberate word choice. This matters for editorial work, marketing copy, fiction, anything where craft is the deliverable. For pure boilerplate (form letters, meeting summaries), the gap closes.
For code, Claude is excellent. It writes code that compiles on the first try more often than ChatGPT. It handles ambiguous requirements better — it'll generate code AND explain why it interpreted the request that way. The trade-off: Claude is more cautious. Asked to write security-sensitive code, it'll often add disclaimers and recommend you have a security review. Right call most of the time, occasionally annoying when you know what you're doing.
For analysis and reasoning, Claude is the model we trust on first reading. Its outputs are calibrated — confident when it should be, uncertain when uncertainty is honest. For high-stakes work (legal interpretation, financial analysis, technical decisions), Claude's hedging is the feature, not the bug.
Three real workflows, end-to-end
Drafting a 3000-word essay from a one-sentence brief
Brief: "Write a 3000-word essay arguing that engineering managers should code at least 4 hours per week, even at senior levels." Pasted that into Claude with one followup: "Use a personal-essay voice. First person. Include 3 specific examples of how this works in practice. Audience is engineering leaders at 100-1000 person companies."
Six minutes later, a complete essay. The argument was structured. The examples were specific and plausible. The voice had personality (not the bland LinkedIn-essay default). The conclusion didn't trail into platitudes.
We made three edits: tightened the intro, swapped one example for a personal anecdote, removed a hedging paragraph. Total edit time: 35 minutes. Total writing time from blank page would have been 4-5 hours. The output isn't done — we still added our own anecdote — but the structure and 80% of the prose was Claude's.
The thing GPT couldn't have done here: the voice consistency. GPT-5 drafts of the same brief had moments of greatness interspersed with phrases like "It's worth noting that..." that read as machine-written. Claude's draft sounded like a single human writing through.
Debugging a race condition in Postgres transactions
Pasted ~150 lines of TypeScript handling concurrent inventory updates. Described the symptom: "Two requests update the same row, we read the updated value, but sometimes the read returns the pre-update value. Happens roughly 1 in 200 requests."
Claude's response: walked through the transaction isolation levels in Postgres (Read Committed is default, our code assumed Repeatable Read), identified the specific lines where the race could happen, proposed two fixes (raise isolation to Serializable for those transactions, or add a SELECT FOR UPDATE on the inventory row). It also flagged a secondary issue we hadn't asked about: the retry logic on conflict was using exponential backoff but the initial timeout was too low for our DB latency.
Both fixes worked. The secondary issue was real and would have caused production problems later. Total time from paste to working code: 12 minutes. By myself, this would have been a half-day of reading Postgres docs and squinting at trace logs.
Becoming briefly competent on a 400-page technical spec
We needed to evaluate whether a feature we wanted to build was HIPAA-compliant. The full technical safeguards document is 400+ pages of legal-engineering hybrid prose. Reading it cold would have been a multi-day project.
Created a Claude Project, uploaded the HIPAA Security Rule PDF, added instructions: "You're a HIPAA technical reviewer. When I describe an architecture, identify HIPAA implications. Cite specific section numbers." Then described our proposed feature in 3 paragraphs.
Over an hour of back-and-forth: we got a structured analysis of every section that applied, specific risks the feature created, and a draft compliance plan. Each claim referenced an exact section of the spec, which we then verified by reading just those passages. The hour with Claude beat reading the full document by an order of magnitude — and we got coverage we'd have missed.
We then took the conclusions to actual security counsel. They confirmed 90% of the analysis, pushed back on one specific interpretation, and saved us the cost of paying them to read the spec from scratch.
Real prompt → real output
We asked Claude:
150 words, hit the brief, the closing line is the kind of unexpected human touch Claude does well. We'd publish this with one or two small tweaks.
Performance, measured
Across 100 representative prompts (50 writing, 30 code, 20 reasoning), Claude vs ChatGPT vs Gemini:
bench --task=all --metric=quality,speed,cost n=100
Claude wins writing and code. ChatGPT's o1 reasoning model edges out Claude on pure reasoning problems. Gemini trails on quality but wins on speed and free-tier generosity.
Privacy and safety — what actually happens
Anthropic's pitch is that safety is their organizing principle. The practical implication: Claude is the most conservative big model on what data gets used and how.
Key facts:
- Training data — by default, Claude does NOT use your conversations for training, even on consumer tiers. This is the opposite of ChatGPT's default. Opt-in is required to share data for training.
- Retention — conversations stored for 30 days by default for trust-and-safety review. Deleted permanently after that unless flagged. Pro and Team users can adjust.
- Compliance — SOC 2 Type II, HIPAA available on Enterprise plans, ISO 27001. The compliance bench is strong.
- Content policies — more conservative than ChatGPT on some topics (medical, legal advice tends to come with strong disclaimers). Less conservative than GPT on others (fictional violence, mature creative writing).
For enterprise data, the API is the path. Claude API holds zero-retention agreements possible on Enterprise tiers; many large companies route through Claude API specifically because of this posture.
