What a Dollar of AI Buys: One Month Across Claude, Qwen, Gemini and Codex
In 30 days my agents processed 46.9 billion tokens. At list prices that is $15,692. I paid $745. What each plan returned per dollar, and why the cheapest worker still needs a reviewer.
My last two posts priced my Claude usage at list rates ($15,144 in a month) and compared one rival plan against it (Qwen3.8 vs Opus 5.5). This one counts everything I run, on every plan I pay for, and answers the question: what does each dollar actually buy?
The month in one line
In the 30 days to 19 October, my agents on two Macs processed 46.9 billion tokens. At each model's list price that comes to $15,692. I paid $745.
| Plan | What I paid | Tokens | At list prices | List value per $1 |
|---|---|---|---|---|
| Claude Max, 2 seats | $400 | 44.9B | $15,168 | $37.90 |
| Qwen Token Plan, 1 seat | $200 | 1.29B | $414 | $2.07 |
| Gemini API, pay as you go | $103 | 623M | $103 | $1.00 |
| ChatGPT Plus (Codex) | $20 | 1.7M | $5 | $0.25 |
The rest of the $745 is an Ollama subscription I barely touched this month and about $1.50 of one-off API calls to other providers. My local models on Ollama handled another 10 million tokens for nothing.
Subscriptions win by a mile, while they last
Both Claude seats hit their weekly limit every week this month, so $15,168 is close to the most $400 of Claude buys. Per million tokens, that is under one cent. The Qwen seat works out at 15.5 cents and the Gemini API at 16.6 cents. A Claude seat used to its limit delivers tokens about seventeen times cheaper than anything else I pay for.
I explained in the workforce post why that number should not be read as the real price of the work. Flat-rate subscriptions that let a heavy user burn thousands of dollars of compute are customer acquisition, carried by the API. A business that wants this volume without a person at each keyboard pays the per-token column, not mine. The weekly limit is the other catch: a seat gives you a week's worth at a time, and when it runs out you wait or you buy another seat.
The Qwen seat is the opposite shape. Its 250,000 credits are a monthly pot you can spend as fast as you like, and I spent 232,000 of them in 27 hours. It returned about two dollars of list-price work per dollar. Fine value, but a pot, not a plan.
Per token, the gap between models is smaller than the price pages suggest
Every model here spent this month doing the same thing: re-reading its own context. Between 97% and 99% of each model's tokens were cache reads. So I priced each model on its own mix of my tokens rather than on the headline input and output rates:
| Model | List price per million tokens of my work |
|---|---|
| Gemini 4 Argon | 16.6¢ (33.2¢ after 30 November) |
| Claude Sonnet 5.5 | 19.0¢ |
| Claude Opus 5.5 | 30.6¢ |
| Qwen3.8-Max | 31.7¢ |
| Claude Opus 5 | 60.9¢ |
| Claude Fable 5.1 | 62.0¢ |
| Claude Mythos 5.1 | 68.1¢ |
Qwen3.8-Max's output costs less than a third of Opus 5.5's, and on my workload it still comes out slightly dearer, because its cache reads cost 25% more. When using agents, look up the cache-read cost first.
Gemini 4 Argon is the cheapest per token right now, but only because of a launch discount. I checked the bill. Google's billing export lists Argon at $4 per million input tokens, $20 output and $0.20 cached, then takes half of it back with a promotional credit. For the days I checked, my agent framework's own estimate came within 0.5% of what Google actually charged. When the credit ends on 30 November, Argon lands at 33 cents, level with Opus 5.5 and Qwen.
What the money actually bought
Tokens are not work, so here are two jobs with receipts, both run on Gemini 4 Argon in my own agent framework.
On 18 October it worked through 13 leftover items from a code review: 155 model calls, $6.06. Claude reviewed every one before it merged. Six of the thirteen needed follow-up fixes.
On 19 October it fixed four defects in the tool that tracks which agent wrote which commit. It took 31 minutes, cost $2, and came with a test for each fix. The review found three problems in the result.
That is the pattern from the Qwen experiment again. The cheap model writes fast and mostly right. The expensive one catches what would have hurt. At 31 cents per million tokens of list price, Opus 5.5 reviewing two dollars of Gemini work is money well spent.
Where I land
Right now, the cheapest frontier tokens I can buy are Claude subscription tokens, used to the limit, by more than an order of magnitude. That is a subsidy, and I am using it while it lasts. On pay-as-you-go prices, Gemini 4 Argon is the cheapest capable model until its discount ends, after which Sonnet 5.5 is. Opus 5.5 and Qwen3.8-Max cost about the same per token of real work. The older Claude models cost twice as much and no longer earn it.
If I had to pay list price for all of it, I would get a server or Mac Studio that's capable of running a bigger model locally. I would keep Opus as the reviewer and move the bulk writing to the model I'm not paying tokens for.
Honest hedges. List price is not cost: it includes the margin that pays for the next model. ChatGPT Plus barely got used this month, so its figure says nothing about what it is worth. And none of these numbers measure quality.
Token counts are from Claude Code's and Codex's own session logs on two Macs, one count per API request, and from my agent framework's per-call ledger, for the 30 days to 19 October 2026. Both Claude seats share these logs, as I switch accounts on the same machines. Per-token list prices are from OpenRouter on 19 October 2026, with cache writes at Anthropic's 5-minute rate. Claude Mythos 5.1 is priced at Fable 5.1's rate, which Anthropic lists at the same price, and the qwen3.8-max-preview model at Qwen3.8-Max's rate. Gemini spend was checked against Google Cloud's billing export (net of the promotional credit). Plan prices: Claude Max 20x and Qwen Token Plan at $200 a month per seat, ChatGPT Plus and Ollama Pro at $20.