Category: Insights & News

In-depth comparisons and trend analysis of AI tools, with verified plan and pricing facts.

  • Anthropic’s Own Postgres MCP Server Is Archived and Has a Read-Only Bypass: 5 Database & Backend MCP Servers to Trust Instead (2026)

    Anthropic’s Own Postgres MCP Server Is Archived and Has a Read-Only Bypass: 5 Database & Backend MCP Servers to Trust Instead (2026)

    Last Verified: September 2026 — commit activity, star counts, and install commands checked against live GitHub sources

    You connect Claude Code to your database through the Postgres MCP server every list recommends, assuming read-only means read-only. That server is Anthropic’s own reference implementation, it’s archived, and it has a documented read-only bypass. The same lists rank a “1.7k-star” MongoDB integration above an official server that ships weekly.

    We ran the same maintenance check from our first curation guide against the database and backend category. Two names got cut for being stale or unsafe; five made it through.

    Quick Start

    Install any of these with the Claude Code CLI. Swap in your own credentials before running.

    claude mcp add supabase -- npx -y @supabase/mcp-server-supabase@latest --access-token=$SUPABASE_ACCESS_TOKEN
    claude mcp add redis -- uvx --from git+https://github.com/redis/mcp-redis.git redis-mcp-server --url redis://localhost:6379/0
    claude mcp add mongodb -- npx -y mongodb-mcp-server --connectionString="$MDB_URI"
    claude mcp add postgres -- uvx postgres-mcp --access-mode=restricted "$DATABASE_URI"
    claude mcp add clickhouse -- uvx mcp-clickhouse

    ⚡ Copy This Prompt

    Paste this into Claude Code to let it install only what your stack actually uses:

    I work with [Postgres / Supabase / Redis / MongoDB / ClickHouse -- name your stack].
    Install only the matching MCP server(s) from this list, using the credentials I have
    available in my environment variables. After installing, run one safe read-only
    command through each server (e.g. list tables, list keys, or list databases) to
    confirm the connection actually works. Tell me plainly if a connection or a
    permission check fails -- don't guess or paper over an error.
    What You’ll NeedNotes
    Claude Code CLI installedAny recent version supports claude mcp add
    Connection string or access token per databaseNever paste raw credentials into chat — use environment variables
    uv / uvx installedRequired for the Redis, Postgres, and ClickHouse servers (Python-based)
    Node.js + npmRequired for the Supabase and MongoDB servers (JS-based, run via npx)

    1. Supabase MCP (supabase-community/supabase-mcp)

    Official · ~2.6k stars

    Supabase’s own team maintains this server, and it’s the one we’d point most readers to first — it covers the whole BaaS surface (Postgres tables, auth, storage, edge functions) through one connection instead of stitching together separate tools.

    Why it made the cut: official repo, consistent commit cadence (roughly 20+ commits/month in recent audits), and it ranks near the top of every independent “best database MCP” comparison we checked.

    Watch out for: by default it can run destructive operations against your project. Pass --read-only if you’re pointing it at anything other than a scratch/dev project.

    2. Redis MCP (redis/mcp-redis)

    Official · Redis Inc.

    Smaller star count than the others on this list, but it’s the vendor’s own server, not a community reimplementation — and an unofficial alternative with more stars (GongRzhe/REDIS-MCP-Server) was archived in March 2026, which makes this the safer long-term pick despite the lower number.

    Why it made the cut: actively released by Redis Inc. itself; covers strings, hashes, streams, and pub/sub through one consistent tool surface.

    Watch out for: star count alone would make this look like the “unpopular” option. Don’t let that fool you — vendor backing matters more than stars for a cache/session-store integration you’re trusting with production data.

    3. MongoDB MCP (mongodb-js/mongodb-mcp-server)

    Official · GA · ~1,000 stars

    MongoDB’s official server is now generally available, with 41+ tools spanning database operations, Atlas cluster management, and performance advisory — and it ships new releases roughly every 1-2 weeks.

    Why it made the cut: vendor-official, GA status, and read-only mode plus environment-variable credentials are the security defaults, not an opt-in.

    Watch out for: several older community MongoDB servers still show up in search results with similar names. If the repo isn’t under the mongodb-js or mongodb-developer org, you’re looking at an unofficial fork with a smaller maintenance team behind it.

