Summary

  • Under Vercel’s default pricing, MISS and STALE incur transformation and cache-write usage. A HIT fetched from the shared global cache incurs read usage; a recently accessed copy in the same regional cache avoids that particular read cost.
  • Allowed widths, qualities and formats shape the variants a site can request. Longer caching needs a replacement procedure, while bypassing optimisation can increase data transfer.

The photograph has finished its sitting. Its employment prospects are excellent. A phone wants one width, a larger display another, and the application has opinions about quality. In the editorial library it remains one photograph. In Vercel’s image service, the requests can give it several working lives.

That is why counting pictures makes a poor standalone budget. Vercel’s default image-optimisation pricing separates transformations, cache writes and cache reads. Delivery adds Fast Data Transfer and CDN Requests. Reusing a finished image saves a specific job. It does not persuade every other job to resign.

A hit has more than one way home

Vercel’s pricing documentation measures retrieval from the shared global cache in 8KB read units. It explicitly says this is not billed on every cache HIT. An image accessed recently, several hours earlier, in the same region is cached there and does not incur that particular read cost.

The distinction is where the ready copy comes from. A high hit rate alone cannot supply the global-cache read total: a regional copy and a retrieval from shared storage both fit the successful outcome. Delivery still has its separately measured resources. The word cache describes useful reuse, rather than an exemption for the entire journey.

On MISS, the service fetches, transforms, caches and serves the image. On STALE, it serves the cached version while revalidating in the background. The documentation assigns transformations and cache writes to both states; writes are also measured in 8KB units. A visitor can see a familiar picture while fresh work happens behind it. A calm shop window is a poor witness to an idle workshop.

Regional use matters to this explanation without supplying an automatic geographic surcharge. Which variants are repeatedly requested, where, at what size and through which retrieval path determines the work observed. “More international visitors” is a question to investigate, rather than a ready-made percentage for the forecast.

The variants are ordered in the design meeting

The service description includes width, quality, project and the normalised Accept header in distinguishing cached results. Local images use a content hash; remote images use their absolute source URL. One recognisable photograph can therefore have several requested outputs.

There is a worthwhile service here. A small screen need not receive the same oversized file as a large one. The commercial choice is which versions earn their place. Allowing another width creates another possibility; it does not eagerly manufacture every possible combination. Actual requests bring the work into being.

This also explains why an unchanged picture library can accompany changing usage. A revised page component can request different widths or qualities from the same originals. Editorial has added nothing, yet the technical order has changed. An audit confined to the library misses the meeting in which some of the spending was authorised.

Vercel’s usage-management guide recommends reviewing allowed formats, qualities, sizes and source patterns. Limiting the range of variants and shrinking a resulting file act on different parts of the problem. One narrows possible transformations; the other can reduce data read, written and transferred. A quality reduction must also be judged in the delivered picture. Saving the detail a buyer needed to inspect is a remarkably diligent way to optimise the wrong result.

Longer reuse needs an exit for the old picture

For images that rarely change, a longer cache lifetime reduces repeated transformations and writes. The saving is avoided rework. The same decision has another owner on the day a picture needs correcting.

Replacing a remote source at the same URL does not immediately replace an already cached output. The previous version remains until expiry or invalidation. Remote TTL uses the larger of the upstream max-age and minimumCacheTTL. Redeploying the application alone does not invalidate the image cache. For local files, changing the content and redeploying can change the content hash; explicit purge methods also exist.

That lifetime is separate from recent regional residency. A long TTL does not promise an equally long period with a nearby copy and no global-cache reads. Joining those rules gives one setting control over a second mechanism it does not specify.

The operating decision is who replaces the image and how they check what the visitor receives. An upload receipt closes an internal action. For a corrected product picture, the version on the customer’s screen completes the job. A proposal to keep results longer should include that route out.

Switching it off still leaves a delivery

For files that gain little from optimisation, such as some very small images and vector graphics, Vercel recommends selective bypassing. Those images avoid the associated transformations and image-cache reads and writes. The same guide warns that bypassing everything can increase Fast Data Transfer. A larger original sent directly remains a larger delivery. Spending is quite capable of moving to the neighbouring column.

Before applying the calculation, establish the team’s actual pricing basis. The legacy-pricing document counts unique source images requested during the billing period. Certain Enterprise teams created before 18 February 2025 can retain that basis until their contract expires or move to the default model. The choice is not a general entitlement for every current team.

A particular operator’s net saving needs its contract, variant sizes, regional requests and measured resource usage. The public rules already support a better purchase: distinguish making an image, retrieving the finished result and delivering it. Counting originals remains a useful library task. Appointing that number to explain the whole distribution business is an ambitious promotion.