Platforms

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.

Richard TangRichard Tangthree-minute read
Share
Screenshot of Greptile's product interface for adding custom review rules, showing a rule description and the repositories and file paths a rule applies to.
1 / 6Slide 1 of 6
Greptile's product, where a team tells the review agent what to check in its code. The company now charges $1 per review credit, and Parcha's bill for the same work went from $180 to more than $7,000 a month.

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 5, open-sourced AgentRun, an Apache-2.0 platform, and built a reviewer on a cheap model that costs about 18 cents a review 13. The launch drew 138 points on Hacker News this week 4.

A meter that grows with your own agents

The meter itself is not new; we covered it last week 6. Greptile's Pro plan charges $1 per credit, 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−4.27% — Nvidia, down 4.27 percent today and Coinbase−6.77% — Coinbase, down 6.77 percent today 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.

Parcha's Greptile bill grew 36x in ten months

  • Monthly Greptile bill
  • Estimate
$0$2,000$4,000$6,000$8,000DecemberOctoberNovember to date$6,451
Data
Monthly Greptile bill
December$180
October$6,451
November to date (estimate)$7,018
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.

Parcha's tuned reviewer caught more of Parcha's real bugs than Greptile did

  • All real bugs caught
  • Serious bugs caught
0%20%40%60%80%Parcha AgentRunGreptile26%42%62.5%
Data
All real bugs caughtSerious bugs caught
Parcha AgentRun42%62.5%
Greptile26%12.5%
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−0.70% — GitHub, down 0.70 percent today Copilot, though cubic does not name the models behind its score 7. Cohere bought cubic on November 13, with terms undisclosed 78. It came days after a Series E that values the company at about $25 billion 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.

Get Platforms by email, free

The cloud, data and developer software AI is built on. In our daily email, every weekday.

Loading NVDA

Reader comments

0 comments

    More in Platforms

    All in this topic →

    Platforms

    Pump.co's claimed spend under management jumped from $600 million to $10 billion in six weeks

    The Y Combinator-backed startup charges customers $0, says cloud providers pay it, and is hiring account executives and partnership leads anyway. For FinOps rivals that bill monthly, and for anyone underwriting the next round, the homepage's new $10B+ figure is the first claim to test.

    Platforms

    LangChain raises $300 million at $3 billion, and the cap table now includes its own customers

    LangChain's agent-governance layer just repriced 2.4 times in 13 months on a usage number its own lead investor calls apples to oranges, and two of its new backers sell competing agent controls. Whose money moves: anyone pricing AI application dependencies against what the model labs will bundle next.

    Platforms

    DXC booked 80 cents of new orders for every dollar of revenue last quarter

    DXC spent last week selling OASIS, the agent platform it says cuts incident-resolution time 95% inside banks and hospitals. Its September order book, 80 cents of new work per dollar billed, says the money has not followed the demos, and the core it must replace is still running off.

    Platforms

    Palantir's Karp says a wealth tax 'is going to need surveillance' as the stock sits at 219 times earnings

    Recorded on November 10 for Walter Isaacson's podcast and carried to more than 14,000 Reddit points, the interview landed in a week that took 17.55 percent off Palantir. At 219 times earnings, holders own the founder's story as much as his contract book.

    Platforms

    Amtrak's booking site failed during the FAA's four-hour national ground stop

    Amtrak blamed a "systemwide update", the same failure mode that grounded the country's flights: a change pushed into a system that is never allowed to stop. The money Congress is asked to add will buy change discipline on live systems, and the vendors selling staged rollouts and tested failover are the layer positioned to be paid for uptime.

    Platforms

    Databricks, valued at $190 billion, routes OpenAI's ChatGPT and Codex through Unity Gateway

    The November 13 playbook makes Databricks' Unity Gateway the place where OpenAI's models are permission-checked, logged and budgeted, weeks after a $190 billion funding round. The listed read-through is Snowflake, which is growing again, and an Anthropic IPO that prices on November 18.

    Subscribe to TheInference for free