Vibe coding: what it is, the tools, and what an hour of agent work costs

Vibe coding is a development style where you describe the task in words and an AI agent writes the code. A prototype in an evening is real, but so is code nobody has read and a token bill that grows with every agent step. Below: what it is, which tools to use, what you're risking, and what an hour of agent work costs, with a formula and limits.

Vibe coding: what it is and where the term came from

The term was coined by Andrej Karpathy, one of the co-founders of OpenAI, in February 2025. The original idea was to give in to the "vibe" completely: don't read the code, accept the agent's changes, and keep going based on results and follow-up requests. Today the word is used more broadly for any development where the agent writes most of the code from a natural-language description.

In practice there are two different modes:

• Pure vibe coding. You don't read the code. Fine for prototypes, personal scripts, throwaway utilities — where a breakage is cheap.
• Development with an agent. The agent writes and runs the code, you review the diff and the tests. This is the mode for anything that reaches users.

Tools for vibe coding

All four popular tools work with models through an API and connect to a single Zukko key:

Mind the address: Claude Code adds /v1 itself, so it needs the base URL without it; the others need /v1. For Claude Code, be sure to set ANTHROPIC_DEFAULT_HAIKU_MODEL (for example, claude-sonnet-5): background requests go to the Haiku-slot model (by default claude-haiku-*), which is not in the Zukko catalog, and the gateway answers an unknown model name with 404 model_not_found. The format difference matters for money too, more on that below; details in the Anthropic or OpenAI format docs↗.

How an agent works: it reads files, runs commands, edits code, runs tests, and repeats the loop until the task is done. Every step is a separate model request carrying the whole accumulated context: system prompt, tool definitions, history, contents of the files it has read.

Risks of vibe coding

• Unread code. Bugs, vulnerabilities, secrets in commits. Anything going to production needs a person who reads the diff.
• Commands on your machine. The agent can delete files, rewrite git history, push. Work in a separate branch, commit before each task, and don't auto-approve dangerous commands.
• Access to secrets. The agent reads project files, including .env. Give it a separate API key, not your main one.
• Loops. The agent fixes one test, breaks another, and goes in circles. Every lap means requests with a large context, and you pay for all of them.
• Growing context. The longer the session, the more each next step costs. New task, new session.

What an hour of agent work costs

Take an example: the agent makes 60 requests an hour, roughly one step a minute. Each request has 50,000 input tokens: 45,000 read from cache, 3,000 new ones written to cache, 2,000 as regular input; the response is 1,000 tokens. This is an assumption for the calculation; check your own numbers in the Usage section of the dashboard.

Cost per request (prices are per 1M tokens, hence the division by 1,000,000):

(fresh input × input price + cache read × read price + cache write × write price + output × output price) / 1,000,000

Prices as of October 7, 2026:

Notes on the table:

• GPT doesn't bill cache writes separately, so the 3,000 new tokens are counted at the input price together with the 2,000 fresh ones: 5,000 × 0.27.
• For Claude, the cache discount applies only in the Anthropic format, which is how Claude Code works. If Claude is connected through the OpenAI format (/v1/chat/completions), all input is billed at the input price — that's the "no cache" row. For GPT, caching shows up in both formats.
• The choice of model moves the hourly cost the most: with the same request profile, an hour on Fable 5.1 costs $6.405, and on Sonnet 5 it costs $0.393.

As a reference for model strength, the catalog has an intelligence index (the Artificial Analysis Intelligence Index): Claude Opus 5.5 — 58, Claude Fable 5.1 — 53, GPT-6.1 Sol — 52. Make the final choice on your own tasks: give two models the same assignment and compare the result and the charge. A detailed walkthrough is in the article on agent cost with caching↗.

The amounts are an example calculation at one day's prices. For the current list and prices, call GET https://api.zukko.pro/v1/models or open the Models page in the dashboard.

How not to burn your budget

1. A separate key for the agent, with limits. Every Zukko key has a daily and monthly dollar limit. Once it's used up, you get a 429 with code key_daily_limit_reached or key_monthly_limit_reached, and the agent stops. A looping agent on Sonnet 5 without cache would spend 8 × $1.32 = $10.56 over an 8-hour night — a $3 daily limit cuts that off much earlier.
2. Allowed models. In the key's allowed_models, keep only the models you're willing to pay for. Requests to the rest get a 403 model_not_allowed_for_key instead of a surprise Fable bill.
3. Cache. Use the Anthropic format for Claude. Don't change the start of the prompt mid-session: caching works on an identical prefix.
4. Short sessions. One task, one session. A long history gets more expensive with every step.
5. Check actual spend. See real charges in the Usage section of the dashboard or via the API.

Limit setup is described in the key limits and cache docs↗, and what to do if the agent does go into a loop is covered in the article on a looping agent↗.

To get started, create a separate key for the agent in the dashboard, set a daily limit, and connect the tool of your choice.

FAQ

Can I vibe code without knowing how to program?

For a prototype or a personal utility, yes. For a product other people use, you need someone who reads the code and understands what the agent did.

Which tool should I choose?

It depends on your habits. If you live in the terminal, Claude Code or Codex CLI; if in an editor, Cursor or Cline. They all connect to one Zukko key, so you can try several.

Why does Claude cost more in one tool than in another?

Usually because of the format. The gateway gives Claude's cache discount only in the Anthropic format. Tools that go through the OpenAI format pay the input price for all input.

How much does vibe coding cost per month?

Multiply the agent's working hours by the hourly cost. Using the example above, 40 hours on claude-sonnet-5 with cache is 40 × $0.393 = $15.72; on gpt-6.1-sol it's 40 × $0.1953 = $7.81. Your request profile may differ, so check against the stats in the dashboard. A detailed estimate for Claude Code is in the article how much Claude Code costs per month↗.