ChatGPT The default choice

Where ChatGPT came from

OpenAI was founded in 2015 as a non-profit by Sam Altman, Elon Musk, Greg Brockman, Ilya Sutskever, and others, with a $1 billion pledge. The pitch was research-first — build AGI for the benefit of humanity, publish freely, no commercial pressure. Within five years the non-profit pivoted to a "capped-profit" structure, took a $1B investment from Microsoft, and started shipping product.

The lineage that produced ChatGPT runs through GPT-2 (2019, considered "too dangerous to release" at the time), GPT-3 (2020, the first model that broadly worked), InstructGPT (2022, fine-tuned to follow instructions), and finally ChatGPT — launched November 30, 2022 as a research preview. It hit 100 million users in two months. No app in history had ever grown that fast.

The strategic moves since: Microsoft's $10B follow-on in early 2023, the chaotic November 2023 board firing-and-rehiring of Sam Altman, GPT-4 in March 2023, the GPT Store in early 2024, voice mode through the year, GPT-4o (the "o" stands for omni — multi-modal), and finally GPT-5 and o1 reasoning in 2025. The product has shipped roughly monthly since launch.

The thing that makes ChatGPT specifically — not the GPT models, the product — defensible is the platform layer. Custom GPTs, the GPT Store, the API, the mobile apps with voice and vision, the enterprise tier with admin and SSO. Anthropic's Claude and Google's Gemini are competitive on raw model quality. None of them has the platform.

What ChatGPT actually is

ChatGPT is a conversational interface to a stack of OpenAI models, deployed across web, mobile (iOS + Android), and native desktop apps for Mac and Windows. The user-facing surface is a single chat window. The behind-the-scenes routing decides which model and which tools to use based on your request and your subscription tier.

The major capabilities, as of mid-2026:

• Text generation — writing, summarization, brainstorming, analysis (GPT-4o on Free, GPT-5 on Plus+)
• Reasoning — o1 family for math, code, and multi-step logic (Plus and Pro)
• Voice — real-time conversation on mobile with sub-second latency
• Vision — upload images, ask questions about them, generate from them
• Image generation — DALL-E 3 built in (Plus); plain English prompts
• Code interpreter — runs Python in a sandbox, handles data analysis, charts
• Web browsing — fetches current information when needed
• Custom GPTs — your own configured assistants with persistent instructions, knowledge files, and tool access
• Canvas — side-by-side document editor for collaborative writing and code

chatgpt.com · main-interface.png

First five minutes in action

Open chat.openai.com. Sign in with Google, Microsoft, or email. You're chatting in under two minutes. The Free tier gives you GPT-4o with usage caps that reset every few hours, basic voice mode on mobile, image upload, DALL-E for a couple of generations per day, and access to public Custom GPTs.

What works immediately:

• Type a question — get a coherent answer that's usually 70-90% of the way there
• Upload a photo of a confusing error message, a chart, a recipe — get explanation, transcription, or summary
• On mobile, tap the headphones icon for voice mode — talk like you're calling a friend who happens to know everything
• Drag a CSV into the chat — get instant analysis, charts, statistical summaries

The thing new users underestimate: follow-up prompts are the actual feature. The first answer is usually mediocre. The second answer, after "make that more concise" or "add code examples" or "what would you change for a non-technical reader," is the answer you actually wanted. Single-shot prompting is for power users with practiced prompts. Iterative chat is what ChatGPT is built for.

Setup, custom GPTs, and the GPT Store

Worth doing in your first week: explore Custom GPTs. These are personalized ChatGPT instances with persistent instructions, optional knowledge files, and optional tool access (web browsing, DALL-E, code interpreter). You can use other people's (the GPT Store has tens of thousands), or build your own in five minutes.

How to build one: click your name → My GPTs → Create. The builder is a meta-chat — you describe what the GPT should do, ChatGPT interviews you, and the system writes the instructions automatically. A typical Custom GPT setup takes 10-15 minutes including testing. Examples that work well:

• "Code reviewer" — instructions: "Review code I paste for bugs, suggest improvements, and explain rationale. Always quote line numbers."
• "Email drafter" — instructions: "Draft replies in my voice (warm, brief, action-oriented). Always end with a specific next step."
• "Recipe converter" — instructions: "Convert any recipe to vegan, gluten-free, or 30-minute versions. Show me what changed and why."
• "Decision journal" — instructions: "Help me think through decisions. Ask 3 clarifying questions before suggesting frameworks."

