The latest AI news and updates of Arena — Community-driven AI model evaluation platform with Arena leaderboards spanning code, text, vision, and search. Covering Arena's latest analysis, product updates, and launches from the past 90 days.
Arena19h agoArena launched its Fullstack Code Arena leaderboard, ranking 39 AI models on end-to-end web development tasks like multi-step reasoning and tool use. Kimi K3 (Max) takes the top spot with 1,664 points, followed by GPT-5.6 Sol (xHigh) at 1,633 and Claude Fable 5 at 1,623. These rankings are based on 22,969 community-driven blind battles.
Arena19h agoArena added Meta’s Muse Spark 1.1 to its Document and Vision leaderboards, where it reshapes the cost-performance Pareto frontier at $3.50 per million tokens. The model ranks #11 in Document Arena, up from #21, and #14 in Vision Arena, where it also secured a third-place finish in the Chinese language category.
ArenaJul 28Arena added OpenAI’s GPT-5.6 Terra and Luna (xHigh) to its Agent Arena leaderboard, where they rank #15 and #17. Arena’s evaluation reveals that Luna’s steep test-time scaling allows the smaller model to outperform the pricier Terra at lower reasoning efforts, proving that increased reasoning effort on cheaper models can yield superior outcomes on complex, long-horizon agentic tasks.
ArenaJul 27Arena reports that Anthropic’s Claude Opus 5 with Max reasoning has debuted at #1 on both the Frontend Code Arena and the factuality-weighted Text Arena. The model’s High reasoning configuration ranks #3 in frontend coding and #2 in factuality-weighted text. These preliminary results are based on community-driven blind battles measuring real-world agentic and reasoning performance.
ArenaJul 27Arena ranks Moonshot AI’s Kimi K3 (Max) as the top open-weight model across its Agent, Frontend Code, and Text leaderboards. In the Agent Arena, the model achieved a +9.75% net improvement, surpassing the previous leader, GLM-5.2 (Max). Kimi K3 also leads five individual agentic performance signals, including confirmed task success and tool hallucination rates.
ArenaJul 22Arena ranks Tencent’s Hy3 model fifth among open-weight systems on its Agent Arena leaderboard, with a net improvement of -2.2% across 8,000 sessions. The model shows strength in bash recovery but struggles with steerability. In the Frontend Code Arena, Hy3 debuts as the second-ranked open-weight model, placing sixteenth overall across all evaluated systems.
ArenaJul 20Arena added Thinking Machines’ Inkling model to its Agent Arena leaderboard, where it debuts at #9 among open-weight models and #30 overall. The model ranks as the top U.S. open-weight system for long-running agentic tasks, showing strong CLI error recovery but trailing in user sentiment, steerability, and task completion.
ArenaJul 20Arena ranks Moonshot AI’s Kimi K3 fourth on its Agent Arena leaderboard with a 9.6% net improvement, tying Claude Opus 4.8 and GPT-5.6 Sol. The model leads in confirmed task success across 8,000 sessions. Moonshot AI plans to release the full model weights by July 27, which would establish Kimi K3 as the top-ranked open-weight model.
ArenaJul 16Arena ranks Moonshot AI’s Kimi K3 first on its Frontend Code Arena leaderboard with 1,679 points, a 17-place jump from its predecessor. The model leads in six of seven coding domains, including Data & Analytics and Brand & Marketing. Moonshot AI plans to release the full model weights by July 27.
ArenaJul 15Arena launched a factuality-weighted leaderboard toggle for its Text and Search Arenas, using 2 million verified claims from 170,000 real-world battles. The ranking combines human preference with factual accuracy, showing that OpenAI models consistently improve in factuality over time, while other providers often see scores decline when factuality is prioritized.
ArenaJul 14Arena added Meta's Muse Spark 1.1 to its Agent Arena leaderboard, where it ranks 17th among models and 5th across all labs. The model shows a net improvement of 0.17% over the average model across 9,262 real-world agentic sessions. It currently ranks above Gemini 3.1 Pro and Qwen-3.7 Plus, but below Grok 4.5 and GLM 5.2.