    4. Postgres MCP Pro (crystaldba/postgres-mcp)

    Community · ~3.2k stars

    This is the one with the asterisk. It’s the most popular general-purpose Postgres server, with index tuning, EXPLAIN plan analysis, and health checks built in — genuinely useful tools the official (now-archived) Anthropic server never had. But development has slowed: our check found roughly one commit in the last 12 weeks and a backlog of open issues, some unanswered since early 2025.

    Why it made the cut anyway: “slower” is not “abandoned” — it’s still shipping, still the most feature-complete option, and it beats the alternative of using Anthropic’s archived reference server, which has a documented read-only bypass and receives no security patches at all.

    Watch out for: if you need a strict read-only default for a production database, look at HenkDz/postgresql-mcp-server instead — smaller (~200 stars) but built specifically so writes, DDL, and arbitrary SQL all require an explicit opt-in flag.

    5. ClickHouse MCP (ClickHouse/mcp-clickhouse)

    Official · ~800 stars

    The odd one out on most “top 5 database MCP” lists, and that’s exactly why it’s here — if your backend does analytics or event logging at scale, a transactional-database server won’t help you. ClickHouse’s own team maintains this one, distributed as a Docker Hub Verified Publisher image.

    Why it made the cut: official, actively maintained, and it’s the only analytics-database option that passed our maintenance check — most of the community ClickHouse servers we found were single-author projects with no recent activity.

    Watch out for: it expects ClickHouse-specific connection env vars (host, port, user, password) rather than a single connection-string argument like the others on this list.

    Verify Each One Actually Works

    $ claude mcp list
    supabase — connected
    redis — connected
    mongodb — connected
    postgres — connected
    clickhouse — connected

    Then ask Claude Code to run one read-only action through each — list tables, list keys, list databases — before you trust it with anything that writes.

    Common Mistakes to Avoid

    • Judging by star count alone. Redis’s official server has a fraction of the stars of some abandoned forks. Vendor backing and commit recency matter more than the number.
    • Pointing a write-capable server at production on day one. Test every new MCP connection against a dev/staging database first.
    • Assuming “modelcontextprotocol/*” means officially supported. Anthropic archived 13 of its 20 original reference servers in 2025, including Postgres, SQLite, and Redis — those repos still run, but they get no security patches.
    • Skipping the read-only flag on Supabase. Without it, the server can alter schema and data in your connected project.

    Q&A

    Why is Postgres MCP Pro included if development slowed down?

    Because “slower than ideal” still beats the alternative: Anthropic’s own reference Postgres server is archived and has a known read-only bypass vulnerability. If you need a stricter security posture, we also name a smaller alternative (HenkDz/postgresql-mcp-server) in that section.

    Should I use the official or community server when both exist?

    Default to official when the vendor maintains one — as with Redis, Supabase, MongoDB, and ClickHouse here. Community servers can still be worth using, but check commit recency first; we cover that process on our Methodology page.

    Can I install more than one of these at once?

    Yes — each runs as an independent MCP connection. If you use Playwright or GitHub MCP alongside these, see our Playwright MCP guide and GitHub MCP guide for the same install pattern.

    Official Resources

  • Stop Installing Abandoned MCP Servers: 5 Community-Built Picks That Passed a 2026 Maintenance Check (and 2 Popular Ones That Failed)

    Stop Installing Abandoned MCP Servers: 5 Community-Built Picks That Passed a 2026 Maintenance Check (and 2 Popular Ones That Failed)

    Last Verified: September 2026 — commands and maintenance status checked against live GitHub sources

    You install a popular MCP server, it breaks, and then you find out the repo was archived months ago. Most “top MCP servers” lists are just GitHub star counts copy-pasted into a listicle. That’s a bad filter. A popular Obsidian MCP project (jacksteamdev/obsidian-mcp-tools, 800+ stars) is now archived — read-only, no more fixes. Mem0’s own self-hosted MCP server (mem0ai/mem0-mcp) is archived — the README opens with a notice that it’s no longer maintained, and Mem0 now points users to its cloud-hosted server instead. Stars measure past hype, not whether a project still works.

    We checked five servers instead of just ranking them: recent commit activity, how open issues get handled, and whether the practical claim (“free search,” “read-only by default”) actually holds up in the current version. These five passed.

