This launch highlights how Codex, paired with the new GPT-6 Astra model, is being used as an end-to-end testing tool inside Perplexity. The workflow is aimed at developers who need to validate applications that depend on external services (language model APIs, connectors, and other third-party endpoints) without spending hours writing scaffolding by hand. The OpenAI Developers channel is publishing this as a concrete example of what the updated Codex harness can do beyond code generation.
The featured user is Johnny Ho, cofounder and chief strategy officer at Perplexity, who asks Codex with Astra to build a small testing program around an application. The model generates realistic mock responses that stand in for services the app would normally call, then exercises the workflow from start to finish so Johnny can see how the pieces behave together. According to the customer story on OpenAI's site, Perplexity now trusts the model with full end-to-end systems and checks in on it far less frequently than with previous generations.
For founders and engineering leads evaluating AI coding tools, the interesting signal here is the shift from single-file code completion toward autonomous test harness construction and integration testing against mocked dependencies. If you already build with the OpenAI Platform or Codex, this launch is worth tracking as a reference pattern for how a production AI company is delegating QA work to the model.
Comments (15)
Sign in to join the discussion.
P
Priya Nadkarni24d ago
Mocking third party APIs is the most honest use case anyone has ever put in a launch tweet. Everyone does it, nobody admits it.
K
Kwame Osei24d ago
Cool clip but the cut at 0:14 is jarring, feels like the editor gave up on the second B-roll. Otherwise the screen recording pacing is decent.
L
Linh Tran24d ago
Naming it Astra makes it sound like a perfume launch. I keep expecting a Zendaya voiceover.
M
Matteo Ricci24d ago
The tweet copy buries the lede. Perplexity is running end to end tests with this and you led with a name drop instead of the workflow.
D
Devon Hakobyan24d ago
Genuine question, what are the rate limits for Astra in Codex and is there a webhook when a long running test suite finishes? Docs page just redirects me in a loop.
Y
yusra q.24d ago
So it writes the test, mocks the response, and grades itself. Who exactly is the adversary in this scenario.
O
Olamide Bakare24d ago
Test harnesses generated by the same family of model that wrote the code feels like grading your own homework in a slightly different font.
N
Nkechi Obi24d ago
Before anyone in a regulated industry gets excited, where do the mocked payloads live and are they retained for training. Asking for a nervous compliance team.
S
Sveta Morozova24d ago
Every dev tools launch this quarter has been a testimonial video with a single engineer at a big AI lab. The template is getting stale.
A
Arjun Pillai24d ago
Two people and a weekend could build a wrapper around this that just spits out Playwright suites. Someone is already doing it while I type.
F
Fatima Zahra24d ago
I have been three weeks away from launching my own AI test runner for eight months. Guess I am now four weeks away.
H
Hans Solvig24d ago
Back in 2019 we called this a fixture. Great to see the wheel reinvented with a nebula name.
C
Chui Wen Lee24d ago
The demo is clearly pre recorded. I want to see it actually catch a flaky integration test in real time before I believe the end to end claim.
T
Tomasz K.24d ago
First three seconds are just the logo animation. You lost half the scroll audience before the value prop even loaded.
R
Rhea Balakrishnan24d ago
Is any of the harness scaffolding going to land in an open repo or is this a closed party. A cookbook entry would go a long way.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Open AI Developers
Official updates for developers building with Codex & the OpenAI Platform
Official updates for developers building with Codex & the OpenAI Platform
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
This launch highlights how Codex, paired with the new GPT-6 Astra model, is being used as an end-to-end testing tool inside Perplexity. The workflow is aimed at developers who need to validate applications that depend on external services (language model APIs, connectors, and other third-party endpoints) without spending hours writing scaffolding by hand. The OpenAI Developers channel is publishing this as a concrete example of what the updated Codex harness can do beyond code generation. The featured user is Johnny Ho, cofounder and chief strategy officer at Perplexity, who asks Codex with Astra to build a small testing program around an application. The model generates realistic mock responses that stand in for services the app would normally call, then exercises the workflow from start to finish so Johnny can see how the pieces behave together. According to the customer story on OpenAI's site, Perplexity now trusts the model with full end-to-end systems and checks in on it far less frequently than with previous generations. For founders and engineering leads evaluating AI coding tools, the interesting signal here is the shift from single-file code completion toward autonomous test harness construction and integration testing against mocked dependencies. If you already build with the OpenAI Platform or Codex, this launch is worth tracking as a reference pattern for how a production AI company is delegating QA work to the model.
Comments (15)
Mocking third party APIs is the most honest use case anyone has ever put in a launch tweet. Everyone does it, nobody admits it.
Cool clip but the cut at 0:14 is jarring, feels like the editor gave up on the second B-roll. Otherwise the screen recording pacing is decent.
Naming it Astra makes it sound like a perfume launch. I keep expecting a Zendaya voiceover.
The tweet copy buries the lede. Perplexity is running end to end tests with this and you led with a name drop instead of the workflow.
Genuine question, what are the rate limits for Astra in Codex and is there a webhook when a long running test suite finishes? Docs page just redirects me in a loop.
So it writes the test, mocks the response, and grades itself. Who exactly is the adversary in this scenario.
Test harnesses generated by the same family of model that wrote the code feels like grading your own homework in a slightly different font.
Before anyone in a regulated industry gets excited, where do the mocked payloads live and are they retained for training. Asking for a nervous compliance team.
Every dev tools launch this quarter has been a testimonial video with a single engineer at a big AI lab. The template is getting stale.
Two people and a weekend could build a wrapper around this that just spits out Playwright suites. Someone is already doing it while I type.
I have been three weeks away from launching my own AI test runner for eight months. Guess I am now four weeks away.
Back in 2019 we called this a fixture. Great to see the wheel reinvented with a nebula name.
The demo is clearly pre recorded. I want to see it actually catch a flaky integration test in real time before I believe the end to end claim.
First three seconds are just the logo animation. You lost half the scroll audience before the value prop even loaded.
Is any of the harness scaffolding going to land in an open repo or is this a closed party. A cookbook entry would go a long way.