Impact / Economics / Free vs Paid

The Off-Worker Arms

Worker requests bind the serving ceiling, and only a view that never runs the Worker avoids them. Two arms do that. The assets arm publishes a rendered page as a static asset, which the asset layer answers with no Worker. The R2 arm mirrors the page to a bucket behind a custom domain. Both are scored by offWorkerArm() in scripts/measure/free-envelope.ts at the shipping default, and bun scripts/measure/free-envelope.ts prints every row below.

The assets arm is not built. The R2 arm ships, but the canonical wrangler.jsonc binds no bucket, and serving from it needs R2 enabled on the account and a custom domain you bought, since the off-Worker path is a bucket on a custom domain. Every row of the first table is derived from the model, each at 99% of views answered off the Worker, which is every view that is not a miss.

Static asset requests do not bill. The Observability events agree on invocations (measured): a drive on 2026-11-02 sent 198,000 asset requests (3,000 views of 66 paths, one laptop, 151 s) and recorded one fetch event at 0 ms of CPU in that window, against 20,000 control requests to a deny-listed path that each ran the Worker. The one documented exception is Workers Cache, under which assets bill at the Worker request rate. The shipped config has no cache block, so that row describes a configuration the product does not ship. A view is 66 requests, the demo site's first load (n=1), counted as the page and 65 assets.

The Cache Rule row uses an absorption of 7/8, which is the measured hit rate in R2 Page Origin Caching: 7 of 7 hits on a cold object after one miss, and 1 miss, 7 hits and 2 dynamic across 10 requests on a second object, so n is small. Without a rule absorption is 0, and the best mirror share is 77%, which gives 432,900 views a day rather than 336,700. At 30.44 days a month the not-billed rows are 3.04 million views today, 10.2 million for R2 without a rule (13.2 million at the 77% share), 82.0 million for R2 with one and 90.5 million for the assets arm. Read the last two as ceilings of the model, not as traffic any site has served.

What the table says, in order:

• Under the Workers Cache exception, no arm helps. A page that loads its 65 assets through the Worker meter spends 66 requests a view, so the free plan's own ceiling is 1,515 views a day, about 46,000 a month, and the arms move it by 0% (assets) and 1.5% (R2). That figure is a first-load bound and a browser that caches the assets spends fewer; the cache rate has not been measured.
• With assets unbilled, the arms lift the Worker meter and another meter binds. The assets arm reaches the DO-request meter at 2.97 million views a day, because a miss still costs three object requests. The R2 arm reaches its Class B read allowance of 10 million a month first.
• A month is 30 days in the ceilings and 30.44 in the bills. scripts/measure/free-envelope.ts divides a monthly allowance by 30 (DAYS_PER_MONTH), which is how 1,000,000 Class A writes become 33,333 a day. The bills in Economics use DAYS_MONTH = 30.44 in scripts/economics/bill.ts, the same allowance as about 32,852 a day. The conventions differ by 1.5%. Each figure here uses the convention of the script that prints it, and neither is a measurement.
• Mirroring to R2 caps regeneration at a lower ceiling. Every fill writes one R2 object, a Class A operation. The free Class A allowance of 1,000,000 a month is 33,333 a day, which binds before the 50,916 a day that rows allow, so a free-account site cannot regenerate more than 33,333 pages a day with the mirror on. That is a 34.5% lower ceiling. It does not reduce the regeneration a site does: a site under 33,333 a day regenerates the same amount with or without the mirror, and the mirror buys serving headroom in return. The shipped R2_WRITES_PER_MONTH guard stops at 900,000, which is 30,000 a day. On the managed plan, Class A beyond the allowance is billed rather than capped. The paid R2 rate is not recorded in scripts/economics/rates.ts, so it is not stated here. The assets arm writes no R2 object and leaves the regeneration ceiling where it is.

R2 is priced here only from the free-tier operation counts in config/quotas.yml. The R2 storage allowance and every paid R2 rate are missing from scripts/economics/rates.ts and from config/quotas.yml, and are not filled in from memory, so a paid R2 bill is not modelled. The page-path headroom comes from the documented 20,000 files a Worker version on free and 100,000 on paid, against the 5,385 files wrangler read for the canonical config.

The publishing figures for the assets arm are from the 2026-11-02 drive, on one paid throwaway with one changed asset per republish. Four of seven attempted republishes completed; the three that failed died in wrangler with fetch failed from the laptop, which says nothing about the platform.

The serial cap of 5,268 publishes a day is 86,400 s over the slowest measured deploy. It is a throughput bound from four runs and not a quota, and a changed script, as opposed to a changed asset, was not tried. The reset result is the one a republish-per-save design depends on, and n=4 with one object turning over by itself in the no-deploy windows does not separate the two.

A save does not make a visitor wait for a render at this size. In a drive of 10 saves and 1,000 anonymous visits on a six-node site where each save invalidated three paths, 0.3% of visits met a render, 0% a 503, 27.2% a stale page and 72.5% a current one. The stale answers lasted up to 29.2 s, p50 6.6 s. Both numbers describe one small site and not a save that invalidates every page.

What A Binding Or An R2 Read Adds

The shipped path makes one Worker invocation a view, and one more Durable Object invocation when the view is served from the object. A Worker-to-Worker call is a separate invocation with its own CPU and no request fee (published, Workers pricing page). An R2 binding read is a subrequest of the invocation that makes it, so it adds no invocation.

• Public bucket on a custom domain runs no Worker: 0 invocations and 0 Worker CPU a view. This is the lower bound and the one the 13.2 million views a month assumes.
• One extra Worker invocation a view, bracketed at 1.10 ms of front-worker CPU (measured, a cached object serve on deployed paid workers, 2026-09-14, n not recorded). That is 2.81 mJ a view at the model's watts and PUE (derived (modelled)). It stands in for an R2 read through a Worker or a service hop, whose callee CPU is unmeasured, so it is a bound and not a reading.

Through a Worker, the mirror no longer offloads the Worker meter, and the free ceiling falls to the Worker bound in the table above. The whole-year effect on the energy headline is in Service Bindings And R2.

On a paid account, R2 costs $0.36 per million Class B reads, $4.50 per million Class A writes and $0.015 per GB-month of storage (published, R2 pricing page, last updated Oct 30, 2026). The page gives the free tier as 10 million Class B and 1 million Class A a month, and does not say what a paid account keeps. A 17.2 KB page costs $0.000258 a month per 1,000 mirrored pages (derived (modelled), the page size is one page, n=1). Whether a service-bound invocation counts toward the free plan's 100,000 requests a day is unmeasured.