---
title: "GPT-6.2 Astra in my state-machine review finds more state bugs and comments far outside state"
pubDate: 2026-11-18
description: "A first run of QuickReview's state-machine review on OpenAI's GPT-6.2 Astra, on the same 34 pull requests and with the same judge as the published runs. It found 52 of the 55 state bugs in the answer key, against 44 on GPT-6.1 Sol, and it left 448 comments against 159, many of them not about state, for $21.06 a pull request against $2.78."
author:
  name: "Mateusz Jakubczak"
  image: "/blog-assets/authors/mateusz_jakubczak.jpeg"
image:
  url: "/blog-assets/posts/gpt-6-2-astra-in-my-state-machine-review_bg_light.png"
  alt: "Two code review panels side by side on a cream background, each showing the same eleven lines of an upload function and a cancel function. The left panel, GPT-6.1 Sol, has a single red comment, \"stuck in uploading\", on the early return. The right panel, GPT-6.2 Astra, has a comment on nine of the lines: three red state-bug comments, \"stuck in uploading\", \"done overwrites idle\" and \"cancel before put() starts\", and six grey comments such as \"add a JSDoc comment\", \"rename url to uploadUrl\" and \"extract a helper\". A legend below reads \"state bug\" and \"other comment\"."
tags: ["engineering"]
minutesToRead: 6
faq:
  - question: "Does GPT-6.2 Astra make QuickReview's state-machine review find more state bugs?"
    answer: "In one run on 34 AACR-Bench pull requests, judged by GPT-6.1 Sol, the state-machine review on GPT-6.2 Astra found 52 of the 55 state bugs in the answer key. The same review on GPT-6.1 Sol found 44, and the default review on GPT-6.1 Sol found 45. Astra found every state bug that the Sol state review found, and 8 more."
  - question: "How much does the state-machine review cost on GPT-6.2 Astra?"
    answer: "$21.06 a pull request at list prices, against $2.78 for the same review on GPT-6.1 Sol counted the same way, about 7.6 times as much. GPT-6.2 Astra's list price is four times GPT-6.1 Sol's. A review took 15.1 minutes on average on Astra and 12.0 on Sol."
  - question: "Is the state-machine review on GPT-6.2 Astra better than the default review?"
    answer: "This run does not show that. On the 55 state bugs in the answer key it found more than the default review on GPT-6.1 Sol (52 against 45). Over all 186 answer-key defects, state or not, the default review on Sol found 161 and the state review on Astra 152, and the default review has not been run on Astra."
---

## TL;DR

**OpenAI released GPT-6.2 Astra on 17 November. I ran QuickReview's state-machine review on it, on the same 34 pull requests and with the same judge as the runs I published on 16 November. It found more state bugs: 52 of the 55 in the answer key, against 44 for the same review on GPT-6.1 Sol, and every one that Sol found. It also commented far outside state. The state review is meant to post state findings only, and on Astra it left 448 comments against 159 on Sol, more than the default review's 397. Of the answer-key defects those comments found, 100 are not state bugs, against 48 on Sol. It cost $21.06 a pull request against $2.78.**

On 16 November I wrote about [a state-machine review for the bugs line-by-line review approves](https://quickchat.ai/post/implicit-state-machines-hide-bugs-that-line-by-line-review-approves). It reads a change through eight narrow reviewers, which QuickReview calls lenses, each looking for one kind of state bug, and a validator replays every claimed sequence of events against the code before anything is posted. Every review in that post ran on GPT-6.1 Sol.

GPT-6.2 Astra is available in ChatGPT, Codex and the API. Its list price is $8 per million input tokens, $40 per million output tokens and $0.40 per million cached input tokens, four times GPT-6.1 Sol's $2 and $10. I wanted to measure what a newer and more expensive model does to the state review while the model is still new.

## The setup

| | This run | The runs it is compared with |
|---|---|---|
| Review | QuickReview 1.5.1, the `state-machines` profile | The same state-machine review; the default review with its automatic state lens off |
| Reviewer model | GPT-6.2 Astra on every tier (medium reasoning, high for the deep tier) | GPT-6.1 Sol |
| Pull requests | 34 from AACR-Bench | The same 34 |
| Judge | GPT-6.1 Sol, high reasoning | The same |
| Runs | One | One each, published with the last post |

The `state-machines` profile runs the eight state lenses and the validator, and no other lens ran on any of the 34 pull requests. All 34 reviews completed. The run went through a ChatGPT login, so nothing was billed, and every cost below is a list-price equivalent. All the data behind this post is public in QuickReview's [benchmark dataset on Hugging Face](https://huggingface.co/datasets/quickchatai/quickreview-bench/tree/main/state_machines): every raw comment of the three reviews, every judge verdict, and each review's cost and time.

QuickReview runs its review sessions in pi, a coding-agent runtime. pi added GPT-6.2 Astra to its model catalog on 17 November, after its latest release, so this run gave pi the same entry through its `models.json`.

Astra's lens passes are slower. The median state-lens pass took 6.3 minutes, against 3.5 on Sol, and many run past QuickReview's default lens deadline of 8 minutes. The run set `lens_pass_deadline_min = 20` under `[engine]`, and that is the setting to raise if you run the lenses on Astra.

