Code for as little as possible: VS Code, OpenCode and app builders

Read a sample · Chapter 01 of 06

Almost every tool charges in one of four ways. Learn to read them before you compare prices.

AI-assisted coding has two costs: the tool (editor, agent or app builder) and the model that writes the code. The tool is often free; the model is where bills come from.

Scroll sideways to see every column.

Prices move. Google's Gemini API pricing page (source 20, checked on 26 September 2026) lists Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 per million output tokens until 31 December 2026, then $1.50 and $7.50. Limits can be hidden: Google's rate limits are shown in Google AI Studio, not on the page (source 21).

A worked reading. GitHub's Copilot documents (sources 1 to 3, checked on 26 September 2026) say:

• Copilot Free: $0, 2,000 code completions a month, and AI credits of unstated size. Verified students can get the free Copilot Student plan.
• Copilot Pro: $10 USD a month, with 1,000 base credits plus a 500-credit flex allotment that "may change over time". One credit is $0.01, spent on each request's tokens.
• Credits reset at 00:00 UTC on the 1st and do not carry over. When they run out, you upgrade, set a budget for extra use, or wait.

Read a sample · Chapter 02 of 06

Per-token prices look tiny. A month of coding requests may not be. Estimate before you spend.

A token is a small piece of text a model reads or writes: about 4 characters for Google's Gemini models, and 100 tokens are about 60 to 80 English words (source 23). Other models differ; your usage page gives the true count. Prices are per million tokens, in three kinds:

• Input: your prompt plus files, instructions and the conversation so far.
• Cached input: input that qualifies for a cache hit under that provider's rules. Merely sending text recently does not ensure a hit.
• Output: what the model writes, usually dearer per token.

Coding requests are big. v0 counts chat history and source files as input (source 33), and agent features can make several model calls per task (source 3). OpenCode's Go page (source 14) estimates a typical request for one model at 1,000 input, 50,000 cached and 220 output tokens, though usage varies. This workbook borrows that shape until you measure your own.

Build the calculator. Save this as plain text named costcalc.py, open a terminal in its folder and run python costcalc.py (or python3). It sends nothing anywhere.

# costcalc.py: estimate a monthly bill for AI-assisted coding.
# Offline, standard library only. Prices are per 1 million tokens,
# copied from the provider's own page, in that page's currency.


def request_cost(new_in, cached_in, out, p_in, p_cached, p_out):
    return (new_in * p_in + cached_in * p_cached + out * p_out) / 1_000_000


def month(label, per_day, days, tokens, prices, cur="USD"):
    one = request_cost(*tokens, *prices)
    total = one * per_day * days
    print(f"{label:<14} {one:.6f} per request x "
          f"{per_day * days} = {total:.2f} {cur} a month")
    return total


def electricity(watts, hours_a_day, days, pence_per_kwh):
    kwh = watts * hours_a_day * days / 1000
    return kwh, kwh * pence_per_kwh / 100  # kWh and pounds


if __name__ == "__main__":
    per_day, days = 60, 20          # your usage: an assumption
    shape = (1_000, 50_000, 220)    # new input, cached input, output tokens
    no_cache = (51_000, 0, 220)     # the same request, nothing cached
    a = (2.00, 0.20, 10.00)         # Price A, USD per 1M tokens
    b = (0.30, 0.006, 1.20)         # Price B, USD per 1M tokens
    month("A, cached", per_day, days, shape, a)
    month("A, no caching", per_day, days, no_cache, a)
    month("B, cached", per_day, days, shape, b)
    kwh, pounds = electricity(200, 2, 20, 26.32)
    print(f"Local model: {kwh:.1f} kWh, about {pounds:.2f} GBP a month")

With Python 3.12.10 on Windows 11 it printed:

A, cached      0.014200 per request x 1200 = 17.04 USD a month
A, no caching  0.104200 per request x 1200 = 125.04 USD a month
B, cached      0.000864 per request x 1200 = 1.04 USD a month
Local model: 8.0 kWh, about 2.11 GBP a month

The prices, in US dollars per million tokens, show a spread; they are not recommendations. Price A: Claude Sonnet 5 on Anthropic's pricing page (source 26, checked on 26 September 2026): $2 input, $0.20 cache hits, $10 output. Price B: deepseek-flash at peak times on DeepSeek's pricing page (source 24, checked on 26 September 2026): $0.30 input, $0.006 cache hits, $1.20 output.

Check line one by hand: 1,000 x 2 + 50,000 x 0.20 + 220 x 10 = 14,200; divided by a million, $0.0142 a request. Sixty requests on 20 days is 1,200, and 1,200 x 0.0142 = $17.04. The usage is an assumption, not a measurement.