ArenaJul 14Arena shared research from PhD candidate Melissa Pan demonstrating that full system configuration can cut agent costs by 89% while maintaining 100% accuracy. The findings show that optimizing the entire system configuration outperforms LLM routing alone. The research introduces BRANE, a system designed to achieve these cost-efficiency gains in agentic pipelines.
ArenaJul 13Arena ranks OpenAI’s GPT-5.6 Sol second on its Agent Arena leaderboard, based on 7,800 real-world agentic sessions. The model achieves a 1.6% net improvement over GPT-5.5 (xHigh), though it trails the top-ranked Claude Fable 5 in user satisfaction, scoring 10.9% in praise versus complaint compared to 17.3%. The leaderboard evaluates performance on long-horizon tasks using causal tracing.
ArenaJul 11Arena.ai reports that ByteDance’s Seedream 5.0 Pro model has debuted on its leaderboards, securing the second spot in Multi-Image Edit with 1,415 points. The model also reached fourth in Image Edit and eleventh in Text-to-Image, marking significant performance gains over the previous Seedream 4.5 version. The model is now available via the BytePlus API.
ArenaJul 10Arena.ai ranks OpenAI’s GPT-5.6-Sol-xHigh joint first on its Code Arena: Frontend leaderboard with a 1,636 score. The model climbed from eighteenth place, pricing at $5 per million input tokens and $30 per million output tokens. It also secured top rankings in subcategories including Data & Analytics, Brand Marketing, Consumer Product, and Gaming.
ArenaJul 9Arena.ai ranked SpaceXAI's Grok-4.5 third on its Code Arena: Frontend leaderboard with a 1,572 score. This marks a significant improvement from the predecessor Grok-4.3, which previously held the 62nd spot. The model now sits alongside GLM-5.2 (Max) and Claude Opus 4.8 (Thinking) in community-driven blind evaluations of frontend web development and agentic coding tasks.
ArenaJul 7Arena.ai reports that Meta’s Muse Image model now ranks second across Text-to-Image, Single-Image Edit, and Multi-Image Edit leaderboards, trailing only OpenAI’s GPT Image 2. Additionally, Meta’s Muse Video model debuted at third place in the Text-to-Video Arena with a score of 1,459, based on community-driven head-to-head evaluations.
ArenaJul 5Arena published a video demonstrating Claude Fable 5 generating 60+ complex 3D scenes, including modern cities, art, and world wonders, evaluated by researcher Peter Gostev. Arena describes the model as being in a league of its own for 3D generation tasks.
ArenaJul 2Arena launched Fullstack Code Arena, expanding its evaluation platform from frontend prototypes to fullstack development. The update adds database integration, third-party API access, and persistent dev servers with hot reloading. AI models now operate as agents using bash and web search tools to build, iterate, and deploy real-world applications directly to Vercel.
ArenaJul 2Arena.ai ranked Anthropic's Claude Sonnet 5 (Thinking) at #6 on its Code Arena: Frontend leaderboard. The model outscored its predecessor Sonnet 4.6 by 29 points and the prior-generation flagship Opus 4.6 (Thinking) by 9 points. It also placed #11 in Document, #17 in Search, #21 in Vision, and #32 in Text arenas.
ArenaJun 30Arena ranked Google DeepMind’s newly released Gemini Omni Flash second on its Video Edit leaderboard with a score of 1,347. The model sits nearly 40 points above the third-ranked HappyHorse 1.0 in a category currently featuring seven models.
ArenaJun 30Arena.ai added Google DeepMind's Gemini 3.1 Flash Lite Image to its Text-to-Image leaderboard, where it debuted at #5 with a 1251 score at $0.034 per image. The model lands on the Pareto frontier, offering near-flagship quality at budget pricing. It also entered the Image Edit Arena at #9 for multi-image and #15 for single-image editing.
ArenaJun 28Arena.ai released a token-efficiency analysis for Agent Arena, mapping performance improvement against median output tokens. Anthropic’s Fable leads with a 14.1% improvement at token levels similar to Opus 4.8 Thinking. OpenAI’s GPT-5.5 models outperform leading Claude models in efficiency, while Grok Build 0.1 consumes over 20,000 tokens for a negative net improvement.