The GPT Store has the same problem app stores always have — most are bad, the top 1% are genuinely useful. The most reliable category is "tools your team uses daily that need an AI layer" — code reviewer, meeting note synthesizer, copywriter. Generic productivity GPTs are usually worse than just asking ChatGPT directly with a clear prompt.

chatgpt.com · gpt-store.png

The model family (GPT-4o vs GPT-5 vs o1)

OpenAI ships three main model families inside ChatGPT, and choosing the right one for your task is the second-biggest skill gain after learning to write good prompts.

GPT-4o — the default workhorse

Fast, multi-modal, available on Free tier. Handles text, voice, vision, and code interpreter. The model behind most ChatGPT interactions. "Omni" because it can take in text, images, audio, and produce all three. Use for conversational work, writing, basic coding, voice mode, vision tasks. About 1.5-second response for typical prompts.

GPT-5 — the heavy lifter

Plus and Pro tiers. Better reasoning, longer context window (200K tokens), better at complex multi-step tasks, more reliable for production code. Slower per response (3-6 seconds typical). The default when you need the answer to actually be right on the first try. Plus users get 80 messages per 3 hours on GPT-5 — generous for normal use, tight for heavy days.

o1 — the reasoning specialist

Available on Plus and Pro. Uses internal chain-of-thought before responding. Excellent for math, complex code, multi-step logic problems where the path matters as much as the destination. Slow (10-30 seconds per response is normal) and not chatty — feels more like submitting a job than having a conversation. Use sparingly, when you have a hard problem and want it solved correctly.

How it actually feels (the hands-on bit)

The biggest single shift ChatGPT introduced is that you stop using search for many of the questions you used to Google. "What's the difference between X and Y" — used to be a 20-minute reading dive across three blog posts. Now it's a 30-second exchange that lands at the answer plus pointed clarifying questions.

For writing, ChatGPT is the strongest first-drafter we've ever used. Email replies, README files, marketing copy, social posts — all faster to start from a ChatGPT draft and edit than to write from scratch. Quality on long-form nuance (essays, technical documentation) lags Claude — Claude writes with more voice. ChatGPT writes with more polish.

For code, ChatGPT is excellent for greenfield work (write a new function, scaffold a new project) and decent for debugging. Worse than Cursor or Claude Code for working inside an existing codebase — it doesn't read your repo, so suggestions assume defaults that may not match your conventions. Best workflow: ChatGPT for "write me a function that does X" in a clean context; an editor-integrated AI for everything else.

For research and learning, ChatGPT is the best on-demand tutor most people have ever had. Ask it to explain quantum mechanics, the bond market, or how a transformer model works — at any level from "explain like I'm five" to "give me the math." The one rule: verify any specific claim. ChatGPT confidently states wrong dates, wrong names, and wrong attributions. Use it for the explanation, not for the facts.

Three real workflows, end-to-end

chatgpt.com · code-interpreter.png

Real prompt → real output

We gave ChatGPT this prompt:

Eight seconds later:

200 words, hit the brief, included the specific differentiator, didn't drift into "revolutionary, game-changing, paradigm-shifting." We'd ship this with one or two minor edits.

Performance, measured

Across 100 representative prompts (50 writing, 30 coding, 20 reasoning), here's how ChatGPT compares to Claude and Gemini:

Pricing for consumer tiers is identical across the big three. Quality differences are real but narrow. Claude leads writing slightly. Claude leads code. Gemini leads speed and Google integration. ChatGPT leads everything else — voice, image gen, ecosystem, tools, mobile UX.

Privacy and security — what actually happens to your data

This is the section that matters most for anyone using ChatGPT at work. The short version: by default, conversations on the consumer tiers (Free, Plus, Pro) can be used to improve OpenAI's models unless you opt out. Conversations on Team and Enterprise tiers are not used for training and are contractually protected.

How to opt out on consumer tiers: Settings → Data Controls → "Improve the model for everyone" → toggle off. This affects all future conversations from that account. It does not retroactively remove past conversations from training data.

Where the data lives:

• Conversation history — stored on OpenAI servers, US region by default, accessible to you via the sidebar. Can be deleted; deletion is typically processed within 30 days.
• Memory — separate feature where ChatGPT remembers facts about you across sessions. Stored on OpenAI servers. Can be reviewed and edited in Settings → Personalization.
• Custom GPT knowledge files — uploaded files stay associated with the GPT. Read by GPT-4o or GPT-5 during chats. Encrypted at rest.
• Voice recordings — voice mode audio is transcribed and stored as text. Audio itself is deleted within 30 days unless used for trust and safety review.

What ChatGPT doesn't see on consumer tiers: nothing. Anything you type or upload, OpenAI's systems see. If your security policy is "no proprietary code, no PII, no client data goes to external AI tools" — that policy precludes consumer ChatGPT, period. The path forward is Team or Enterprise plans, where training is off by default and additional admin controls are available.