Claude vs ChatGPT
a/claude b/chatgpt
The two leaders. They split the market: ChatGPT is the platform; Claude is the model. Most power users keep both.
claude wins at
- long-form writing quality and voice
- code quality (especially first-try compile rate)
- handling ambiguous prompts
- artifacts mode — side-by-side editing
- 200K context window standard, 1M on enterprise
- privacy default — no training on your data
chatgpt wins at
- voice mode — real-time conversational
- image generation (DALL-E 3 built in)
- code interpreter with Python sandbox
- custom GPTs ecosystem
- browser/plugin breadth
- polished mobile apps
Verdict: Use both at $20/mo each ($40 total). Claude for serious writing/code, ChatGPT for voice, images, voice mode on mobile. Power users do this. Half-power users pick one and miss out on the other half.
Claude vs Gemini
a/claude b/gemini
Google's Gemini is the price competitor (most generous free tier) and the Workspace integration play. Quality has caught up since 2024.
claude wins at
- writing quality and voice
- code reliability
- nuanced reasoning
- artifacts and projects features
- narrower scope = more polish
gemini wins at
- free tier generosity (Gemini 2.5 Pro free)
- Workspace integration (Gmail, Docs, Drive)
- 2M-token context window
- response speed
- image generation included free
Verdict: If you work in Google Workspace, Gemini's integrations earn their keep. Otherwise Claude's quality edge is worth the $20.
Claude vs Open-Source (Llama, Mistral)
a/claude b/open-source
Meta's Llama and Mistral's open-weight models close the quality gap each release. For organizations with self-hosting needs, this is a real alternative.
claude wins at
- quality on hard tasks
- writing voice (open models feel generic)
- zero infrastructure burden
- safety calibration out of the box
- reliability — no hosting outages on you
open-source wins at
- data sovereignty — runs on your hardware
- cost at scale (no per-token fees)
- customization via fine-tuning
- no content policy refusals
- compliance-friendly air-gapped deployments
Verdict: Claude for most use cases. Open-source when you have hard regulatory requirements, high-volume API use, or specific fine-tuning needs.
Where Claude gets it wrong
Refuses some things that aren't dangerous
Claude's content policies sometimes flag innocuous requests. Asking for help with a research paper on extremist movements can trigger a refusal. Workaround: explain the context. Claude responds well to "I'm a researcher, this is for an academic paper" framing.
No voice mode
For mobile use, voice is a major gap. ChatGPT's voice mode is the killer feature for thinking while walking. Claude on mobile is just chat. If voice matters, that's a real argument for the paid ChatGPT subscription.
No built-in image generation
You can't generate images in Claude. Period. You can describe images you want and Claude can write a prompt for Midjourney or DALL-E to execute elsewhere. Workable but a context-switch.
No real-time web search
Claude doesn't browse the web by default. If you need current information (news, prices, recent events), Claude will say "I don't have access to information past my training cutoff." Workaround: use the API + tool use to give Claude a search tool, or use Perplexity for current-info tasks.
Usage limits hit hard on Pro
Claude Pro at $20 hits message limits faster than ChatGPT Plus. Heavy daily users frequently bounce against the cap. Max tier ($100/$200) raises limits significantly; it's the right tier for power users.
Project knowledge files cap at 5
You can only attach 5 files per Project. For knowledge-heavy contexts (full books, large codebases), this is genuinely tight. Workaround: combine into one PDF, or use the API with bigger context.
Power-user tips
If you're going to have more than 3 conversations on a topic, make it a Project. Persistent instructions + knowledge files transform Claude from "generic assistant" into "domain expert who's read your stuff."
Don't regenerate documents — highlight the part you want changed and ask for the edit. Claude rewrites just that section. Massive time savings.
"Before writing this, ask me 3 questions you'd need answered to do it well." Claude often comes back with questions that reframe the brief entirely. Worth the extra prompt.
Upload your writing style guide as a Project file. Every future conversation in that project gets it. Cuts down on "make it more X" iterations dramatically.
Available on Max. Claude can drive your screen. Currently best for repetitive tasks where you describe the goal and Claude executes the clicks. Rough edges but real productivity for specific workflows.
On Mac, the keyboard shortcut opens Claude in a popup overlay. Faster than alt-tabbing to the browser.
If you find yourself running the same prompt daily — script it. Claude's API is fast to set up; a one-line cron + Python script can automate work you've been doing by hand.
Right-click any conversation in the sidebar → Pin. Pinned chats stay at top. Useful for ongoing projects you return to.
Common gotchas
- Free tier message limits are tight. About 30-40 messages per 4-hour window before forced wait. Plan for Pro within a week of regular use.
- Projects cap at 5 knowledge files. Combine related files into one PDF if you hit the limit.
- Claude doesn't browse the web. Use Perplexity for current-events queries; Claude for evergreen analysis.
- Computer Use is slow and rough. Each click takes a few seconds. Useful for batch tasks, not real-time work.
- Artifacts get cluttered. Old artifacts pile up in your chat. There's no centralized library yet.
- Mobile lacks Artifacts. Side-by-side editing is desktop-only.
