Azure SQL serverless cost: when auto-pause beats provisioned

Quick answer

Azure SQL Database serverless bills compute per vCore-second (about $0.000145 per vCore-second, roughly $0.52 per vCore-hour) and auto-pauses after an idle delay so you pay only for storage while paused. That makes it cheaper than provisioned vCore for intermittent or unpredictable workloads that sit idle much of the day. Provisioned General Purpose bills a flat rate whether used or not, roughly $370 per month for 2 vCores, so it is cheaper for databases active more than about 25 to 30 percent of the time. The rule: serverless for spiky and idle-heavy databases, provisioned for steady ones.

Azure SQL Database offers two compute models within the vCore tier: provisioned, which reserves a fixed number of vCores and bills a flat hourly rate, and serverless, which scales compute between a minimum and maximum and bills per second of use, pausing entirely when idle. The choice hinges on one question: how much of the time is the database actually doing work?

How serverless bills

Model	Billing basis	Idle cost
Serverless	Per vCore-second used	Storage only when paused
Provisioned GP	Flat per vCore-hour	Full compute rate, always

Serverless charges roughly $0.000145 per vCore-second, about $0.52 per vCore-hour, scaled by the vCores actually in use at each moment. When the database has no activity for the configured auto-pause delay (as low as one hour), compute stops billing entirely and you pay only for storage at about $0.115 per GB per month. Provisioned General Purpose, by contrast, bills a flat rate around $370 per month for 2 vCores regardless of whether a single query runs.

Where serverless wins

Serverless is cheaper whenever the database is idle for meaningful stretches: development and test databases that no one touches overnight or on weekends, internal tools used a few hours a day, and workloads with unpredictable, bursty traffic. A dev database active four hours a day on 2 vCores costs roughly 4 vCore-hours times $0.52 times 30, about $62 per month plus storage, versus roughly $370 per month provisioned. The auto-pause is what makes the difference, and it is the same idle-elimination logic behind Aurora Serverless v2 on AWS.

Where provisioned is cheaper

The crossover comes when the database is busy enough that serverless never gets to pause and often runs near its maximum vCores. A production database active most of the working day, or one that serves steady round-the-clock traffic, will accumulate serverless vCore-seconds that exceed the flat provisioned rate. As a rough guide, once a database is genuinely active more than about 25 to 30 percent of the time, provisioned General Purpose becomes cheaper, and a reservation on provisioned compute cuts it further. The broader tier tradeoff is covered in the DTU versus vCore guide.

Tuning serverless to save more

Two settings drive the serverless bill. The auto-pause delay controls how quickly an idle database stops billing compute; a shorter delay saves more but adds a warm-up latency on the next connection, so match it to how tolerant the workload is of a cold start. The minimum and maximum vCore range sets the floor and ceiling: a lower minimum saves during quiet periods, a higher maximum handles spikes but raises the peak rate. For workloads that cannot tolerate the resume latency at all, provisioned is the safer choice despite the higher idle cost.

Choosing the model on cost

Estimate the database's active hours per day and its vCore needs during those hours, then compare the serverless vCore-second total against the flat provisioned rate. Idle-heavy databases favor serverless; steady ones favor provisioned plus a reservation. Price both models against the resource catalog before you deploy so the crossover is a known number rather than a discovery on the first full month's bill.

FAQ

How does Azure SQL serverless bill?

It bills compute per vCore-second used, roughly $0.000145 per vCore-second or about $0.52 per vCore-hour, scaled by the vCores active at each moment. When the database is idle for the configured auto-pause delay, compute stops billing entirely and you pay only for storage at about $0.115 per GB per month. This makes it cheap for databases that sit idle for meaningful stretches.

When is Azure SQL serverless cheaper than provisioned?

When the database is idle for meaningful periods, such as dev and test databases overnight and on weekends, internal tools used a few hours a day, or bursty unpredictable workloads. A database active four hours a day on 2 vCores costs roughly $62 per month on serverless versus about $370 provisioned. The auto-pause eliminating idle compute is what drives the saving.

When is provisioned Azure SQL cheaper than serverless?

When the database is active enough that serverless rarely pauses and often runs near its maximum vCores. As a rough guide, once a database is genuinely active more than about 25 to 30 percent of the time, the accumulated serverless vCore-seconds exceed the flat provisioned rate, and a reservation on provisioned compute widens the gap further in provisioned's favor.

What is the auto-pause delay in Azure SQL serverless?

It is the idle period after which serverless stops billing compute, configurable as low as one hour. A shorter delay saves more by pausing sooner, but adds a warm-up latency on the next connection when the database resumes. Match the delay to how tolerant the workload is of that cold start; latency-sensitive workloads may prefer provisioned to avoid resume delay entirely.

How does C3X help with Azure SQL cost?

C3X prices Azure SQL Database configurations from Terraform before you deploy, so the cost of serverless versus provisioned vCore, including the vCore range and tier, is visible in the pull request. That lets you compare the two compute models against your expected active hours at design time and choose the cheaper one before the first monthly bill arrives.

What to do next

Compare Azure SQL serverless and provisioned before you deploy. C3X reads your Terraform and prices your resources against a live catalog. Start with the quickstart.