@HavenFeng: Today, we’re introducing [schema]: a harness reaching 99% RHAE with Opus 4.8 + Fable 5 and 95.35% with GPT-5.6 Sol on A…
Summary
Introducing [schema], a harness that achieves 99% RHAE with Opus 4.8 + Fable 5 and 95.35% with GPT-5.6 Sol on ARC-AGI-3 Public set, designed to make an LLM think like a physicist.
View Cached Full Text
Cached at: 07/16/26, 10:24 PM
Today, we’re introducing [schema]: a harness reaching 99% RHAE with Opus 4.8 + Fable 5 and 95.35% with GPT-5.6 Sol on ARC-AGI-3 Public set.
[schema] makes an LLM think like a physicist.
ARC-AGI-3 gives an agent a 64×64 grid plus legal actions, no rules, stated goal, or reward. The agent must discover both what the world is and how it works like a physicist:
- State grounding -> identify objects, relations, and goals.
- Mechanism discovery -> infer how these states change.
[schema] handles the state and mechanism in one editable program, a symbolic world model. It designs experiments to verify hypotheses, backtests the program against history, and plans inside its world at zero action cost.
[schema]’s saturation of the ARC-AGI-3 public set is only a starting point. There is much more to explore!
Full blog: http://schema-harness.github.io Agent traces: http://huggingface.co/datasets/schema-harness/arc-agi-3-schema-traces…
Amazing team effort with @guanningzeng, @Jiani_Wang_, @wenjie_ma, @shaofeng_y27736, @ChenyangWa70207, @lustralisk95, @akanazawa, @wodenimoni, @xiuyu_l and @Zanette_ai
@arcprize
@HavenFeng 129 days running opus 4.8 in production at http://monday.com. schema proves the harness matters more than the model - my 20 cron jobs prove the same thing daily. what does that 99% become when you swap opus for K3?
Similar Articles
Schema Harness Achieves ~99% on Arc‑AGI‑3 Public
Schema introduces a new harness that achieves ~99% on the ARC-AGI-3 Public set using frontier models like Claude Opus 4.8 and Fable 5, by improving the process around models rather than modifying weights.
@sethkarten: https://x.com/sethkarten/status/2072034978112889328
Continual Harness is a reset-free, self-improving agentic harness that achieves 20.54% on ARC-AGI-3 at a cost of $774 by storing memories, reusing skills, and refining its prompt, outperforming prior baselines like Hermes and OpenClaw with greater efficiency.
@Saboo_Shubham_: WILD times. Anthropic: Opus 5 beats GPT-5.6 on ARC-AGI-3 Tibo: I change two settings and GPT-5.6 Sol is now SoTA
Anthropic's Opus 5 beats GPT-5.6 on ARC-AGI-3, but Tibo claims GPT-5.6 Sol becomes SoTA with two setting changes involving multi-context reasoning and canonical compaction.
Schema (2 minute read)
Schema is a harness that enables frontier AI models to achieve 99% on the ARC-AGI-3 benchmark by having them write executable programs to model game environments, test predictions, and plan.
Opus 4.8 just broke ARC-AGI-3 (1 minute read)
A new benchmark called LisanBench evaluates LLMs on word chain tasks requiring planning, memory, and constraint adherence, with results showing strong performance from o3 and Anthropic models.