ArenaJun 25Arena.ai reports Z.ai's GLM-5.2 (Max) climbing its Code Arena: Frontend leaderboard to a 1,595 Elo score. This trajectory marks a rapid compression of the performance gap between open-source models and frontier systems like Claude Fable 5, which currently holds a 1,665 score on the platform.
ArenaJun 25Arena added Alibaba's Wan-2.7 I2V to its Image-to-Video leaderboard, where it debuted at number five with an Elo score of 1,434. Based on 1.3 million head-to-head community votes, the model ranks ahead of xAI's Grok Imagine Video (720p) and every Google Veo-3.1 variant currently on the platform.
ArenaJun 22Arena.ai ranks ByteDance's Seed 2.1 Pro Preview eighth on its Code Arena: Frontend leaderboard with a 1539 Elo score. The model, currently in early access, places in the top 10 across five subcategories, including #7 for React and #6 for Brand & Marketing. It will be publicly available in a few weeks.
ArenaJun 18Arena detailed the causal tracing methodology powering its Agent Arena leaderboard, which evaluates AI agents on real-world tasks. The framework uses five signals—confirmed success, praise versus complaint, steerability, bash recovery, and tool hallucination—to calculate net performance improvements. This approach treats agent sessions as multi-component systems to isolate the impact of specific model and tool selections.
ArenaJun 18Arena added 10 models to its Agent Arena leaderboard, where Z.ai’s GLM-5.2 (Max) entered the top 10 with a 9.4% improvement in confirmed task success. Anthropic’s Claude Fable 5 briefly debuted at #1 across nearly every metric before a U.S. government directive suspended access, leaving the model as an inaccessible performance benchmark for the current frontier.
ArenaJun 18Arena ranks Moonshot AI's Kimi K2.7 Code #19 overall and #6 among open models on its Agent Arena leaderboard. The model demonstrates strong confirmed task success, though steerability regressed by 12.25% compared to the K2.6 version. These scores remain subject to wide confidence intervals as the evaluation data stabilizes.
ArenaJun 17Arena’s Agent Arena leaderboard ranks MiniMax M3 at #18 overall and #5 among open models. The model improves over its predecessor, M2.7, with gains in confirmed task success and bash error recovery. It also ties for first place in tool hallucination, demonstrating high discipline in tool use across real-world agentic workflows.
ArenaJun 16Arena.ai ranks Z.ai's GLM-5.2 (Max) second on its Code Arena: Frontend leaderboard with a score of 1,595. The model outperforms Claude Opus 4.7 (Thinking) by 29 points and leads six of seven domain-specific sub-categories, including Data & Analytics and Gaming. It leads competing open-weight models like Kimi-K2.6 and Minimax-M3 by a large margin.
ArenaJun 15Arena.ai added Moonshot AI's Kimi-K2.7-Code to its Code Arena: Frontend leaderboard, where it ranks third among open-weight models and nineteenth overall. The ranking reflects blind community evaluations of the model's performance on real-world agentic frontend coding tasks, including HTML and React development.
ArenaJun 15Arena.ai added Anthropic's Claude Opus 4.8 to its Agent Arena leaderboard. With thinking enabled, the model ties for first place with a 9.1% net improvement in task completion. However, the model shows regressions in steerability and bash recovery, while the non-thinking variant logs one of the highest tool hallucination rates on the platform.
ArenaJun 13Arena.ai added NVIDIA's Nemotron 3 Ultra to the Agent Arena leaderboard, where it ranks #20 overall and #5 among open models. The model shows strong tool-use discipline, tying for #1 in tool hallucination, but struggles with steerability and bash recovery. These scores, based on 2,849 sessions, remain subject to wide confidence intervals as data stabilizes.
ArenaJun 13Arena.ai ranks OpenAI's GPT-5.5 (xHigh) second on its Agent Arena leaderboard with a +10.6% net improvement over the average model. It tops three signals — praise-versus-complaint (+29.4%), bash recovery (+14.1%), and tool hallucination (+2.1%) — but ranks lower on confirmed success (#9, +5.4%) and steerability (#11, +1.9%), across 160,000 real-world agentic tasks over seven days.
