Coding Horizon

NVIDIA Just Proved Your AI Model Is Not Enough

Every figure, name and claim the finished picture puts on screen, chased to a primary source.

The ARC AGI 3 baseline

• Claude Opus 5 scores 30.2% on ARC AGI 3, at High reasoning effort, on the Public Demo set of 25 interactive environments. ARC Prize lists it as the highest performing model on that benchmark as of 24 July 2026, and notes it completed five Public Demo environments no model had previously beaten. Source: ARC Prize, Claude Opus 5 results — https://arcprize.org/results/anthropic-claude-opus-5
• The previous high score on ARC AGI 3 was 7.8%, set by GPT 5.6 Sol at Max reasoning effort. ARC Prize also reports Anthropic’s Fable class models at approximately 20% on the Public Demo environments. Source: ARC Prize — https://x.com/arcprize/status/2080716561539907928
• Opus 5’s ARC AGI 3 run was evaluated only at High effort because of a short testing window, while the previous record holder was tested at Max. Source: ARC Prize, Claude Opus 5 results (above).
• ARC AGI 3 is an interactive benchmark rather than a question and answer test: the system is dropped into small game environments with rules it is not told, and has to explore, infer and act. Source: ARC Prize — https://arcprize.org/arc-agi/3/

The NVIDIA AVO result

Primary source for everything in this section: NVIDIA Technical Blog, “NVIDIA AVO Reaches 100% on ARC-AGI-3, Demonstrating a Frontier-Level General-Purpose Architecture for Long-Horizon Autonomous Agents” — https://developer.nvidia.com/blog/nvidia-avo-reaches-100-on-arc-agi-3-demonstrating-a-frontier-level-general-purpose-architecture-for-long-horizon-autonomous-agents/

• AVO stands for Agentic Variation Operators.
• 100.00 RHAE score across all 25 environments in the ARC-AGI-3 public set, completing all 183 levels.
• The run took 6,624 environment actions, about 12% fewer than VISTA’s reported 7,542 actions for the same 183 levels.
• The model inside the AVO system was Claude Opus 5. NVIDIA cites ARC Prize’s separately reported figure of approximately 30% for Claude Opus 5 at High reasoning effort as the model baseline.
• NVIDIA states the comparison is not a controlled ablation: “the two systems differ in agent backend, observation representation, memory, context management, and other implementation details.”
• One concrete difference given: VISTA used a 512x512 PNG observation, AVO used a text only 64x64 grid.
• The result is on the public set only. ARC Prize’s semi private and private held out sets were not run.
• The supervisor: “The supervisor monitors the broader trajectory for stagnation or repeated unproductive cycles and can redirect the main agent toward alternative strategies when needed.”
• AVO’s origin is engineering search, not games. In GPU kernel optimisation work it explored more than 500 optimisation directions and produced 40 committed kernel versions over seven days, reaching up to 10.5% better performance than FlashAttention 4.
• AVO was first introduced as a general purpose coding agent system in late March 2026.
• No open source release of AVO is announced in the post.

Tools named on screen

• Cursor — AI code editor, https://cursor.com
• Claude Code — Anthropic’s terminal coding agent, https://www.anthropic.com/claude-code
• GitHub Copilot — https://github.com/features/copilot
• NVIDIA — https://www.nvidia.com
• ARC Prize — https://arcprize.org

Caveats

• “Different execution loop” as one of the ways the two setups differ is a fair reading of NVIDIA’s “and other implementation details”, but NVIDIA’s own list names agent backend, observation representation, memory and context management explicitly and does not use that phrase.
• The seventy point gap between 30% and 100% is arithmetic on two numbers NVIDIA itself says are not a controlled comparison. It is not an attributable AVO gain.