ChatGPT vs Claude

ChatGPT vs Google Gemini

ChatGPT vs Perplexity

Where ChatGPT gets it wrong

Confident hallucinations

ChatGPT will state false things with high confidence. Fake citations. Wrong dates. Misattributed quotes. The error rate is low (maybe 1-3% on factual claims) but consistent. Always verify specific facts. The 2026 update reduced hallucination significantly but didn't eliminate it.

Aggressive content policies

ChatGPT refuses requests Claude and open-source models handle. Medical advice, security research, creative violence in fiction, anything touching public figures. The policies have loosened since 2023 but ChatGPT is still the most restrictive of the major models. If you're writing fiction with morally complex characters, this gets annoying fast.

Usage limits surprise heavy users

Plus tier limits GPT-5 to 80 messages per 3 hours. Sounds generous until you're in a long analysis session and hit it mid-flow. The model silently switches to GPT-4o, often without warning, and the answer quality drops. Pro tier ($200/mo) removes most caps. Team and Enterprise have higher limits than Plus.

"As an AI" boilerplate creep

GPT-5 reduced this dramatically but still occasionally inserts "As an AI assistant, I should note..." disclaimers nobody asked for. Trim them from your prompts: "Skip disclaimers. Get to the answer." Or build a Custom GPT with these instructions baked in.

Memory feature feels half-baked

The "memory" feature, where ChatGPT remembers facts about you across sessions, is useful when it works and creepy when it surfaces something weird. It's still inconsistent — sometimes it remembers your dog's name, sometimes it forgets you're a developer. Better to keep memory off and just put context in your prompt for important sessions.

Image generation lags Midjourney

DALL-E inside ChatGPT is convenient but the quality gap to Midjourney v6 is real. ChatGPT image gen is great for "draft me an illustration" — bad for production-grade visuals. Use Midjourney or Stable Diffusion when output quality matters.

chatgpt.com · dall-e-image-gen.png

Power-user tips

Common gotchas (8 things new users miss)

1. Free tier rate limits feel arbitrary. "You've reached the GPT-4o limit" can hit at 20 messages or 60 depending on server load. Plan around mid-day if you're on Free.
2. Mobile and desktop sync is sometimes laggy. A conversation you start on desktop may not show up on mobile for several minutes. Refresh both.
3. Voice mode burns battery. 30 minutes of voice mode costs about 25% of an iPhone's battery. Plan accordingly on travel days.
4. Image upload limits are tier-dependent. Free: 1-3 per day. Plus: 50 per day. Pro: unlimited. Easy to hit.
5. Custom GPTs don't transfer between accounts. If you switch from personal to Team account, you rebuild your custom GPTs. Annoying.
6. The web browsing tool is selectively triggered. ChatGPT decides when to browse. Sometimes it should and doesn't. Add "search the web" explicitly when you need current info.
7. Code interpreter sandboxes reset. Files you uploaded for analysis disappear after a few minutes of inactivity. Save outputs immediately.
8. Subscription doesn't share across team members. Plus is single-user. If two people on your team need access, you need Team (or two Plus subs).

Best practices — daily habits of power users

• Set custom instructions once, use forever. Add: your profession, your communication style preference, "skip disclaimers and caveats."
• Build 3-5 Custom GPTs you use weekly. Code reviewer, email drafter, content critic. Each saves 2-3 minutes per use.
• Switch models per task. GPT-4o for chat, GPT-5 when stakes go up, o1 for hard problems. Most users default to GPT-4o for everything — wrong.
• Iterate, don't restart. Bad answer? Reply with feedback. Don't open a new chat. The context is the value.
• Use Canvas for documents over 500 words. Lets you edit in place, not regenerate the whole thing.
• Pin reusable prompts. Star or save them in a notes app. The 10 prompts you'll write 50 times each year — find them, save them.
• Verify facts that matter. ChatGPT is right about 97% of the time. The 3% includes the thing you'd be embarrassed to repeat.

Pricing, in real terms

Free: GPT-4o with limits, basic voice, limited DALL-E. Good for casual use. You'll hit limits if you're using it as a work tool.

Plus ($20/mo): The default for individuals. GPT-5 and o1 access, generous limits (80 messages per 3 hours on GPT-5), voice mode unlimited, 50 DALL-E images/day, custom GPTs, file uploads.

Team ($25/user/mo, billed annually): Worth it the moment 3+ people on your team need ChatGPT. Higher limits, no training on your data by default, shared workspace, admin console.

