Google introduces Gemini 4 Neon, positioned for frontier-level intelligence and fast end-to-end task completion across complex software engineering, knowledge work, and science workflows. It scores 61 on the Artificial Analysis Intelligence Index for #3 overall, ranks #2 on the Vals Index on the Pareto frontier, and places #2 on the Agent Arena steerability leaderboard.
Steerability
可控性
The ability of a model to precisely adjust output attributes (e.g., style, motion) according to user instructions.
Curated Intelligence
Google has released Gemini 4 Neon, positioned as a frontier-intelligence model built for fast end-to-end task completion across complex software engineering, knowledge work, and science workflows. Neon scores 61 on the Artificial Analysis Intelligence Index, ranking third overall, ranks second on the Vals Index and sits on its Pareto frontier, and places second on the new Agent Arena steerability ranking for adapting to mid-task user corrections.
Thinky Machines released Inkling 2 Small as open weights today; it ranks #2 among all open models and #22 overall in Agent Arena, with a +2.9% net improvement across 11.5k real-world agentic sessions. It leads open models in steerability (+10.9%) and costs a median $0.26 per task, roughly 4.7x less than top-ranked Kimi K3.1 Pro.
Anthropic released Claude Fable 5.2 (Max), which takes #1 on Arena.ai's Agent Arena with a +30.4% net improvement, roughly 1.6x the #2 model Opus-5.5, and ranks first in Code, Chat and Work. Versus Fable 5.1 it climbs from #5 to #1, cuts median cost per task from $7.7 to $7.3, and lowers API pricing 20% to $8/$40 per million input/output tokens. Results draw on 4.7k real-world agentic sessions, with the largest gains in user satisfaction and steerability.
Moonshot AI has released the open weights for Kimi K3.1 Pro, which now ranks #1 among open models in Agent Arena. The new version shows a 5.1% net improvement over its predecessor, leading specifically in Bash reliability, user sentiment, and steerability across long-horizon agentic tasks.
Tencent Hunyuan's Hy4 ranked #14 overall on Agent Arena across 8,500+ real-world long-horizon agentic sessions, up 3.6 points from its preview and joining the Pareto frontier. At $0.55 per task it costs roughly 62% less than Kimi K3.1 Pro Max ($1.46) while scoring within 1 point. Tencent says Hy4 will tie Kimi K3.1 Pro for #1 among open models once weights are released; it is currently proprietary.
OpenAI's latest model, GPT-6.1 Terra (Max), ranks #8 in the Agent Arena benchmark, showing a 10.6% net improvement over previous versions. It significantly optimizes task execution costs (median $0.62) and excels in Bash reliability and steerability, reshaping the Pareto frontier of performance versus cost.
Google DeepMind released Gemini 4 Carbon, ranking #2 in Agent Arena with a 16.2% net improvement on real-world tasks. The model sits on the Pareto frontier for cost and efficiency, showing significant gains in task completion and steerability over its predecessor, Argon.
Google DeepMind's Gemini 4 Carbon (High) debuted strongly on Arena, ranking #1 in Text, #2 in Agent, and #4 in Code: WebDev. It demonstrates significant improvements in real-world agentic success rates and steerability while offering a 72% cost reduction compared to top competitors, reshaping the frontier model Pareto frontier.
Meta's latest model, Muse Wonder Full (Max), has climbed to #4 in the Agent Arena, marking a significant leap over previous versions. It achieved a +11.3% net improvement across real-world agentic sessions and ranked within the top 10 for code, chat, and work categories, demonstrating enhanced steerability and tool reliability.