The simple and definitive guide to choosing the right LLM for the job.
The one question that matters.
Before comparing features, ask this: does the task need words, or does it need action?
“Write me a report.”
“Build me a tool.”
Everything else is detail.
The four players, plain English.
The most well-known, broadest general-purpose assistant. Good at a bit of everything — writing, research, brainstorming, data analysis in a pinch.
Built with a strong emphasis on careful reasoning and following complex instructions precisely. The assistant professionals reach for when the task is long, nuanced, or has to be exactly right.
Deeply wired into Google's ecosystem (Docs, Sheets, Gmail, Search). Also ships “Antigravity,” an agentic coding environment that can even run other companies' models inside it.
Built into X (Twitter), positioned as fast, current, and a little less filtered. Good for real-time information and commentary.
Three jargon terms worth actually understanding.
Every AI has a limit on how much text it can hold in its head during one conversation, measured in “tokens” (roughly: pieces of words). A small context window means it forgets the start of a long document by the time it reaches the end. A large one means it can hold an entire contract, codebase, or research report at once.
Some models can be told to slow down and show their work before answering — checking logic, catching contradictions, double-checking math — instead of just generating the first plausible-sounding response.
A chat assistant writes words back to you. An agent can actually take actions: browse the web, edit files, run code, use other software, and check its own output along the way.
Stop picking a favorite AI. Pick based on the job.
Words vs. action, short vs. long, routine vs. high-stakes. Answer a few questions and we’ll point you at the model that matches what you’re actually asking it to do.
Does the task need words, or does it need action?
They move as you answer.
How token charges work, in plain English.
When you build on an AI model directly (through its API) you don’t pay a flat fee. You pay by the token, a chunk of text about three-quarters of a word long. A thousand tokens is roughly 750 words; a 40-page proposal is around 25,000–30,000. Every request runs two meters at once.
Your question, plus your instructions, any files or documents, and the entire conversation so far. The model has no memory between requests, so the whole history gets sent again every time.
The answer, the code, and any behind-the-scenes “thinking” it does first. Output costs about five times more per token than input, so long answers and heavy thinking add up fastest.
Turn six re-sends turns one through five. You pay for the whole stack again each time, which is why a long, meandering session costs far more than several short, focused ones. Most providers now cache repeated content and charge roughly a tenth of the normal price to re-read it, which helps a lot but doesn’t make it free.
Price per million tokens on Anthropic’s API, October 2026. OpenAI, Google, and xAI use the same per-token model with their own rates; the pattern (cheap fast models, pricier smart ones, output costing more than input) holds everywhere.
Rough illustrations at the rates above, before caching discounts.
The takeaway: for most small-business agents, the model bill is measured in cents per task. The real cost is building it well, and choosing the smallest model that does the job.
Vibe coding, and why it eats tokens.
Vibe coding means building software by describing what you want in plain English and letting an AI agent write, run, and fix the code. You judge the result by whether it works, not by reading every line. AI researcher Andrej Karpathy coined the phrase in early 2025, and tools like Lovable, Claude Code, and Gemini Antigravity are built around it.
It’s agent mode applied to software, and it changes the token math. One request like “add a login page” isn’t one model call. It’s a loop that can run 10 to 50 times, and every pass re-sends your code and the conversation so far.
- 1Read the files+ tokens
- 2Plan the change+ tokens
- 3Write code+ tokens
- 4Run & test+ tokens
- 5Read the errors+ tokens
- 6Fix and repeat+ tokens
About a million input tokens and ~40,000 output on Sonnet 5.5.
“Still broken.” “Try again.” Six million input tokens of the same files, re-read over and over.
Building straight on the API. You see the meter, you pay exactly for what you use. Best for agents you run for customers.
Claude Code and similar tools on a plan: a set amount of usage per time window. Predictable, but heavy sessions hit the ceiling and pause.
Lovable and other app builders: each message costs credits based on how complex it is. Credits are tokens wrapped in a simpler unit.
- →Be specific up front. A vague request means rework, and rework means more loops.
- →Build in small steps. One feature per request, checked before the next.
- →Start fresh per feature. A new session drops the history you'd otherwise keep paying for.
- →Stuck twice? Stop. Re-describe the problem instead of typing “fix it” a third time.
- →Keep a project brief. A short file of how the project works saves the agent re-discovering it every session.
- →Match the model to the job. Small edits don't need the most expensive model or maximum thinking.
Lovable vs. Claude Code: when to use which.
Both turn plain English into software. The difference is how much they decide for you. Lovable makes the choices and hands you a running app; Claude Code gives you the full toolbox and expects you to drive.
You chat, it builds a working web app with a database, sign-ins, and hosting already set up. You never see a terminal.
- ✓Prototypes, demos, and MVPs to show a client this week
- ✓Internal tools, portals, simple dashboards, landing pages
- ✓Non-developers, or anyone who wants zero setup
- !Credits burn fast in debugging loops
- !Complex logic, back-end jobs, and unusual integrations get hard
- !You work inside its stack (React web apps)
An agent that works on a real codebase on your machine: any language, any stack, reading and changing files, running tests, using your tools.
- ✓Real software: integrations, automations, AI agents, data jobs
- ✓Existing codebases, and projects that outgrew a builder
- ✓Anything that has to run somewhere other than a web page
- !You handle setup, hosting, and deployment yourself
- !Helps to know the basics (Git, how deploys work)
- !More power means more room to make a mess
Rule of thumb: if it’s a web app a client will click around in, start with Lovable. If it has to talk to other systems, run on a schedule, or act on its own, start with Claude Code.
Want help putting the right AI to work?
Liquid Technology Solutions helps consultancies and B2B teams pick, wire up, and build with AI — chat, agents, and CRM. Take the PDF with you, or book 30 minutes and bring the task you’re trying to solve.