    Quick Start

    Install all five with Claude Code’s CLI. Each is a separate claude mcp add call — run only the ones you actually need:

    claude mcp add --transport stdio desktop-commander -- npx -y @wonderwhy-er/desktop-commander
    claude mcp add --transport stdio n8n-mcp -- npx -y n8n-mcp
    claude mcp add --transport stdio slack -- npx -y slack-mcp-server
    claude mcp add --transport stdio duckduckgo -- uvx duckduckgo-mcp-server
    claude mcp add --transport stdio mysql -- npx -y @benborla29/mcp-server-mysql

    Restart Claude Code after adding — MCP connections load at session startup. Total time: about 5 minutes for all five, or under a minute each.

    ⚡ Copy This Prompt: Let Claude Code Pick and Verify the Right Ones For You

    I want to install MCP servers from this list, but only the ones relevant to my actual workflow: Desktop Commander (terminal + filesystem control), n8n-mcp (build n8n workflows via natural language), Slack MCP (read/post to Slack), DuckDuckGo MCP (free web search, no API key), MySQL MCP (read-only database access).
    
    1. Ask me which 1-3 of these match what I actually do day to day.
    2. Run the "claude mcp add" command for only those.
    3. After I restart, run one real action through each newly connected server (not just a connection check) and show me the output.
    4. If any server fails to connect or the real-action test fails, tell me exactly what's wrong instead of saying it worked.

    This matters because installing all five “just in case” is the exact mistake every 2026 MCP guide warns against — more servers means more standing access an agent can act through, and no one uses half of it.

    What You’ll Need

    Requirement Why You Need It Time
    Claude Code installed All five connect via claude mcp add 0 min
    Node.js or uv/uvx Each server ships as an npx or uvx package ~2 min if missing
    Per-server credentials (Slack token, MySQL connection string) Only Slack and MySQL need them — the other three need none varies

    The Five, One at a Time

    1. Desktop Commander — terminal + filesystem control

    ~9.7k stars · MIT

    github.com/wonderwhy-er/DesktopCommanderMCP

    Gives Claude direct terminal control, file search, and diff-based file editing — the thing people actually reach for instead of chaining together separate read/write/bash tools.

    claude mcp add --transport stdio desktop-commander -- npx -y @wonderwhy-er/desktop-commander

    Why it made the cut: commit activity through September 2026, cross-platform (macOS/Windows/Linux), works with 15+ AI clients beyond Claude.

    Watch out for: it’s terminal access — same blast-radius reasoning as giving Claude Code Bash. Don’t run it in a context you wouldn’t give shell access to.

    2. n8n-mcp — build n8n workflows in natural language

    ~22.8k stars

    github.com/czlonkowski/n8n-mcp

    If you followed our n8n setup guide, this is the complementary piece: instead of clicking nodes together manually, you describe the workflow and Claude builds it in n8n directly.

    claude mcp add --transport stdio n8n-mcp -- npx -y n8n-mcp

    Why it made the cut: largest install base on this list by a wide margin, and the maintainer has published multiple security advisories (GHSA) in 2026 — a sign of active review, not neglect.

    Watch out for: solo-maintained project; the author has been open that maintaining it competes with paid work, so response times on issues can be slow.

    3. Slack MCP Server — the community server that replaced the official one

    ~1.8k stars

    github.com/korotovsky/slack-mcp-server

    Anthropic’s own reference Slack server was retired to an archive repo in 2026. If you want Slack in your agent today, this is the de facto replacement, not an official one.

    claude mcp add --transport stdio slack -- npx -y slack-mcp-server

    Why it made the cut: it’s the specific project the official docs now point to, with active 2026 development.

    Watch out for: you’re trusting a community maintainer with a Slack token — scope that token’s permissions down before connecting.

    4. DuckDuckGo MCP Server — free web search, no API key

    ~1.4k stars

    github.com/nickclyde/duckduckgo-mcp-server

    Every other search MCP wants a paid API key (Brave, Google, Bing). This one doesn’t — it’s DuckDuckGo search plus page-content fetching, free.

    claude mcp add --transport stdio duckduckgo -- uvx duckduckgo-mcp-server

    Why it made the cut: concrete proof of active maintenance — DuckDuckGo started fingerprint-blocking the default HTTP client in mid-2026, and the maintainer shipped a fallback within the same release cycle. That’s the difference between “still maintained” and “still has a green checkmark on npm.”

