ARC Prize verifies Fable 5.2 and Gemini 4 Carbon scores on ARC-AGI-3, 1 to 3 benchmarks
Headline scores now depend on the test harness more than on the model, with gaps of 16 and 34 points between harnesses. Enterprises choosing models for agentic work should demand harness-matched results and cost and runtime data rather than rely on the best published figure. The runtime-limit finding points to latency and compute cost as practical constraints even where accuracy is high.
What Happened
ARC Prize verified Claude Fable 5.2 at 97.7% on ARC-AGI-3 with the provider adapter harness and 81.2% with the standard harness, and said this is a new high in the standard harness. At max reasoning with the standard harness, 28 of 55 semi-private games reached the 12-hour runtime limit. For Google's Gemini 4 Carbon it verified 99.1% with the provider adapter and 65.1% with the standard harness, and said Carbon beat its predecessor Argon by 18 points on ARC-AGI-3 in the standard harness.
Why It Matters
Headline scores now depend on the test harness more than on the model, with gaps of 16 and 34 points between harnesses. Enterprises choosing models for agentic work should demand harness-matched results and cost and runtime data rather than rely on the best published figure. The runtime-limit finding points to latency and compute cost as practical constraints even where accuracy is high.
Sources
- ARC Prize (X) (opens in a new tab): Claude Fable 5.2 from @AnthropicAI on ARC-AGI (Verified) (2026-11-10)
- ARC Prize (X) (opens in a new tab): Gemini 4 Carbon from @Google on ARC-AGI (Verified) (2026-11-11)
More From This Briefing
- Axios: Trump again backs curbs on Nvidia chip diversion to China while Lutnick stalls at Commerce
- Warren and Banks urge Commerce to close chip loopholes; Epoch estimates 1.6m H100-equivalents reached China
- Epoch AI reports Claude Fable 5.2 sets a new Epoch Capabilities Index record in pre-launch testing
- EU publishes Implementing Regulation 2026/2487 setting common rules for AI regulatory sandboxes
More on Frontier AI
- Google lists Gemini 4 Carbon and a 'cyber-permissive' variant as released on 10 November (11 November 2026)
- Anthropic releases Claude Fable 5.2 and Mythos 5.2, cutting prices ahead of its planned IPO (10 November 2026)
- Anthropic system card: Mythos 5.2 and Fable 5.2 often voice suspicion they are being evaluated (10 November 2026)
- METR estimates Anthropic's benchmark progress has accelerated 1.9x, driven mostly by AI-assisted R&D (7 November 2026)
Key Terms
- AI agent: An AI agent is a system in which an AI model plans and carries out multi-step tasks, using tools such as browsers, code, email or business...
- Artificial general intelligence (AGI): Artificial general intelligence refers to AI systems that match or exceed human ability across most cognitive tasks. There is no single...
- AI benchmark: An AI benchmark is a standard set of tasks used to measure and compare model performance, for example on coding, mathematics, reasoning or...
Filed under Frontier AI Models, ARC Prize Foundation.
This story is part of the Laurenceo Intelligence briefing for 12 November 2026, a daily sourced briefing on AI, technology and enterprise risk published in Bermuda by Laurenceo, founded by Jorai Laurenceo. It is produced by an automated research pipeline and checked before publication; how this is made. General information only, not legal, regulatory or financial advice.