Decide
Economics
What this work actually costs. Everything here is computed from the model pricing table and the tool prices this site already keeps — nothing is copied, so nothing carries its own source: the source is the row the figure came from.
Computed over 21 current text models and 18 tools with a published monthly price.
Is the subscription worth it
A subscription is one figure a month, an API is billed per call — they cannot be compared until both are on the same scale. That scale here is calls per working day at which the two even out. Below that line the API is cheaper, above it the subscription is.
Computed on a mid-priced model (Claude Sonnet 4.6), twice over. A chat turn is 4,000 input and 800 output tokens — $0.024 a question. Working with a document is 50,000 and 5,000 — $0.225, an order of magnitude more.
| Tool | Monthly | Chat turns a day | Document calls a day |
|---|---|---|---|
| Gemini | $4.99 | 9.5 | 1.0 |
| Codex | $8.00 | 15 | 1.6 |
| GitHub Copilot | $10.00 | 19 | 2.0 |
| Zed | $10.00 | 19 | 2.0 |
| Jules | $19.99 | 38 | 4.0 |
| Microsoft Copilot | $19.99 | 38 | 4.0 |
| Claude | $20.00 | 38 | 4.0 |
| Claude Code | $20.00 | 38 | 4.0 |
| Cursor | $20.00 | 38 | 4.0 |
| Devin | $20.00 | 38 | 4.0 |
The gap between the last two columns is the whole argument. Someone using a plan for conversation needs dozens of questions a day to break even; someone using it on documents breaks even in a handful of calls. And the arithmetic only holds while you stay inside the plan's limits — which are counted in messages and hours rather than tokens, and that is where the comparison stops.
What one job costs
A price per million tokens is a measure nobody feels. These are four jobs people actually do, each with a different balance of reading and writing — which is exactly where models diverge.
| Model | Translating a page | Reviewing a code change | An 800-word article | Summarising a long document |
|---|---|---|---|---|
| GPT-5.6 LunaOpenAI | $0.0019 | $0.0034 | $0.0043 | $0.025 |
| Gemini 2.5 FlashGoogle | $0.0038 | $0.0062 | $0.0073 | $0.038 |
| Gemini 3 FlashGoogle | $0.0048 | $0.0085 | $0.0108 | $0.0624 |
| Gemini 3.7 FlashGoogle | $0.0061 | $0.0116 | $0.0154 | $0.093 |
| Claude Opus 4.5Anthropic | $0.0405 | $0.0775 | $0.1025 | $0.620 |
| Claude Fable 5Anthropic | $0.081 | $0.155 | $0.205 | $1.240 |
| Claude Mythos 5Anthropic | $0.081 | $0.155 | $0.205 | $1.240 |
| GPT-5.6 CyberOpenAI | $0.1188 | $0.2125 | $0.270 | $1.560 |
- Translating a page
- 800 words in, the translation out. The target language usually costs more tokens than English for the same content. 1,100 → 1,400
- Reviewing a code change
- A diff of around 500 lines with the surrounding files in context, and a review with comments out. 8,000 → 1,500
- An 800-word article
- A handful of sources read in, a finished text out. This is the job this site does every day. 15,000 → 1,100
- Summarising a long document
- A contract or report of a hundred-odd pages in, a page of conclusions out. 120,000 → 800
An hour of agent work
An agent does not ask once, it asks twenty times, and each step re-reads what it already knows. The input is large, the output small, and this is where prompt caching stops being a detail — the gap between the two columns below is the whole bill, not a discount on it.
| Model | No cache | With cache | Difference |
|---|---|---|---|
| GPT-5.6 LunaOpenAI | $0.1392 | $0.042 | 3.3× |
| Gemini 2.5 FlashGoogle | $0.22 | $0.0742 | 3.0× |
| Gemini 3 FlashGoogle | $0.348 | $0.105 | 3.3× |
| Gemini 3.7 FlashGoogle | $0.51 | $0.1455 | 3.5× |
| Claude Opus 4.5Anthropic | $3.40 | $0.97 | 3.5× |
| Claude Fable 5Anthropic | $6.80 | $1.94 | 3.5× |
| Claude Mythos 5Anthropic | $6.80 | $1.94 | 3.5× |
| GPT-5.6 CyberOpenAI | $8.70 | $2.625 | 3.3× |
A dash instead of a figure means the vendor has not published a cached input price, not that caching is absent. For those models only the upper bound is shown.
What is assumed
Model and tool prices are copied from vendor pages and carry a check date; everything else here is an estimate, and these are the estimates. A token is about 0.75 of an English word, so 800 words come to roughly 1,067 tokens. A month has 22 working days. An agent hour is 20 steps of 30,000 input and 800 output tokens each, with 90 percent of the input identical between steps. Your own work will differ — the order of magnitude will not, and that is what this table is read for.