AI Aulis

AI NEWS IN ENGLISH · 18 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

32 stories · next issue tomorrow morning · How the paper is made →

← Issue 18 November 2026Paper

GPT-6.2 Astra scored up to 70.3 per cent or 100 per cent on ARC-AGI-3, depending on the harness

AI-generated contentWritten by AI

developers.openai.com, openai.com, aws.amazon.com, arcprize.org · 4 sources · Sources published: 17 November 2026 · Some source publication dates are unknown

OpenAI’s GPT-6.2 Astra scored up to 70.3 per cent in ARC-AGI-3 results verified by ARC Prize when it could carry forward notes it chose to keep. When the harness preserved the model’s reasoning state between requests, its best score was 100 per cent.

ARC Prize published the results on 17 November, the day OpenAI released the model. ARC-AGI-3 is a benchmark in which an AI agent works its way through game environments. A harness drives the agent: the program that gives the model its tools and runs its steps.

ARC Prize tested the model with two harnesses. The Standard harness lets the model keep notes it chooses itself. The Provider Adapter harness preserves the model’s opaque reasoning state between requests and uses compaction for long conversations. This lets the model reuse its earlier work, ARC Prize said.

The table shows verified ARC-AGI-3 results by reasoning effort.

Reasoning effort Standard harness Provider Adapter
Max 70.3% 99.8%
XHigh 56.5% 100.0%
High 58.0% 99.8%
Medium 44.0% 97.8%
Low 38.7% 95.1%

Source: ARC Prize, verified results, 17 November 2026.

At max effort the gap is 29.5 percentage points; at the other levels it is 42–56 percentage points. A single ARC-AGI-3 score therefore says nothing about a model’s ability unless you know which harness was used. With the Standard harness, scores also did not rise steadily with reasoning effort: XHigh scored lower than High.

The gap also shows in the 25 public demo environments for which ARC Prize reports per-game results. In BP35, the model scored 2.2 per cent at max effort with the Standard harness and 100 per cent with the Provider Adapter harness.

On the older ARC-AGI-1 and ARC-AGI-2 benchmarks, the model achieved its best results at max reasoning effort: 98.5 per cent and 95.8 per cent on ARC Prize’s semi-private test set.

OpenAI advises users of GPT-6.2 models to keep the start of the prompt unchanged and append new instructions to the end of the conversation. This can improve cache reuse and lower costs and latency, and it preserves reasoning across turns, the company said.

GPT-6.2 Astra is available in OpenAI’s API as gpt-6.2-astra. It costs $8 per 1 million input tokens and $40 per 1 million output tokens. ChatGPT Plus, Pro, Enterprise and Edu users get the model in ChatGPT Work and Codex; for eligible Enterprise and Edu users, it becomes available once a workspace admin enables it. AWS said the model is also available on Amazon Bedrock.

Previous coverage

Report an error in this story