In pounds (this workbook's own arithmetic; your bank sets its own rate): the European Central Bank's reference rates for 25 September 2026 (source 29) were 1.1403 dollars and 0.86045 pounds to the euro, so $1 = 0.86045 / 1.1403 = £0.7546, and $17.04 is about £12.86.

Read a sample · Chapter 03 of 06

A free editor plus a free allowance is a cheap start, if you know the limits and the data terms.

VS Code is Microsoft's code editor. Its FAQ (source 6, checked on 26 September 2026) says it is free for private or commercial use and needs no account for the core editor. The product has a Microsoft licence (source 7); its source code is MIT-licensed. Three ways to add AI, as published that day:

Scroll sideways to see every column.

What leaves your computer. Telemetry is on by default in Copilot's free version; set telemetry.telemetryLevel to off to stop it (source 8). From 24 April 2026, Copilot Free, Pro and Pro+ inputs, outputs and code context train GitHub's models unless you opt out under Privacy in your Copilot settings (source 5). With your own key the provider's terms apply (source 8); a local model can run offline (source 9).

Subscription or own key? GitHub prices Claude Sonnet 5 at Price A's rates (source 4, checked on 26 September 2026), and base credits match the subscription price (source 3). So $10 of base credits buys about what $10 on your own key would. The extra is the flex allotment, unlimited completions and a 10% discount with automatic model selection (source 3).

Read a sample · Chapter 04 of 06

An open-source agent for your terminal. The agent costs nothing; the model behind it might.

OpenCode is an MIT-licensed coding agent for the terminal, desktop or an editor (source 16). Its home page (source 10, checked on 26 September 2026) says it is "100% free to use", with some free models; other providers may cost more. Your options, checked that day:

Scroll sideways to see every column.

Sharing a session makes a public link and copies it to OpenCode's servers; it is manual by default and can be disabled (source 15).

A model on your own computer. Tools such as Ollama (MIT-licensed, source 19) run open-weight models, whose trained weights you can download; OpenCode can connect to one (source 12). Ollama's pricing page says running models on your own hardware is always unlimited (source 17). Unlimited is not free:

• Hardware: the model must fit your memory; the GPU memory maths workbook shows how to check.
• Context: Ollama defaults to 4,096 tokens (source 18); OpenCode suggests 16k to 32k if tool calls fail (source 12), which needs more memory.
• Time: a local model may be slower. Time a real task.
• Electricity: kWh (kilowatt hours) = watts x hours / 1,000, times your price per kWh.

Ofgem's cap for 1 October to 31 December 2026 (source 28, checked on 26 September 2026) averages 26.32 pence per kWh for electricity by Direct Debit on a standard variable tariff in Great Britain, with no VAT. The calculator's last line assumes an extra 200 watts for 2 hours on 20 days: 8 kWh x 26.32 = 210.56 pence, about £2.11. Measure your own machine with a plug-in energy monitor.

Read a sample · Chapter 05 of 06

App builders turn a description into a running, hosted app. You pay for the convenience in credits or tokens.

An app builder generates a whole app from a description and hosts it. Four, alphabetically, checked on 26 September 2026. Retained dollar prices are US dollars; a regional Lovable subscription quote was not confirmed:

Scroll sideways to see every column.

Credits vary per task. Lovable's FAQ (source 31) gives examples: "Make the button gray" 0.50 credits, "Remove the footer" 0.90, sign-up and login 1.20, a landing page with images 1.70. Daily free credits expire each day.

A builder can pay off when you need a hosted prototype fast, would otherwise pay separately for hosting and a database, or value the time saved above the fee. It is weaker as a project grows (Bolt says bigger projects use more tokens per message), when you iterate a lot, or when the code must run elsewhere: check ownership and export. Lovable says you own your code, subject to third-party rights in the models (source 31); check each provider's spending controls and whether they actually stop further charges, including hosting costs.

Wiring it yourself means a free editor, an assistant from chapter 3 or 4, and your own hosting, as in the Build and launch a website with AI workbook.

Read a sample · Chapter 06 of 06

Decide with your own numbers, data and hardware, not a headline price.

These exercises are offline and need no sign-up. If you are under 18, work through them with an adult or teacher and check the age and account conditions of any online service before trying it. A student discount does not by itself establish eligibility to open an account.

1. Name the task and its size: a one-page tool, an app with logins, or a large codebase?
2. Set a monthly cap you could lose without harm.
3. Decide what may leave your computer. If nothing, only a local model fits. Never paste secrets or other people's personal data.
4. Check your hardware if a local model is an option.
5. Price two routes with the calculator, using prices you copied, dated.
6. Check spending controls first: confirm whether a budget actually stops requests or only sends an alert. Turn auto-reload off and set alerts. If there is no enforceable limit within your budget, keep to a route with no paid billing attached.
7. Review after a week against the usage page; change route if needed.

For step 3, the terms read on 26 September 2026 say:

Scroll sideways to see every column.