AIHigh

GPT 6.2 Astra reviews split between testers and engineers

Users say opinions on OpenAI's GPT 6.2 Astra are unusually split: some testers dislike it, while systems programmers praise it. Others call the new models capable but incremental.

A laptop screen showing lines of program code
A laptop screen showing lines of program code. Photo: James Harrison, via Unsplash

Confirmed

  • OpenAI made GPT-6.2 Astra and Sol available on its API
  • UK AISI testing is published in the system card
  • UK AI Security Institute reports a jailbreak-robustness regression
  • A 30-person early-access team said on 17 November that nobody loved Astra 6.2

Not yet known

  • Whether Astra is a broad improvement or only an incremental upgrade; users disagree
  • Whether the model is the start of a plateau, as one user fears
  • How Astra compares with rival models; it is being benchmarked, with no results given

What we know

New: one user says he has never seen opinions so split on a model, with testers who outsource their thinking finding it impenetrable and serious engineers enjoying it. The post quotes a developer who calls Astra the best model for C systems programming. Another user says the new models are capable and efficient but an incremental upgrade, and hopes it is not the start of a plateau. Background: on 17 November a 30-person early-access team said nobody loved Astra 6.2.

How the story developed

18 Nov · 10:44 UTC

Users say opinions on OpenAI's GPT 6.2 Astra are unusually split: some testers dislike it, while systems programmers praise it. Others call the new models capable but incremental.

18 Nov · 00:44 UTC

A 30-person team with early access says nobody loved OpenAI's Astra 6.2.

17 Nov · 21:44 UTC

OpenAI's GPT-6.2 Astra is live on third-party platforms and is being benchmarked against rival models.

17 Nov · 20:44 UTC

OpenAI's GPT-6.2 Astra and Sol are available, with UK AISI testing published in the system card.

17 Nov · 19:44 UTC

As OpenAI's GPT-6.2 Astra and Sol roll out, the UK AI Security Institute reports a jailbreak-robustness regression, and Epoch AI measures higher hidden reasoning.

17 Nov · 18:44 UTC

OpenAI made GPT-6.2 Astra and Sol available on its API, and Microsoft announced six in-house MAI models at Ignite the same day.

What people are posting

Posts on X this story draws on. They are individual users' views and have not been verified by Global Eye 24.

@banteg

“never seen opinions so split on a model”

View post →
@mzeller

“Yet another incremental upgrade, nothing night and day”

View post →

By the Global Eye 24 desk, with AI assistance. Updated as the story develops. Spot an error? Write to corrections@globaleye24.com.