## 52 of the 55 state bugs in the answer key

The answer key is the list of labelled defects a review is scored against. The last post relabelled the 186 defects on these 34 pull requests with a strict rule, which left 55 real state bugs.

![Bars of the state bugs in the answer key found on 34 AACR-Bench pull requests, of 55, judged by GPT-6.1 Sol: the state-machine review on GPT-6.2 Astra 52 (95%); the state-machine review on GPT-6.1 Sol 44 (80%); the default review on GPT-6.1 Sol with its state lens off 45 (82%)](../../assets/blog/posts/gpt-6-2-astra-in-my-state-machine-review/fig1-state-bugs-on-the-answer-key.svg)

**On GPT-6.2 Astra the state-machine review found 52 of the 55.** On GPT-6.1 Sol the same review found 44 and the default review 45. Astra found every state bug that the Sol state review found, and 8 more. It missed three, one each in ClickHouse, LVGL and mpv, and the state review on Sol missed the same three.

## Far more comments, many of them not about state

The state-machine review is meant to post state findings and nothing else, so that a team that turns it on gets a short list. On Sol it was the short review, with 159 inline comments, 4.7 a pull request. On Astra it left 448, 13.2 a pull request, more than the default review's 397 (11.7) on Sol.

Many of those comments are about other things. Over all 186 answer-key defects, state or not, the state review on Astra found 152. Of those, 52 are state bugs and 100 are not. The same review on Sol found 92, of which 48 are not state bugs, and the default review on Sol found 161. On Astra the state review behaved much like a general review that also finds more state bugs.

## Cost and time

![Bars of the list-price cost per pull request of the state-machine review on 34 AACR-Bench pull requests: $21.06 on GPT-6.2 Astra at 15.1 minutes on average, $2.78 on GPT-6.1 Sol at 12.0 minutes on average](../../assets/blog/posts/gpt-6-2-astra-in-my-state-machine-review/fig2-cost-per-pull-request.svg)

The Astra run cost $716.05 for the 34 pull requests, or $21.06 each, with every model call of the review priced at list and the judge left out. That is the method behind the $2.78 the last post printed for the same review on Sol, so Astra cost about 7.6 times as much. A review took 15.1 minutes on average on Astra and 12.0 on Sol.

## What it adds up to

On these 34 pull requests, the state-machine review on GPT-6.2 Astra found more of the answer key's state bugs than either review on GPT-6.1 Sol. It did not stay on state. It posted almost three times as many comments as on Sol, more than the default review, and cost about 7.6 times as much. Over all 186 answer-key defects the default review on Sol still found more, 161 against 152, and the default review has not been run on Astra yet.

The raw comments of the Astra run and the two Sol runs, the judge's verdict on each comment and each answer-key defect, and every review's cost are in the `state_machines` folder of QuickReview's [benchmark dataset on Hugging Face](https://huggingface.co/datasets/quickchatai/quickreview-bench/tree/main/state_machines). QuickReview is open source at [github.com/quickchatai/quickreview](https://github.com/quickchatai/quickreview), and the state-machine review ships in 1.5.0 and later.

## Frequently asked questions

### Does GPT-6.2 Astra make QuickReview's state-machine review find more state bugs?

In one run on 34 AACR-Bench pull requests, judged by GPT-6.1 Sol, the state-machine review on GPT-6.2 Astra found 52 of the 55 state bugs in the answer key. The same review on GPT-6.1 Sol found 44, and the default review on GPT-6.1 Sol found 45. Astra found every state bug that the Sol state review found, and 8 more.

### Does the state-machine review on GPT-6.2 Astra stay on state bugs?

No. It left 448 inline comments on the 34 pull requests, against 159 for the same review on GPT-6.1 Sol and 397 for the default review on Sol. Of the 152 answer-key defects its comments found, 100 are not state bugs, against 48 of 92 for the same review on Sol.

### How much does the state-machine review cost on GPT-6.2 Astra?

$21.06 a pull request at list prices, against $2.78 for the same review on GPT-6.1 Sol counted the same way, about 7.6 times as much. GPT-6.2 Astra's list price is four times GPT-6.1 Sol's. A review took 15.1 minutes on average on Astra and 12.0 on Sol.

### Is the state-machine review on GPT-6.2 Astra better than the default review?

This run does not show that. On the 55 state bugs in the answer key it found more than the default review on GPT-6.1 Sol (52 against 45). Over all 186 answer-key defects, state or not, the default review on Sol found 161 and the state review on Astra 152, and the default review has not been run on Astra.

## Related guides

- [Implicit state machines hide bugs that line-by-line review approves](https://quickchat.ai/post/implicit-state-machines-hide-bugs-that-line-by-line-review-approves)
- [Same code reviewer, three models, three different reviewers](https://quickchat.ai/post/same-code-reviewer-three-models-three-different-reviewers)

---

Not answered on this page? Ask Quickchat AI directly: https://app.quickchat.ai/v1/api_core/public-question/quickchat/ask/?q=your+question&placement=www-markdown-footer-ask-v1&format=html
