How many neurons does a Workers AI request use?
Link to section: How many neurons does a Workers AI request use?About 35, for a short customer question on Llama 3.1 8B Instruct, the model in Cloudflare's own Workers AI quickstart. Cloudflare bills every request in , its unit for the computing a model does, at $0.011 per 1,000 after the first 10,000 each day. So the free allowance covers roughly 280 of those requests a day, and a business sending 1,000 a day pays about $8.36 a month.
Those two numbers, how far the free neurons go and what the month costs once they run out, are what you need before you put an AI feature in front of customers. Cloudflare publishes the per-model rates but no calculator, so the one below does the arithmetic from those rates.
- Workers AI gives every account 10,000 neurons a day free, reset at 00:00 UTC, then charges $0.011 per 1,000.
- A 500-token question with a 300-token answer uses from about 7 neurons on Llama 3.2 1B to about 75 on Llama 3.3 70B.
- On every text model below, an output token uses more neurons than an input token, so answer length drives the bill.
- A $50 monthly cap pays for about 128,000 of those requests on Llama 3.1 8B Instruct, on top of the free ones.
Work out your own neuron bill
Link to section: Work out your own neuron billEnter the length of a typical question and answer in tokens (a token is a piece of a word; 300 tokens is roughly 225 English words), how many requests you expect a day, and the most you want to spend in a month. The card shows the neurons per request, how many requests the free allowance covers, the 30-day bill and whether your cap lasts the month.
Neurons per request, and what a month costs
Pick a model, set a typical question and answer length, and enter your traffic and spending cap.
- Each request uses
- 35.3 neurons
- Free allowance covers
- 282 a day
- 30-day bill at your traffic
- $8.36
- Your cap
- Lasts the month
Rates from Cloudflare's Workers AI pricing page, checked 4 November 2026: $0.011 per 1,000 neurons after 10,000 free a day, reset at 00:00 UTC. The bill assumes the same traffic every day for 30 days. Token counts are your estimate. Every model listed charges more neurons per output token than per input token.
What a short request costs on each model
Link to section: What a short request costs on each modelMethod: each row applies Cloudflare's published neuron rates for that model, checked on 4 November 2026, to one request with a 500-token question and a 300-token answer. The token counts are our assumption for a short customer-service exchange, not a measurement. Cloudflare also lists a dollar rate per million tokens for each model, and both columns give the same cost to within a tenth of a cent per 1,000 requests.
| Model | Neurons per request | Free requests a day | Cost per 1,000 paid requests |
|---|---|---|---|
| Llama 3.2 1B Instruct | 6.7 | 1,491 | $0.07 |
| Llama 3.1 8B Instruct (fp8-fast) | 12.5 | 798 | $0.14 |
| gpt-oss-20b | 17.3 | 578 | $0.19 |
| Llama 3.1 8B Instruct | 35.3 | 282 | $0.39 |
| Llama 3.3 70B Instruct (fp8-fast) | 74.8 | 133 | $0.82 |
Speech is priced per audio minute instead of per token. Whisper uses 41.14 neurons a minute, so the free allowance covers about 243 minutes of transcription a day.
The spread matters more than any single row. The same customer question costs about eleven times as much on the 70B model as on the 1B model, so test the smallest model that gives acceptable answers before you pay for a larger one.
Why does the answer cost more than the question?
Link to section: Why does the answer cost more than the question?Generating a token takes more computing than reading one, and Cloudflare's rates show it. On Llama 3.1 8B Instruct an output token uses about 2.9 times the neurons of an input token; on Llama 3.3 70B the gap is about 7.7 times (204,805 against 26,668 neurons per million tokens). A chatbot that answers in three paragraphs when one would do is spending most of its neurons on the extra two.
Workers AI text models accept a max_tokens setting that limits
how long a reply can run. Set it to the longest answer your feature should
ever give, then enter that number as the answer length above. The
calculator's bill is then a ceiling, not a guess.
What happens when the free neurons run out?
Link to section: What happens when the free neurons run out?On the free Workers plan, requests past 10,000 neurons fail with an error until the allowance resets at 00:00 UTC, according to Cloudflare's pricing page. On Workers Paid, the requests keep running and each 1,000 neurons past the allowance costs $0.011. The free allowance resets every day, so unused neurons do not carry over to the next one.
For a small business that means a choice to make before launch: stay on the free plan and accept that a busy day stops the feature, or move to Workers Paid and set a spending cap you are comfortable paying. The calculator above shows which day a cap would run out at your traffic.
Frequently asked questions
Link to section: Frequently asked questionsIs a neuron the same as a token?
No. A token is a piece of text going into or out of a model. A neuron is Cloudflare's measure of the computing a request needs. Each model has its own rate for how many neurons a million input or output tokens use, which is why the same question costs different amounts on different models.
Does the calculator include the Workers Paid subscription?
No. It covers neuron charges only. Check Cloudflare's Workers pricing for the plan fee, which is billed separately from Workers AI usage.
Why might my bill differ from the estimate?
Real questions and answers vary in length, and the calculator assumes the same traffic every day. Compare the first week of your account's actual neuron use with the estimate, then adjust the token counts until they match.
Will these rates change?
They can. The figures here were checked against Cloudflare's pricing page on 4 November 2026, and new models arrive with their own rates. The link in the calculator's note goes to the current page.
Where this fits
Link to section: Where this fitsPer-request pricing is the reason Cloudflare's AI bet carries so little risk for Cloudflare, and the reason it suits a small business testing its first AI feature. Our Google vs. Cloudflare AI adoption piece sets that pricing against Google's choice to cut into its own search ad revenue. If you want help scoping an AI feature with a cap that fits, tell us what you are trying to build.
Sources: Cloudflare Workers AI pricing; Workers AI REST API quickstart.
