Cloud Run Billing Modes Compared, With Real us-central1 Prices

Published by Rushikesh Dhule, LessBill · November 18, 2026 · Prices checked on this date
Every Google Cloud Run service has a billing setting that most teams never touch: request-based (the default) or instance-based. Pick the wrong one and the same container costs roughly a third more than it needs to.
The short answer: request-based billing is cheaper when an instance is busy less than about 70-75% of the time. Above that, instance-based billing wins. The rest of this post shows where that number comes from, using the list prices on Google's pricing page today, and when the rule does not hold.
What each setting actually bills
Google's billing settings documentation describes the two options. You may know them by their old names:
- Request-based billing (formerly "CPU only allocated during request processing"). You pay for CPU and memory only while an instance is starting, shutting down, or handling at least one request. You also pay a per-request fee. Between requests, CPU is throttled.
- Instance-based billing (formerly "CPU always allocated"). You pay for the whole life of each instance, from start until it is shut down, with a minimum of one minute. There is no per-request fee, and the per-second rates are lower. CPU stays available after a response is sent, so background work keeps running.
Two details catch people out. First, with request-based billing, an instance handling 50 concurrent requests is billed once for that time, not 50 times; concurrency is what makes the default cheap for busy APIs. Second, Cloud Run jobs, and any service that uses a GPU, always use instance-based billing, so this choice only applies to ordinary services.
The prices side by side
These are the default (no discount) rates for Tier 1 regions such as us-central1 (Iowa), taken from the Cloud Run pricing page on November 18, 2026:
| Item (us-central1) | Request-based | Instance-based |
|---|---|---|
| CPU, per vCPU-second | $0.000024 (active) | $0.000018 |
| Memory, per GiB-second | $0.0000025 (active) | $0.000002 |
| Idle minimum instance | $0.0000025 per vCPU-s + $0.0000025 per GiB-s | Same rate as above (always billed) |
| Requests | $0.40 per million | None |
| Monthly free tier | 180,000 vCPU-s, 360,000 GiB-s, 2M requests | 240,000 vCPU-s, 450,000 GiB-s |
So for a 1 vCPU, 1 GiB instance, request-based billing costs $0.0000265 per busy second, and instance-based costs $0.000020 per second whether busy or not. Instance-based is about 25% cheaper per second, but you pay for every second.
Worked example: one service, two traffic shapes
Take a service with 1 vCPU and 1 GiB in us-central1. A 730-hour month is 2,628,000 seconds. I leave out the free tier here because it is shared across your whole billing account, so it rarely belongs to one service.
Instance-based, one instance up all month: 2,628,000 × $0.000020 = $52.56.
Shape A: an internal API, busy 30% of the time, 5 million requests a month.
- Busy seconds: 2,628,000 × 0.30 = 788,400
- Compute: 788,400 × $0.0000265 = $20.89
- Requests: 5 × $0.40 = $2.00
- Request-based total: $22.89, which is $29.67 less than instance-based.
If you add one minimum instance to avoid cold starts, the 70% idle time (1,839,600 seconds) is billed at the idle rate of $0.000005 per second, or $9.20. The total becomes $32.09, still well under $52.56. A warm instance on request-based billing is cheap; on instance-based billing it costs full price.
Shape B: a steady backend, busy 90% of the time, 25 million requests a month.
- Busy seconds: 2,628,000 × 0.90 = 2,365,200
- Compute: 2,365,200 × $0.0000265 = $62.68
- Requests: 25 × $0.40 = $10.00
- Request-based total: $72.68, versus $52.56 on instance-based. Switching saves $20.12 a month, or 28%, for each instance like this.
The break-even rule
Ignoring the request fee, the two settings cost the same when busy time × $0.0000265 = total time × $0.000020. That gives a break-even of 0.000020 ÷ 0.0000265 = 75.5% busy time. The CPU-to-memory ratio barely moves this: at 1 vCPU and 512 MiB it is about 75.3%.
The request fee pulls the break-even down. One fully used 1 vCPU / 1 GiB instance costs $69.64 a month on request-based compute rates, so every 10 million requests per instance per month ($4.00) lowers the break-even by about 6 points, to roughly 70%. My working rule:
- Busy under 60%: stay on request-based billing.
- Busy 60-75%: run the numbers with your real request count.
- Busy over 75%, or lots of small requests: switch to instance-based billing.
"Busy" means the share of instance time spent handling at least one request. Cloud Run publishes this as the container/billable_instance_time metric in Cloud Monitoring. Compare it with the total time your instances were up.
How to check and change the setting
See which setting a service uses (the annotation run.googleapis.com/cpu-throttling: 'true' means request-based):
gcloud run services describe SERVICE --region us-central1 --format export | grep cpu-throttling
Switch to instance-based billing, or back again:
# instance-based billing (CPU always allocated)
gcloud run services update SERVICE --region us-central1 --no-cpu-throttling
# request-based billing (the default)
gcloud run services update SERVICE --region us-central1 --cpu-throttling
In Terraform, the same switch is cpu_idle = false (instance-based) or true (request-based) in the container's resources block. Before changing anything by hand, look in Recommender: according to Google's cost optimization guide, it reviews the last month of traffic and suggests a move to instance-based billing when that would be cheaper.
Committed use discounts change the maths
Steady services that run all month are also good candidates for commitments. On the pricing page, a 1-year Cloud Run CUD lowers instance-based CPU from $0.000018 to $0.00001494 (17% off), and a 3-year Compute Flexible CUD lowers it to $0.00000972 (46% off). Request-based rates only get 17% off, even with a 3-year flexible commitment. So if you plan to commit, instance-based billing gets cheaper relative to request-based, and the break-even moves lower. We covered how commitments work across clouds in our guide to Savings Plans, Reserved Instances and Spot.
When this advice does not apply
- Spiky or bursty traffic. Instance-based instances stay up (and billed) while idle until Cloud Run scales them in, which the docs say can take up to 15 minutes. If your traffic comes in short bursts, the idle tail can erase the lower rate. Google itself recommends request-based billing for sporadic traffic.
- Background work. If your app runs goroutines, async tasks or an OpenTelemetry agent after sending a response, request-based billing throttles that CPU. You may need instance-based billing for correctness, whatever the price.
- Tier 2 regions. Places such as London, Frankfurt and Singapore have higher rates. The ratio stays similar, but recalculate the dollar amounts before you trust them.
- Low concurrency. With concurrency set to 1, every request needs its own instance time, so busy share and cost both climb quickly. Raising concurrency often saves more than changing the billing setting.
If you also run functions on AWS, the same "pay per request or pay for capacity" question appears there; see our breakdown of where Lambda bills really come from.
The bottom line
Request-based billing is the right default for most Cloud Run services because idle time is free and concurrent requests share one bill. Once a service is busy more than about three quarters of the time, or handles tens of millions of small requests, instance-based billing is usually 20-30% cheaper. Check the billable instance time metric, run the arithmetic above, and change one flag.
How LessBill can help
LessBill.org tracks spend across AWS, Azure and GCP and flags services whose usage pattern suggests a cheaper configuration. You can see the opportunities in one place without opening each console.
Sources
- Google Cloud: Cloud Run pricing (rates, free tier, CUDs, billable instance time; checked November 18, 2026)
- Google Cloud: Billing settings for services (setting names, gcloud flags, traffic guidance)
- Google Cloud: Best practices for cost-optimized Cloud Run services (Recommender, GPU billing)
- Google Cloud: Cloud Run metrics (container/billable_instance_time)
Comments
Post a Comment