- No image generation, period. Use ChatGPT or Midjourney for visuals.
- Aggressive defaults on hedging. If you want a direct answer with no caveats, prompt explicitly: "Skip caveats. Give me the direct answer."
Pricing, in real terms
Free: Limited messages, Sonnet model, basic features. Use for evaluation.
Pro ($20/mo): Generous (but not unlimited) usage on Sonnet and Opus, Projects, Artifacts. The default for individuals.
Max ($100 or $200/mo): 5x or 20x Pro limits. Worth it if you hit Pro limits daily. The $200 tier is the heavy-user plan.
Team ($25/user/mo): Centralized billing, no training on your data, shared workspaces. Good for 3-20 person teams.
Enterprise: Custom pricing, SSO/SAML, audit logs, dedicated capacity, 1M-token context. For serious deployments.
Total cost of ownership
Solo, 1 year
- Pro: $20 × 12 = $240
- Time savings on writing alone: 5 hours/week × 50 weeks = 250 hours/year. At $50/hr, $12,500/year.
5-person team, 1 year
- Team: $25 × 5 × 12 = $1,500
- Productivity lift on writing and code: ~20% across team. For $500k blended payroll: $100k recovered.
50-person dev org, 2 years
- Enterprise (negotiated, est $30/user/mo) × 50 × 24 = $36,000
- Compliance overhead negligible (Anthropic's stack is enterprise-ready)
- Recovered productivity at 10% lift on $10M payroll = $2M/year
What's next for Claude
- Voice mode — Anthropic has confirmed they're working on it. No public timeline. Expected within the next year.
- Computer Use expansion — currently beta. Expected to move to GA with broader plan availability.
- Image generation — repeatedly hinted at but not yet committed. Likely 2027.
- Mobile Artifacts — currently desktop-only. Mobile parity expected late 2026.
- Larger context windows — 1M context already on Enterprise. Expected on Pro/Max within a year.
- Multi-modal output — Claude understanding voice, video, and complex documents is expected; outputting voice/video less certain.
What people are saying
Dario Amodei@DarioAmodei · on x.comClaude 4.7 Sonnet's biggest jump isn't on benchmarks. It's on the tasks that don't have benchmarks — writing with voice, thinking through ambiguous problems, knowing when to say 'I'm not sure.'
Ethan Mollick@emollick · on x.comFor writing tasks where the output matters, Claude beats GPT-5 consistently in blind tests with my students. Not by a lot — but reliably. If your work product is words, this is the model.
Simon Willison@simonw · on x.comClaude Artifacts is the feature ChatGPT keeps trying to copy. Side-by-side document + chat, with the document being live-editable from the chat. Once you use this for writing, the single-panel chat feels primitive.
Andrej Karpathy@karpathy · on x.comI switched my primary coding model from GPT to Claude six months ago and haven't looked back. The thing Claude does better: it asks 'wait, what are you actually trying to do' instead of just answering the literal question.
Pieter Levels@levelsio · on x.comClaude Pro at $20 = best ROI of any subscription I pay for. Two prompts a day pays for the month. I use it for code, writing, decision frameworks. The only AI that feels like talking to someone who's actually competent.
FAQ
Claude or ChatGPT?
For writing and code, Claude. For voice, images, and ecosystem, ChatGPT. Most power users keep both — $40/mo total is a small line item if you use them daily.
Is my data used to train Claude?
Not by default on any tier. Anthropic asks you to opt in explicitly if you want to share data for training. This is the opposite of ChatGPT's default behavior.
How does Pro compare to Max ($100/$200)?
Pro has limits that working professionals hit by mid-day. Max removes most of those limits. If you're using Claude as a daily tool, Max pays for itself in unmedded usage.
Can I use Claude with my company's confidential data?
On Pro: yes, but read the data agreement. On Team and Enterprise: yes with stronger contractual protections.
Does Claude have voice mode?
Not yet. If voice is critical, use ChatGPT.
What's the context window?
200K tokens on standard plans (Pro, Team). 1M on Enterprise. That's about 500 pages of text.
Can I bring my own model?
No. Claude only runs Anthropic's models. If you want flexibility to switch between providers, use Cursor (for code) or LibreChat (for general).
Is Claude available via API?
Yes. Anthropic API is the same models, pay-per-token. Used by many products (Cursor, etc.) under the hood.
Why is Claude more cautious than other models?
Anthropic's mission is AI safety. The hedging is intentional. Some users find it annoying; on high-stakes decisions, it's the right calibration.
Can I use Claude offline?
No. Requires internet for all features.
The verdict
The model for people whose work product is words.
Claude is the AI assistant that takes craft seriously. The writing has voice. The code compiles. The reasoning is calibrated. Anthropic's bet on safety as a core design principle produced a model that's genuinely trustworthy — it knows what it knows, it asks when uncertain, it pushes back when the brief is malformed. If you spend your day writing or coding, Claude Pro at $20 is the highest-ROI subscription on this list.
Pair it with ChatGPT ($20 extra) for voice and image generation. That $40/mo combo is the bottom of the AI stack for serious professional work in 2026.