    Watch out for: free tier of a scraped search engine — expect occasional rate-limiting under heavy use, not an SLA.

    5. MySQL MCP Server — read-only by default

    ~1.7k stars

    github.com/benborla/mcp-server-mysql

    Lets Claude inspect schemas and run queries against a real MySQL database. The reason it’s on this list instead of a dozen equivalents: writes are opt-in via an explicit env flag, not the default.

    claude mcp add --transport stdio mysql -- npx -y @benborla29/mcp-server-mysql

    Why it made the cut: the safe-by-default design is the right call for a database an agent can reach — you have to deliberately turn on write access.

    Watch out for: last major commit was March 2026 — still maintained, but slower-moving than the others on this list. Fine for read-only schema inspection; verify current behavior before enabling writes in production.

    Verify Each One Actually Works

    Connection status isn’t proof. Run one real action through each:

    desktop-commander    ran ls in project root, got real output
    n8n-mcp    listed existing workflows from a live n8n instance
    slack    posted a test message to a private test channel
    duckduckgo    searched a query, got real result URLs back
    mysql    ran SHOW TABLES against a real (non-production) schema

    If a server shows “connected” in claude mcp list but fails on a real action, the connection succeeded and the tool call is what’s broken — usually a missing credential or a stale cached version.

    Common Mistakes to Avoid

    • Installing all five because the list exists. Every 2026 MCP guide converges on the same warning: more connected servers means more standing access, most of it idle. Install what matches your actual workflow.
    • Trusting star count alone. A popular Obsidian MCP project with 800+ stars is archived, and its star badge looks exactly the same as before. Check the commit date, not just the badge.
    • Assuming “official” beats “community.” Slack’s official server is gone; the community one is now the standard. Vendor-backed isn’t always current.
    • Skipping the credential scoping step. Slack and MySQL both need real credentials — scope them to the minimum before connecting, not after.

    Q&A

    Why isn’t Playwright, GitHub, or Context7 MCP on this list?

    They already have their own dedicated setup guides — see Playwright MCP, GitHub MCP, and Context7 MCP.

    What got excluded and why?

    A popular Obsidian MCP project (jacksteamdev/obsidian-mcp-tools, archived) and Mem0’s self-hosted MCP server (archived in favor of Mem0’s hosted version) both looked like obvious picks by star count and both failed the maintenance check.

    Is a solo-maintained project (n8n-mcp, Slack MCP) too risky to depend on?

    It’s a real tradeoff, not a dealbreaker — check for recent commits and how issues get triaged before depending on any single-maintainer tool, the same way you’d vet any dependency.

    Official Resources

  • ChatGPT Text Chat Is Unlimited Now. Here’s What Still Runs Out (and How to Make It Last)

    ChatGPT Text Chat Is Unlimited Now. Here’s What Still Runs Out (and How to Make It Last)

    Last Verified: September 24, 2026 — plans and prices checked on each vendor’s official pricing page. Prices are USD and vary by country.

    For years the standard ChatGPT advice was about rationing: batch your questions, don’t waste messages, wait out the cooldown. Most of that is out of date. OpenAI’s pricing page now lists unlimited text chats on the Free plan. If you only type questions and read answers, you won’t hit a wall.

    That doesn’t mean nothing runs out. The limits moved. They’re now on the expensive features: file uploads, image generation, voice, deep research, memory, and the newer reasoning models. This piece covers what still has a cap on each plan, what it feels like when you hit it, and the habits that stretch each one.

    Key Takeaways

    • Text chat is unlimited on Free. Uploads, images, voice, deep research, memory and context, and Codex are all limited.
    • Go ($8) raises those limits but may include ads. Plus ($20) adds GPT-6 reasoning models, Projects, and custom GPTs.
    • The limit most people actually feel is context: long chats drift. Starting fresh with a summary fixes it.
    • Don’t upgrade on a hunch. Note which limit you hit, and how often, for two weeks first.