Pro ($200/mo): For heavy individual users. Effectively unlimited GPT-5 and o1, exclusive access to o1 Pro mode (longer thinking), unlimited DALL-E, early access features. Pays for itself if you'd otherwise burn through Plus limits daily.

Enterprise (custom pricing): SOC 2, SSO, admin and audit logs, longer context, no training on your data, dedicated support.

chatgpt.com · pricing-tiers.png

Total cost of ownership

Solo user, 1 year

• Plus subscription: $20 × 12 = $240/year
• Time savings: even 30 min/day × 250 working days = 125 hours/year. At $50/hr, that's $6,250 of recovered time.
• Verdict: trivial ROI for anyone who uses it daily for work.

5-person team, 1 year

• Team plan: $25 × 5 × 12 = $1,500/year
• Productivity gain: typical 15-20% on knowledge work tasks. For $400k blended payroll: $60-80k/year recovered.
• Plus admin control over data, no training, shared GPT library across team.

50-person org, 2 years

• Enterprise: roughly $30/user/mo at this scale × 50 × 24 = $36,000
• Plus governance overhead (writing policies, training, monitoring): ~$15,000
• Total: ~$51,000 over 2 years
• For a $10M payroll, even 5% productivity lift = $1M/year recovered.

Enterprise adoption playbook

For engineering leads or operations heads rolling ChatGPT out to teams of 10-100:

Phase 1: Policy first (Week 1)

Before any rollout, write a one-page AI usage policy. What's allowed (drafting emails, analyzing public data, generating boilerplate code). What's forbidden (customer data, confidential strategy, regulated content). Who pays. How violations are handled. Get legal to sign off.

Phase 2: Pilot with willing volunteers (Weeks 2-4)

Identify 5-10 people who want to try it. Pay for Team accounts. Have them document workflows that worked, prompts that helped, and where they hit limitations. The pilots become your internal champions.

Phase 3: Build internal GPT library (Week 4)

Before broad rollout, create the 3-5 Custom GPTs your team needs most — Code Reviewer, Customer Support Drafter, Meeting Notes Synthesizer. These show new users immediate value on day one.

Phase 4: Rollout (Weeks 5-8)

Team-wide announcement. Opt-in initially, mandatory only for roles that benefit clearly. Run a 30-minute "Show me how" session with pilots demonstrating actual workflows. Avoid the trap of forcing AI usage — let value speak for itself.

Phase 5: Monitor and govern (ongoing)

Admin dashboard shows usage by user. Identify high-engagement people — make them mentors. Quarterly policy reviews. Track ROI in concrete terms: "X hours saved on Y workflow this quarter."

What's next for ChatGPT

What people are saying

Sam Altman

@sama · on x.com

GPT-5 is the first model where I find myself genuinely surprised by the quality of its reasoning. Not impressed — surprised. There's a difference.

// view on x.com

Ethan Mollick

@emollick · on x.com

For a year I told MBA students to use Claude for writing and ChatGPT for everything else. With GPT-5 the gap closed. ChatGPT is now genuinely competitive with Claude on long-form writing while keeping its lead on tool use, voice, and image gen.

// view on x.com

Linus Lee

@thesephist · on x.com

The thing ChatGPT does that Claude doesn't: it's a platform. Custom GPTs, the API, voice on mobile, vision, DALL-E, data analysis. Claude is one chat. ChatGPT is a stack.

// view on x.com

Simon Willison

@simonw · on x.com

Honest take on GPT-5 limits: Plus tier hits the cap fast. If you're doing real work daily you'll want Team or API. ChatGPT Plus is the demo. The actual productivity tier is one step up.

// view on x.com

Aravind Srinivas

@AravSrinivas · on x.com

ChatGPT remains the easiest AI to recommend to your non-technical friend. The UX work OpenAI has done on conversational AI is still the standard everyone copies.

// view on x.com

The competition we considered

ChatGPT is the default. But we evaluated and rejected (or recommended as supplements) these alternatives. None of them changes the verdict, but some of them are right for specific use cases:

• Claude (Anthropic) — Better writing, better code. Lacks voice, image gen, ecosystem. Use as a daily complement for serious writing or coding.
• Gemini (Google) — Best free tier. Best Workspace integration. Slightly behind on polish and ecosystem.
• Perplexity — Best for research with citations. Not a general assistant.
• Mistral Le Chat (France) — Open weights, EU-hosted, strong on French. Smaller ecosystem.
• Llama 3 / 4 (Meta, self-hosted) — When you absolutely need on-prem inference. Quality is competitive, deployment cost is real.
• Cohere Command — Enterprise-focused, strong RAG capabilities. Not a consumer product.

FAQ

The verdict