WorkPublished Top storyAll news from that day
Ongoing case: AI code review in practice 3 stories
Cloudflare says its AI code reviews now cost 68% less
Cloudflare's engineering blog says its in-house code reviewer now costs $0.77 per review, after the team fixed prompt caching and tested models and effort settings.
Check our sources · 9 facts from 1 source
Image: Cloudflare
Key points
- Cloudflare says the average review now costs $0.77, down 68% from $2.43 in its two most expensive weeks, and median review time fell from 4m 09s to 1m 32s.
- A session's first request found a cached prefix only 14% to 26% of the time; a prompt split on 16 September gave the shared part its own cache breakpoint.
- About 40% of the savings came from the reviewer's code and the rest from models and effort; Sonnet 5.5 was 62% cheaper than Sonnet 5 for code quality.
What happened
Cloudflare published an account of the cost of its own AI code reviewer on 19 October 2026. It says that with code quality at high effort the average review now costs $0.77, down 68% from the $2.43 it averaged over its two most expensive weeks. The median review time fell from 4m 09s to 1m 32s.
About 40% of the savings came from changes to the reviewer's code and the rest from experiments with models and effort. Cloudflare found that the first request of a session found a cached prefix only 14% to 26% of the time, although every session starts with the same 11k to 21k tokens.
The cause was how OpenCode v1 built its system prompt, with the only cache breakpoint placed after a block that included repository-specific details. On 16 September it split the prompt in two at a marker so the shared part got its own breakpoint.
On 3 October it split its Sonnet reviewers 50/50 between Sonnet 5 and Sonnet 5.5, and Cloudflare says Sonnet 5.5 was 62% cheaper per run for code quality. Cloudflare also reports that the code quality reviewer at high effort posts 40% fewer critical and warning findings than in late August.
Underlined sentences link to the facts we checked. Dotted words open a short explainer.
Fact 2 of 9 · from Cloudflare Blog, 19 Oct 2026
the average review now costs just $0.77, down 68% from the $2.43 it averaged over our two most expensive weeks before we started.
Fact 4 of 9 · from Cloudflare Blog, 19 Oct 2026
About 40% of the savings came from changes and optimizations to the reviewer's code, and the rest from experimenting with which models we use and how much effort they put in.
Fact 5 of 9 · from Cloudflare Blog, 19 Oct 2026
The data showed that the first request of a session found a cached prefix only 14% to 26% of the time. Every session starts with the same tool definitions and the same agent prompt, between 11k and 21k tokens that are identical in every repository we review
Fact 8 of 9 · from Cloudflare Blog, 19 Oct 2026
On October 3, we split every Sonnet reviewer 50/50 between Sonnet 5 and Sonnet 5.5. After two and a half days, Sonnet 5.5 was 62% cheaper per run for code quality
What it means for you
Our viewThe figures are Cloudflare's, measured on its own workload. The sharpest finding is about caching: the start every session shares was rarely found in the cache, because of where the only cache breakpoint sat.
Cloudflare also reports 40% fewer critical and warning findings from its code quality reviewer; fewer severe findings alone does not show whether real problems are missed or noise is cut, so we do not read it as a quality gain.
If you pay for AI code review, check how often the first request of a session hits your cache, and track missed defects alongside cost.
It adds no new facts.
Your reaction
We count reactions per story and day, never who reacted. The counts help us choose what goes in the monthly issue. If you are signed in, your own page shows yours too.
We count reactions per story and day, never who reacted. They help choose what goes in the monthly issue.
Check our sources
Every sentence above is checked against this source.
1 Cutting our inference bill by 68% by being obsessive about optimization
Open the source archived copy-
Cloudflare's engineering blog post of 19 October 2026 says that with code quality at high effort the average review of its AI code reviewer now costs $0.77.
citeOctober 19, 2026 Cutting our inference bill by 68% by being obsessive about optimization ... With code quality at high effort, the average review now costs just $0.77
-
The post says that $0.77 is down 68% from the $2.43 the average review cost over its two most expensive weeks before the work started.
citethe average review now costs just $0.77, down 68% from the $2.43 it averaged over our two most expensive weeks before we started.
-
The post says the median review now takes 1m 32s instead of the 4m 09s it took before.
citethe median review now takes just 1m 32s instead of the 4m 09s it took before.
-
The post says about 40% of the savings came from changes and optimizations to the reviewer's code, and the rest from experimenting with which models are used and how much effort they put in.
citeAbout 40% of the savings came from changes and optimizations to the reviewer's code, and the rest from experimenting with which models we use and how much effort they put in.
