---
title: "Parcha's Greptile bill passed $7,000 a month, so it built an 18-cent reviewer and open-sourced it"
description: "The bill is the first customer-published measure of what usage-priced AI code review costs when agents write the pull requests, and it landed the week Cohere bought the reviewer ranked above Greptile on an independent leaderboard. The money at stake is Greptile's $750 million mark and every seat-priced review vendor's renewal."
publisher: "The Inference"
section: "Platforms"
published: 2026-11-16T16:11:29.240Z
modified: 2026-11-16T16:11:29.240Z
canonical: https://theinference.org/article/parcha-s-greptile-bill-passed-7-000-a-month-so-it-built-an-18-cent-reviewer-and-open-sourced-it
language: en
keywords: "AI, Developer Tools, Code Review, Coding Agents, Open Source"
---

# Parcha's Greptile bill passed $7,000 a month, so it built an 18-cent reviewer and open-sourced it

> The bill is the first customer-published measure of what usage-priced AI code review costs when agents write the pull requests, and it landed the week Cohere bought the reviewer ranked above Greptile on an independent leaderboard. The money at stake is Greptile's $750 million mark and every seat-priced review vendor's renewal.

[Greptile](https://theinference.org/markets/companies/greptile)'s customers pay a dollar a review, and this week one of them published the bill. In December Parcha paid $180 a month for the AI agent that reviews its pull requests [1]. Its own coding agents then began opening more of those pull requests. October's bill ran $6,451, and November has already passed $7,000 [1].

Instead of paying, Parcha, a YC-backed agent platform whose [site now announces a rename to Grep AI](https://www.parcha.ai/) [5], open-sourced AgentRun, an Apache-2.0 platform, and built a reviewer on a cheap model that costs about 18 cents a review [1][3]. The launch drew [138 points on Hacker News](https://hn.algolia.com/?q=agentrun) this week [4].

## A meter that grows with your own agents

The meter itself is not new; [we covered it last week](https://theinference.org/article/greptile-drops-the-seat-fee-and-charges-1-a-review) [6]. Greptile's [Pro plan charges $1 per credit](https://www.greptile.com/pricing), one for a standard review and ten for a review of a large change, with no seat fees [5]. The company says it reviews more than 1% of all pull requests opened on GitHub daily for customers including [Nvidia](https://theinference.org/markets/companies/nvidia) and [Coinbase](https://theinference.org/markets/companies/coinbase) [6].

What is new is the receipt. Usage pricing scales with the pull-request volume a coding-agent rollout creates, so the bill compounds with the customer's own automation, and Greptile's answer is that price tracks the tokens a correct review needs [6]. Parcha's answer is to take the tokens off the meter.

*Chart: **Parcha's Greptile bill grew 36x in ten months** Column chart of Parcha's monthly Greptile review bill: $180 in December, $6,451 in October and $7,018 month-to-date in November.*

*Parcha's reported monthly bill for Greptile code review, US dollars; November is month-to-date as reported on November 13, 2026. Source: Parcha's AgentRun README as reported by Stack Futures. [1]*

## Eight cheap reviewers under a verifying parent

Parcha's replacement is a parent agent that dispatches eight reviewers over a read-only checkout of each change: five quick readers of the whole diff and three deep readers of the risky parts. The parent verifies every finding against the code and posts a comment-only review that never approves [1]. Access, not model size, was the bottleneck: with only the diff, the cheap model was wrong about 70% of the time [1]. Given read-only shell on the repository, that error rate fell to 7% [1].

On the 30 pull requests Parcha held out of tuning, its reviewer caught 42% of real bugs; Greptile caught 26% [1]. Of the 8 serious bugs, Parcha's caught 5 and Greptile 1 [1]. The caveats are real: Parcha's own repository, its own pooled answer key of 78 bugs, and no false-positive comparison, so 5 versus 1 is suggestive, not settled [1].

