Stop spending your Opus limit on code review

R
Ryan Monsurate
5 min read
Share:

If you use Opus every day, you know the weekly limit. You hit it, and you wait.

At Farpoint, Opus 5.5 agents write most of our code. Every change they make gets a full code review before it ships. We don't run those reviews on Opus. We run them on Fabric, and we measured what that costs.

How much a review uses

More than you'd guess. A thorough review reads the change, searches it for bugs, checks each suspected bug, then runs the app to test it. It rereads the same code many times along the way.

Across our last ten reviews, a typical one used 52 million tokens. Tokens are what AI models bill by, roughly a word or a piece of a word each. The smallest review used 29 million and the largest used 118 million.

That much reading would take a large bite out of a weekly Opus allowance.

What that would cost on Opus

We priced a typical review at Anthropic's published Opus 5.5 rates.

Type of tokenTokensPrice per millionCost
Reread text (cached input)48.4 million$0.20$9.67
New text (uncached input)3.4 million$4 to $5$13.55 to $16.93
Text the model writes (output)0.55 million$20$11.05
Totalabout 52 million$34 to $38

So about $35 a review. All ten together would have cost $401 to $440.

We took the prices from Anthropic's pricing page on November 9, 2026. New input costs $4 or $5 per million depending on how it's cached, so we show both. We assumed Opus would use the same number of tokens. It wouldn't use exactly that many, so read this as a rough comparison.

What it costs on Fabric

For one reviewTypicalLowest to highest
Cost$0.90$0.55 to $1.81
Tokens52 million29 to 118 million
Timeabout 45 minutes19 to 57 minutes

The ten reviews cost us $10.76 in total, about a dollar each. These are our model costs. Customers pay a margin on top.

Most of the saving comes from caching. When a model rereads text it has already seen, that text is billed at a much lower rate. In our reviews, 91% to 95% of everything the model read was reread text.

Does a cheaper reviewer catch anything?

Yes. Every one of the ten reviews found at least one bug serious enough to fix before release. In total they confirmed 19, about 57 cents per bug.

Opus 5.5 agents wrote all of the code that was reviewed. Here are four of the bugs, all of which would have reached users:

  1. When a chat's chosen model couldn't be reached, Fabric offered a paid cloud model instead. If you said yes and then left the chat, the paid request still ran.
  2. If one small part of a chat failed to load, the whole app showed the crash screen.
  3. If you pasted an image and left the chat before it finished uploading, the image ended up in the chat you had left.
  4. A feature meant to remove old entries for offline Macs did nothing in the exact situation it was built for.

None of these are exotic. They sit just off the main path, where the person writing the code wasn't looking.

How to run it

The review is built into Fabric. Type /code-review-stage-1 followed by a pull request number, and it works through five stages on its own:

  1. Work out what changed and what to check.
  2. Six smaller models search the change for possible bugs at the same time. Each one has to quote the code it's pointing at.
  3. A larger model tries to prove each possible bug wrong. The ones that survive get a severity and a suggested fix.
  4. Write a test plan, then run the real app and follow it.
  5. Confirm every test step has evidence, and give a verdict.

One thing we got wrong

The review finds bugs. It doesn't fix them. At first we ran the whole review again after every fix, and every new run found something else. All ten reviews ended with "not mergeable". One pull request went through eight rounds.

Now we run the full review once, fix everything serious it finds, and then re-check only those fixes. When the re-check passes, the change ships.

What these numbers don't show

This was ten reviews of four pull requests, all on one day, November 7, 2026. We counted a bug as real when the review confirmed it and a fix followed. We didn't double-check each one by hand.

So far the pattern has held. Opus does the work only Opus can do. Fabric checks it, for about a dollar.

R

Ryan Monsurate

Founder of Fabric. Building the future of AI-assisted development.

Related posts

v1.71: Connectors

Fabric now connects to Gmail, Slack, Notion, GitHub and more, lets you see and edit what it remembers, and can run fully offline on Apple Silicon Macs.

F
Fabric Team
6 min read

v1.70: Fabric Studio

Fabric Studio brings a no-code workspace to Fabric, alongside one-click local models and a shared queue that keeps your phone, desktop and web in sync.

F
Fabric Team
7 min read

v1.69: Code Map

Code Map replaces the Project Graph with a single drill-down hierarchy of your whole codebase — and you can now drive desktop chats from your phone.

F
Fabric Team
5 min read