    What Each Plan Limits

    FeatureFreeGo ($8/month)Plus ($20/month)
    Text chatUnlimited (GPT-5.6 Luna)UnlimitedUnlimited, plus GPT-6 reasoning models
    File and image uploadsLimitedMoreMore
    Image generationLimited and slowerMoreMore
    Voice chatsLimitedMoreMore
    Deep researchLimitedMoreMore
    Memory and contextLimitedMoreMore
    Codex (coding agent)LimitedMoreExpanded
    Projects, scheduled tasks, custom GPTsIncluded
    AdsMay include ads

    OpenAI publishes the categories but not exact numbers for most of these, and it adjusts them over time. Pro (from $100/month) offers 5x more usage and GPT-6 Astra, but it’s aimed at people who use ChatGPT for hours a day.

    The Limits You’ll Actually Hit, and How to Stretch Them

    1. Context: the limit nobody sees coming

    Every chat has a working memory. In a very long conversation, the early details get less attention, and you’ll see it: ChatGPT forgets a constraint you set at the start, contradicts an earlier decision, or asks for something you already gave it. On Free, with less memory and context, this happens sooner.

    The fix: when a chat gets long or starts drifting, ask for a handoff summary and start a new chat with it:

    Summarize this conversation for a fresh chat: the goal, every decision we made, the constraints I set, and what's still open. Under 200 words.

    Paste that into a new chat. You keep everything that matters and drop the clutter. This one habit fixes more “ChatGPT got dumber” complaints than any upgrade.

    2. Uploads

    Uploading files and images is limited on Free. Two ways to go further:

    • Paste text instead of uploading. For a document you only need read, copy the relevant section into the chat. It doesn’t count as an upload, and it forces you to pick the part that matters.
    • Combine files. Three short documents in one file is one upload instead of three.

    3. Image generation

    Free image generation is limited and slower. Most of the waste comes from vague first prompts that need four or five rerolls. Write the complete instruction once: subject, style, lighting, and, for edits, what must stay unchanged. Our photo editing prompts and product photo template are built to land on the first or second try.

    4. Deep research

    Deep research runs a multi-step web investigation and writes a report. It’s one of the most limited features because it’s one of the most expensive to run. Save it for questions that genuinely need many sources. For a quick fact, a normal chat with web search is faster and doesn’t use a deep research run. Before starting a run, narrow it down: say what you already know, which sources you trust, and what format you want the report in.

    5. Voice

    Voice chats are limited on Free. If you use voice for thinking out loud, draft your main question by voice, then continue in text. Text chat is where the unlimited allowance is.

    Stop Repeating Yourself: Custom Instructions and Projects

    Retyping the same background (your role, your audience, how long answers should be) wastes time and fills the context window. Two features fix that.

    • Custom instructions (settings, available on every plan): tell ChatGPT once who you are and how you want answers formatted. It applies to every new chat.
    • Projects (Plus and above): group related chats with shared files and instructions, so a client or a long piece of work keeps its own context.

    A good custom instruction is short and specific. For example:

    I'm a marketing manager at a small software company. Default to concise answers: lead with the answer, then up to 5 bullets. Use plain language. If my request is ambiguous, ask one clarifying question before answering.

    Write Prompts That Don’t Need a Second Round

    Even with unlimited text chat, every follow-up costs time and adds clutter that makes context drift sooner. The biggest wins come from putting everything in the first message: the format, a length limit, and what to leave out. Our before/after guide covers the five mistakes that cause most follow-ups, and the RTFC framework gives you a four-slot template to fill in.

    The Make-It-Last Checklist

    Everything above, in one list you can keep next to your browser:

    • Set custom instructions once, with your role, audience, and default answer length.
    • Put format, length, and what to leave out in the first message, not in follow-ups.
    • Paste the relevant section of a document instead of uploading the whole file.
    • Combine small files into one before uploading.
    • Write complete image prompts, including what must stay unchanged, before generating.
    • Use normal chat with web search for quick facts. Save deep research for questions that need many sources.
    • Start voice, finish in text.
    • When a chat drifts, ask for a handoff summary and continue in a new chat.
    • Keep a note of every limit you hit for two weeks before deciding on an upgrade.

    None of this needs a paid plan. On Free, these habits decide whether you hit a limit on day one or never notice it. On a paid plan, they’re the difference between Plus being enough and thinking you need Pro.