*Chart: **Parcha's tuned reviewer caught more of Parcha's real bugs than Greptile did** Column chart comparing Parcha's AgentRun reviewer and Greptile on 30 held-out pull requests from Parcha's repository: Parcha caught 42% of real bugs to Greptile's 26%, and 62.5% of serious bugs (5 of 8) to Greptile's 12.5% (1 of 8).*

*Share of real bugs caught on the 30 pull requests Parcha held out of tuning, on Parcha's own repository; serious bugs are 5 of 8 for Parcha and 1 of 8 for Greptile. No false-positive rate was published. Source: Parcha's AgentRun README as reported by Stack Futures. [1]*

## The independent leader just joined a $25 billion buyer

The ground moved the same week. Martian's independent leaderboard ranked cubic first at 64.9% F1 as of October 6, ahead of Greptile, CodeRabbit and [GitHub](https://theinference.org/markets/companies/microsoft) Copilot, though cubic does not name the models behind its score [7]. Cohere bought cubic on November 13, with terms undisclosed [7][8]. It came days after a [Series E that values the company at about $25 billion](https://www.techtimes.com/articles/330774/20261116/cohere-acquires-ai-code-reviewer-cubic-bug-finder-aisle-air-gapped-results-unpublished.htm) [7]. Greptile had raised $50 million from Menlo Ventures in October [6]. The round marked it at $750 million [6].

## The value moves to whoever earns merge-time trust

Greptile's own logic makes its revenue track inference cost and pull-request volume [6]. A customer has now shown that the product can be rebuilt on a cheap model tuned to one repository, for a fraction of the price [1]. What a managed reviewer still sells is cross-repo context and the confidence to let it gate a merge, which is where the margin now lives [5].

The recipe is public, so the question for Greptile's investors is whether $1 a review holds once every agent-heavy team can tune its own reviewer, and the seat-priced holdouts should ask the same of their seats. Parcha built the missing link between a cheap model and a reviewer a team will trust, and gave it away. On this chain, the builder who turns the capability into something people rely on is the one creating the value.

## Takeaway

A customer published the receipt for usage-priced AI code review, then shipped a working replacement at about 18 cents a review, which turns Greptile's $1-a-review plan into a renewal fight.

## Sources

1. [Stack Futures, Parcha Open-Sources AgentRun After Its Greptile Bill Hit $6,451 a Month, Says a Tuned Cheap-Model Reviewer Costs 18 Cents, November 2026](https://stackfutures.com/blog/parcha-agentrun-code-review-18-cents-greptile-bill-180-to-6451-nov-2026/)
3. [Parcha, Parcha is now Grep AI, accessed 16 November 2026](https://www.parcha.ai/)
4. [Hacker News (Algolia search), Show HN: AgentRun, posted 12 November 2026](https://hn.algolia.com/?q=agentrun)
5. [Greptile, Pricing, accessed 16 November 2026](https://www.greptile.com/pricing)
6. [The Inference, Greptile drops the seat fee and charges $1 a review, 13 November 2026](https://theinference.org/article/greptile-drops-the-seat-fee-and-charges-1-a-review)
7. [Tech Times, Cohere Acquires AI Code Reviewer cubic and Bug Finder AISLE; Air-Gapped Results Unpublished, 16 November 2026](https://www.techtimes.com/articles/330774/20261116/cohere-acquires-ai-code-reviewer-cubic-bug-finder-aisle-air-gapped-results-unpublished.htm)
8. [BetaKit, Cohere buys AI code review startup Cubic as acquisition spree continues, 13 November 2026](https://betakit.com/cohere-buys-ai-code-review-startup-cubic-as-acquisition-spree-continues/)


---

Parcha's Greptile bill passed $7,000 a month, so it built an 18-cent reviewer and open-sourced it — The Inference. Canonical: https://theinference.org/article/parcha-s-greptile-bill-passed-7-000-a-month-so-it-built-an-18-cent-reviewer-and-open-sourced-it
