ARC-AGI

Trial

Platforms

A benchmark suite for evaluating AI systems on general intelligence and adaptive problem-solving.

Why it's here

Placed in Trial: 10 article(s) of evidence from 6 source(s), led by research-stage coverage, with 6 in the last 30 days. Confidence 86%.

Evidence (10)

7Hacker News·9/9/2026research
GPT-6 Astra, Looped Transformers, and Hidden Reasoning
The article reviews early impressions of OpenAI’s GPT-6 Astra, describing it as a strong model with especially notable performance on math, coding, and 3D/rendering tasks. It then explains looped or recurrent transformer blocks and discusses whether they relate to hidden chains of thought, alongside recent research on the topic.
8OpenAI Blog·9/3/2026model_release
OpenAI rolls out GPT-6 Astra
OpenAI is rolling out GPT-6 Astra to a limited set of organizations, with broader availability coming soon to ChatGPT Plus, Pro, Business, Enterprise, the API, and AWS. The model is priced at $10 per million input tokens and $50 per million output tokens, and is reported to score strongly on several benchmarks, especially long-context and security tasks.
9The New Stack·9/3/2026research
GPT-6 Astra Tops ARC-AGI-3 Benchmarks
OpenAI’s GPT-6 Astra posted state-of-the-art results on ARC-AGI-3, reaching 62.7% with the Standard harness and 99.9% with the Provider Adapter harness. The benchmark measures agentic intelligence in novel turn-based environments, and Astra also showed strong action efficiency versus human testers.
9InfoQ·9/3/2026model_release
OpenAI Posts System Card for GPT-6 Astra
OpenAI has published a system card for GPT-6 Astra, outlining safety, capability, and deployment considerations for the model. The Hacker News discussion highlights reported gains on ARC-AGI-3 and the Artificial Analysis Coding Agent Index, indicating strong interest in its performance and safety profile.
5Martin Fowler·9/1/2026research
Fragments on AI prose, long-horizon agents, and CI with agents
Martin Fowler highlights a tool for detecting common LLM prose patterns and notes research suggesting humans are poor at distinguishing AI-generated text from human writing. He also discusses NVIDIA’s AVO-based long-horizon agent work and Paul Stack’s argument that agent-driven development changes where verification should happen in CI pipelines.
8Hacker News·9/1/2026research
Small Transformer Trained in 1.5 Hours Beats Many LLMs on ARC
The author reports training a small transformer from scratch in about 1.5 hours on a 5090 GPU and claims it matches or beats several larger models on ARC-style evaluation. The updated approach adds architectural and training changes such as SwiGLU, RMSNorm, fewer augmentations, and using ARC-2 non-overlapping tasks while filtering duplicates to avoid data leakage.
8The New Stack·8/21/2026research
Nvidia says AVO lifts Claude Opus 5 from 30.2% to 100% on ARC-AGI-3
Nvidia detailed how its Agentic Variation Operators (AVO) system sustains long-running autonomous work and applied it to the ARC-AGI-3 benchmark. The company says AVO raised Claude Opus 5 from a reported 30.2% baseline to 100% on the public ARC-AGI-3 set, arguing that agent system design can matter as much as model capability.
8Hacker News·8/7/2026research
DeepSeek V4 Flash 0731 posts strong ARC-AGI benchmark results
DeepSeek announced V4 Flash 0731 with three reasoning variants and published benchmark results on ARC-AGI-1 and ARC-AGI-2. At maximum effort, the model reportedly scored 89.0% on ARC-AGI-1 and 61.4% on ARC-AGI-2 with low per-task costs. The post includes detailed task-level pass/fail results and links to the paper and model page.
7OpenAI Blog·7/29/2026research
Two settings tripled GPT-5.6 scores on ARC-AGI-3
OpenAI says two API settings significantly improved GPT-5.6 performance on the ARC-AGI-3 benchmark by preserving reasoning state and enabling compaction. The change increased scores and efficiency, showing how inference-time configuration can materially affect benchmark results.
6Hacker News·7/25/2026research
ARC-AGI Leaderboard Tracks AI Reasoning Efficiency
The ARC-AGI leaderboard compares AI systems across ARC-AGI-1, ARC-AGI-2, and ARC-AGI-3, with a focus on performance versus cost per task. It highlights how different model types, including reasoning systems, base LLMs, and Kaggle submissions, perform under varying compute constraints.

Compare

Is this placement right?

Your vote is a weighted signal that nudges the radar — never a direct override.

Ring: Trial
Confidence: 87%
Momentum: cooling
Evidence: 10

Role implications

AssessTrialConfidence68%

Business✦ Emerging
For BD, ARC-AGI is becoming a useful talking point to demonstrate reasoning capability in platform or model conversations with technical buyers. But because it is only at Trial, BD should use it as supporting evidence for trials and POCs, not as a broad commercial promise.
Developers✦ Emerging
ARC-AGI is moving from assessment to trial, so dev teams will start treating it as an additional check on reasoning and synthesis, not just a benchmark to glance at. The recent news shows both small and large models being tested on ARC, so developers need repeatable evaluation workflows and side-by-side model comparisons rather than relying on a single score.
Product→ Steady
PMs will start using ARC-AGI as a reference signal when judging whether a model is trustworthy enough for reasoning-heavy product flows, but it should not become the only decision criterion. The recent results show the benchmark is attracting many different trials, so PMs should read it as an indicator of execution capability rather than a promise of end-user experience.
Testers / QA▲ Rising
With ARC-AGI in Trial, QA/test teams will need to treat it as an important behavioral test set for models that must reason across multiple steps. That increases the need for test cases, model-version comparisons, and stability checks on the same task, rather than only measuring broad accuracy.

as of 9/2/2026

Educated guesses from the trend — you decide.

History

trial9/2/2026
assess8/24/2026
trial8/22/2026
assess7/26/2026