AIToday
Large Language ModelsAI Business & IndustryBen's BitesPublished: Sep 9, 2026, 01:00 JST1 min read

OpenAI's Astra: First GPT-6 Family Model, Beats ARC-AGI-3

OpenAI's Astra: First GPT-6 Family Model, Beats ARC-AGI-3

3 Key Points

  1. What happened

    OpenAI's Astra model is now available, beating the ARC-AGI-3 benchmark and scoring the highest on Zapier's AutomationBench. The newsletter frames it as a first step in the GPT-6 family.

  2. Why it matters

    Astra is 'token-hungry' and 'spiky', described as 'a show horse, not a workhorse'. A user burned through 4Bn tokens on it, and there's a wait for the anticipated consistency improvements that came with GPT-5.3.

  3. What to watch

    The usefulness of Astra hinges on whether its vast token consumption can be managed for practical tasks. Note the emerging capabilities, like generating autonomous UIs and identifying sounds from spectrograms (images of sound frequencies).

WHO IT HITSAI enthusiasts and power users are directly affected as they adjust to Astra's high token consumption and inconsistent output. Also, users of coding tools can expect Claude Code to become more adaptable if Anthropic builds the plugin feature it is currently testing.

Ask the AI about this article →

Summaries like this, in your inbox every morning.

Context & Analysis

The newsletter reflects the current state of the AI field, where new models are pushing boundaries but are not yet fully polished. Astra has generated significant buzz for its capabilities, such as drawing in Canva and playing piano, but its full potential is somewhat hampered by its current limitations.

The author draws a parallel with the release of GPT-5, which was also initially spiky but became 'consistently great' with the GPT-5.3 update a few months later. This suggests that the current issues with Astra might be ironed out in time, as early adopters find better ways to use it.

While OpenAI works on the future with agents, other companies like Anthropic are also exploring new features. The outcome for Astra hinges on how effectively OpenAI and its users can manage its token consumption and turn its raw potential into consistent, workhorse-grade performance.

FAQ
What exactly is the Astra model?
According to Ben's Bites, Astra is the start of the GPT-6 family. It beats the ARC-AGI-3 benchmark and scores the highest on Zapier's AutomationBench.
What are the main criticisms of Astra?
The newsletter describes Astra as a 'token-hungry monster' that is 'not consistently great' and 'a show horse, not a workhorse'.
What is the next milestone for OpenAI's agents?
According to the article, OpenAI has reached its automated research intern goal. The next target is an automated AI researcher by March 2028.

Get the latest Large Language Models news every morning

For example, today's edition would include:

  • Perplexity CEO Aravind Srinivas: Banks want unpluggable AIYahoo Finance AI · 42m ago
  • Meta drops "tokenmaxxing" from engineer reviews after budget blowoutTHE DECODER · 42m ago
  • Mistral AI raises €3 billion in landmark Series D led by SamsungTechCrunch AI · 42m ago

AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. Q&As are published on this page for other readers too.

Related Articles

Next articleGoogle DeepMind unveils AlphaGenome Atlas to map genome variants