    When Upgrading Is Worth It

    If you keep hitting…ConsiderWhy
    Upload, image, or voice limits, but text is fineGo ($8)Higher limits at the lowest price; may include ads
    Harder reasoning tasks where answers fall shortPlus ($20)GPT-6 reasoning models
    Re-explaining the same project every weekPlus ($20)Projects keep files and instructions together
    Plus limits, daily, for workPro (from $100)5x usage; only if AI is core to your job

    Before paying, check whether a different tool fits better. If what you need is Gmail and Docs integration or a coding agent, see our Claude vs ChatGPT vs Gemini comparison.

    Q&A

    Is ChatGPT Free really unlimited now?

    For text chats, yes, according to OpenAI’s pricing page as of September 2026. Uploads, image generation, voice, deep research, memory and context, and Codex are still limited.

    What happens when I hit a limit?

    The feature becomes unavailable until the limit resets, or you’re offered an upgrade. Text chat keeps working.

    Does starting a new chat lose my memory?

    Saved memories and custom instructions carry over to new chats. What you lose is the conversation itself, which is why the handoff summary matters.

    Is Go worth it over Free?

    If you regularly hit upload or image limits and don’t mind possible ads, it’s the cheapest way to raise them. If you mostly type questions, Free is enough.

    Sources

  • Claude vs ChatGPT vs Gemini in 2026: Which One Should You Actually Use?

    Claude vs ChatGPT vs Gemini in 2026: Which One Should You Actually Use?

    Last Verified: September 24, 2026 — plans and prices checked on each vendor’s official pricing page. Prices are USD and vary by country.

    “Which AI should I use?” is the wrong question, and most comparison posts answer it anyway. They crown a winner off a benchmark chart that will be out of date in a month, and you end up paying for a tool that’s great at something you never do.

    In 2026 all three assistants are good enough at plain writing and Q&A that the model itself is rarely what decides it. What decides it is where your work already lives, what kind of work it is, and which limits you’ll actually hit. This piece compares ChatGPT, Claude, and Gemini on exactly those three things, using what each company’s own pricing page says today.

    Key Takeaways

    • Live in Gmail, Docs, and Sheets all day? Gemini. Google AI Pro puts it inside the apps you already have open.
    • Want one assistant for everything (chat, images, voice, custom GPTs)? ChatGPT. The widest feature set, and Free text chat is unlimited.
    • Long documents, careful writing, or code? Claude. Pro includes Claude Code, and even Free includes Projects.
    • All three paid tiers sit around $20/month. Try the free tier on your real work for a week before you pay.

    The Plans Side by Side

    Here’s what each company lists for its main consumer tiers. Features change often, so treat this as a snapshot and check the linked pricing pages before you subscribe.

    ChatGPTClaudeGemini (Google AI plans)
    FreeUnlimited text chats with GPT-5.6 Luna; limited uploads, image generation, voice, deep research, memory, and CodexChat, web search, file creation, code execution, memory, Artifacts, up to 5 Projects; no ResearchGemini app with a Google account
    Entry paidGo, $8/month: more of everything in Free; may include adsAI Plus, $4.99/month: Gemini 3.1 Pro, Deep Research, proofreading in Gmail
    Main paidPlus, $20/month: GPT-6 reasoning models, Projects, scheduled tasks, custom GPTs, more CodexPro, $20/month ($17/month billed annually): more usage, Research, Claude Code, Claude in Chrome and Microsoft 365AI Pro, $19.99/month: Gemini in Gmail, Docs, and Sheets
    Power userPro, from $100/month: 5x usage, GPT-6 AstraMax, from $100/month: 5x or 20x Pro usageAI Ultra, from $99.99/month: Deep Think, Gemini Agent

    Two things stand out. First, the free tiers are real products now, not demos. ChatGPT Free has no cap on text chat, and Claude Free includes Projects, memory, and code execution. Second, the $20 tier is where the three split. ChatGPT spends it on breadth, Claude on depth and coding tools, Google on putting Gemini inside Workspace.

    Pick by Where Your Work Lives

    If your day runs through Google Workspace

    This is the easiest call. On AI Pro, Gemini sits in a side panel inside Gmail, Docs, and Sheets and already sees the email or document you have open. That cuts out the copy-paste round trip, which is where most of the time goes when you use a separate chatbot. If most of your AI use would be summarizing threads, drafting replies, and cleaning up spreadsheets, a standalone assistant has to be much better on quality to make up for that. Most of the time it isn’t. Our Gemini Workspace starter kit has prompts for exactly those jobs.

