Hacker News front pageSep 4, 2026, 03:45 UTC+8

OpenAI GPT-6 Astra hits 99.9% on ARC-AGI-3 for $19K

OpenAI's GPT-6 Astra on ARC-AGI-3

GPT-6 Astra scored 62.7% for $26K on ARC-AGI-3 Semi-Private with the Standard harness, and 99.9% for $19K with the Provider Adapter harness, which preserves opaque reasoning state and uses compaction. Astra beat the median human in action efficiency on 96% of levels. It built compact symbolic world models from unfamiliar environments and invented its own shorthand to track state and plan. The post does not disclose parameter count, architecture, or release date.

Why it matters: GPT-6 Astra hits 99.9% on ARC-AGI-3, the first flagship model near-perfect on this benchmark, with cost dropping from $26K to $19K. Cross-source coverage is guaranteed. Not 95+ because this is the ARC Prize blog, not an OpenAI release, and the post doesn't disclose Astra's arc...

Read the original ↗Export Markdown

Take

The reason this post is worth opening: two numbers that shouldn't go together. Standard harness gave 62.7% at $26K. The provider adapter—which preserves opaque reasoning state and compacts long conversations—hit 99.9% at $19K. That gap says Astra's reasoning chain has a lot of reusable intermediate work, and the right interface saves both money and time.

The behavioral detail is the more interesting part. Astra built compact symbolic world models from unfamiliar environments and invented its own shorthand to track state and plan. That's not brute-force search; it's internal rule-building, and it's a stronger signal than the score alone.

What's missing: parameter count, architecture, release date, and how many inference calls the 99.9% run actually used. ARC Prize defines AGI as acquiring any skill as efficiently as a human. On action efficiency, Astra already beats the median human on 96% of levels. On cost efficiency, we're nowhere close.

Source: Hacker News front page · arcprize.org
Published: Sep 4, 2026, 03:45 UTC+8 · 26 days ago