-
The post says the first request of a session found a cached prefix only 14% to 26% of the time, although every session starts with the same tool definitions and agent prompt of between 11k and 21k tokens.
citeThe data showed that the first request of a session found a cached prefix only 14% to 26% of the time. Every session starts with the same tool definitions and the same agent prompt, between 11k and 21k tokens that are identical in every repository we review
-
The post says OpenCode v1 built one system-prompt block containing the agent prompt followed by an environment section, with the only cache breakpoint at the end of that block.
citeThe problem ended up being in how OpenCode v1 constructed the system prompt: It built one block containing our agent prompt, followed by an environment section with the working directory and the date, and placed the only cache breakpoint at the end of that block.
-
The post says that on September 16 Cloudflare added a marker to the end of every agent prompt and a plugin hook that split the block in two at that marker, so the shared first half got its own breakpoint.
citeOn September 16, we added a marker to the end of every agent prompt and a plugin hook that split the block in two at that marker. The first half now ended on bytes that every job shares, and got its own breakpoint.
-
The post says that on October 3 Cloudflare split every Sonnet reviewer 50/50 between Sonnet 5 and Sonnet 5.5, and that after two and a half days Sonnet 5.5 was 62% cheaper per run for code quality.
citeOn October 3, we split every Sonnet reviewer 50/50 between Sonnet 5 and Sonnet 5.5. After two and a half days, Sonnet 5.5 was 62% cheaper per run for code quality
-
The post says code quality on Sonnet 5.5 at high effort posts 40% fewer critical and warning findings than it did in late August.
citeWith code quality on Sonnet 5.5 at high effort, it posts 40% fewer critical and warning findings than it did in late August.
Every sentence above is checked against this source. Each fact has its own address you can cite.
Source 1 · Cloudflare Blog · 19 Oct 2026
Cutting our inference bill by 68% by being obsessive about optimization
Cloudflare's engineering blog post of 19 October 2026 says that with code quality at high effort the average review of its AI code reviewer now costs $0.77.
October 19, 2026 Cutting our inference bill by 68% by being obsessive about optimization ... With code quality at high effort, the average review now costs just $0.77
The post says that $0.77 is down 68% from the $2.43 the average review cost over its two most expensive weeks before the work started.
the average review now costs just $0.77, down 68% from the $2.43 it averaged over our two most expensive weeks before we started.
The post says the median review now takes 1m 32s instead of the 4m 09s it took before.
the median review now takes just 1m 32s instead of the 4m 09s it took before.
Show all 9 factsShow fewer facts
The post says about 40% of the savings came from changes and optimizations to the reviewer's code, and the rest from experimenting with which models are used and how much effort they put in.
About 40% of the savings came from changes and optimizations to the reviewer's code, and the rest from experimenting with which models we use and how much effort they put in.
The post says the first request of a session found a cached prefix only 14% to 26% of the time, although every session starts with the same tool definitions and agent prompt of between 11k and 21k tokens.
The data showed that the first request of a session found a cached prefix only 14% to 26% of the time. Every session starts with the same tool definitions and the same agent prompt, between 11k and 21k tokens that are identical in every repository we review
The post says OpenCode v1 built one system-prompt block containing the agent prompt followed by an environment section, with the only cache breakpoint at the end of that block.
The problem ended up being in how OpenCode v1 constructed the system prompt: It built one block containing our agent prompt, followed by an environment section with the working directory and the date, and placed the only cache breakpoint at the end of that block.
The post says that on September 16 Cloudflare added a marker to the end of every agent prompt and a plugin hook that split the block in two at that marker, so the shared first half got its own breakpoint.
On September 16, we added a marker to the end of every agent prompt and a plugin hook that split the block in two at that marker. The first half now ended on bytes that every job shares, and got its own breakpoint.
The post says that on October 3 Cloudflare split every Sonnet reviewer 50/50 between Sonnet 5 and Sonnet 5.5, and that after two and a half days Sonnet 5.5 was 62% cheaper per run for code quality.
On October 3, we split every Sonnet reviewer 50/50 between Sonnet 5 and Sonnet 5.5. After two and a half days, Sonnet 5.5 was 62% cheaper per run for code quality
The post says code quality on Sonnet 5.5 at high effort posts 40% fewer critical and warning findings than it did in late August.
With code quality on Sonnet 5.5 at high effort, it posts 40% fewer critical and warning findings than it did in late August.
Topics
The morning email
On the mornings we publish: the three top stories and up to four short ones. Free.
We email you a link to confirm. An issue may include one sponsor, always labelled Sponsored · Advertisement. Our emails count opens and clicks, not who made them. Unsubscribe in one click. What we keep