    One caveat: work and school accounts get Gemini through their Workspace edition, not a personal Google AI plan. If you don’t see the side panel at work, that’s an admin setting, not something you can buy your way around.

    If you work across lots of tools

    ChatGPT is the generalist. One subscription covers text, image generation and editing, voice conversations, custom GPTs, scheduled tasks, and Codex for code. If you want one tab that does a bit of everything, and you don’t want to think about which tool fits which task, ChatGPT asks the least of you. Image editing in particular is a daily-use feature for many people. See our photo editing prompts for how to keep edits under control.

    If you work with long documents or code

    Claude’s strengths are following detailed instructions closely and staying on track across long material. Two product details back that up. Projects (up to 5 on Free) let you keep standing instructions and reference files in one place. Pro includes Claude Code, Anthropic’s agentic coding tool that works directly in your terminal and repository. Most of the setup guides on this site are for Claude Code, because it’s what we use every day.

    Pick by the Task

    If you don’t have a single home base, go task by task. This is our read from daily use of all three, not a benchmark. Your results will vary with your prompts and your material.

    TaskBest starting pointWhy
    Summarizing email and drafting repliesGeminiWorks inside Gmail on AI Pro, no copy-paste
    Spreadsheet cleanup and formulasGeminiSide panel in Sheets sees the actual data
    Image generation and photo editsChatGPTBuilt in on every tier, including Free with limits
    Voice conversationsChatGPTAvailable on Free with limits; more on paid tiers
    Long reports, contracts, and research docsClaudeHolds long material and follows detailed instructions
    Writing that has to sound like youClaude or ChatGPTBoth take a voice sample well; test both
    Coding in a real repositoryClaudeClaude Code included with Pro
    Deep, multi-source research reportsAny paid tierAll three offer a deep research mode; Claude requires Pro

    Notice what’s missing: general Q&A and everyday writing. For “explain this concept” or “tighten this paragraph”, all three are fine, and your prompt matters more than the model. Our RTFC prompt framework will do more for those answers than switching tools.

    The Limits That Actually Bite

    Pricing pages list features. What you’ll actually notice is the limits. Here’s where each one tends to push you toward paying:

    • ChatGPT Free doesn’t cap text chat, but it does cap uploads, image generation, voice, deep research, and how much it remembers. Our ChatGPT limits breakdown covers what runs out and how to make it last.
    • Claude Free is generous on features but has lower usage limits than Pro, and no Research mode. Heavy daily use is where Pro pays for itself.
    • Gemini’s big split is between the app and Workspace. The Gemini app is available free, but the Gmail, Docs, and Sheets integration is what AI Pro sells, so decide whether that’s the part you need.

    Should You Pay for More Than One?

    For most people, no. Two $20 subscriptions is $480 a year, and the overlap is large. A better pattern is one paid plan for your main workflow and the free tier of another for second opinions. Common combinations:

    • Workspace user: Google AI Pro, plus ChatGPT Free for images and voice.
    • Developer: Claude Pro for Claude Code, plus ChatGPT Free for quick general questions.
    • Generalist: ChatGPT Plus, plus Claude Free Projects for long-document work.

    The exception is if AI is central to your job and you regularly hit limits on your main plan. Then the $100 power tiers, or a second paid plan, can make sense. Track how often you actually hit a limit over two weeks before upgrading.

    How to Decide in One Week

    1. Write down the three tasks you’d use AI for most often. Be specific: “summarize client email threads”, not “productivity”.
    2. Run those exact tasks through the free tier of all three, with the same prompt.
    3. Note which answer you’d use without editing and which tool made you copy-paste the least.
    4. Pay for the one that won two out of three. Re-check in three months, because plans change.

    Q&A

    Which is cheapest?

    All three have usable free tiers. Among paid plans, Google AI Plus ($4.99) and ChatGPT Go ($8) are the low-cost options; the main tiers are all about $20/month.

    Which is the smartest model?

    It changes with every release, and it depends on the task. For everyday work, where the assistant fits into your workflow matters more than the latest benchmark.

    Are these prices the same everywhere?

    No. Prices, taxes, and some features vary by country. Check the pricing page while signed in to see your local price.

    Can I switch later?

    Yes. All three are month-to-month on monthly billing. Your chat history stays with each service, so export anything you need before cancelling.

    Sources