ArenaJun 13Arena.ai has ranked Google DeepMind’s Gemini Omni Flash as the top model on its Video Arena leaderboard for both Text-to-Video and Image-to-Video. The model achieved a 1,527 Elo score, securing a 61-point lead over the next-best model, Seedance 2.0. In head-to-head battles, Gemini Omni Flash won 82% of its matches, excluding ties.
ArenaJun 13Arena.ai reports Claude Fable 5 now ranks first in Code Arena: Frontend, winning 72% of battles with a 98-point lead, and first in Text Arena. The model also secured second place in Vision Arena. These rankings follow the model's recent top performance in the Agent Arena, where it outperformed other frontier models by the widest margin recorded on the platform.
ArenaJun 13Arena.ai ranks Anthropic's Claude Fable 5 first on its Agent Arena leaderboard with an 11.2% net improvement. The model leads in confirmed task success and user praise, though it ranks 17th in steerability. It outperforms Opus-4.8 and GPT-5.5 by the widest margin recorded on the platform, demonstrating high capability for complex, multi-step agentic workflows.
ArenaJun 10Arena.ai has made Anthropic's Claude Fable 5 model available in its Agent Mode, allowing users to test its agentic capabilities on real-world tasks and contribute to the Agent Arena leaderboard. This integration enables community-driven evaluation of Claude Fable 5's autonomous planning and tool-use in complex, multi-step workflows.
ArenaJun 9Arena.ai's new Agent Arena leaderboard places xAI's Grok Build 0.1 at #15 and Grok 4.3 (High) at #17. Grok Build 0.1 demonstrates improved bash capability and looks to be successfully completing tasks more often overall than Grok 4.3, though it is slightly less steerable and more prone to tool hallucinations.
ArenaJun 5Arena.ai has integrated Mistral AI's Mistral 3.5 model into its Agent Mode, enabling users to test its performance on complex, multi-step tasks. User sessions contribute to the Agent Arena leaderboard, which evaluates agentic AI models on their ability to autonomously plan and execute real-world workflows.
ArenaJun 5Arena.ai introduced Agent Mode and the Agent Arena leaderboard to evaluate agentic AI models. This provides a new standard for measuring how AI agents perform complex, multi-step tasks in real-world scenarios, moving beyond single-turn chat assessments.
ArenaJun 5Arena.ai's Image Arena Top 10 Text-to-Image leaderboard saw three new models enter its ranks this past month: Reve 2.0 at #2, MAI-Image-2.5 at #4, and Ideogram 4.0 Quality at #9. Ideogram 4.0 Quality is the only open-weights model in the top 10. This shift highlights continuous performance improvements in image generation, with new versions displacing their predecessors.
ArenaJun 5Arena.ai has integrated NVIDIA's Nemotron 3 Ultra model into its Agent Mode, enabling users to run the model for complex, multi-step tasks. These sessions contribute to the new Agent Arena leaderboard, which evaluates agentic AI models on real-world performance using tools like web search and terminal. This expands the range of frontier models available for practical agentic workflows and provides new data for understanding their capabilities in autonomous tasks.
ArenaJun 4Arena.ai introduced Agent Mode, a new feature for its evaluation platform that allows users to test frontier AI models on complex, multi-step tasks using integrated tools. It shifts evaluation beyond single-turn chat to measure how models autonomously plan and execute real-world workflows, providing a new standard for agentic AI performance.
ArenaJun 4Arena.ai introduced Agent Arena, a new leaderboard that evaluates agentic AI models on their ability to perform complex, real-world tasks using tools like web search and terminal. It measures performance across five signals, including task success and error recovery, with OpenAI's GPT-5.5 (High) and Anthropic's Claude-Opus-4.7 (Thinking) leading the initial rankings. It gives a live read on how agents perform in practical, multi-step workflows.
ArenaJun 4Arena.ai has placed Reve 2.0 in the second spot on its Text-to-Image leaderboard following a significant performance jump. The model's 125-point improvement allows it to outperform flagship image generators from Google and Microsoft in human-preference testing.
ArenaJun 4Arena.ai has placed Ideogram 4.0 Quality at number eight on its Text-to-Image leaderboard with an Elo score of 1,204. The ranking establishes the model as the highest-rated open-weights system, rivaling proprietary performance from Google and OpenAI.
ArenaMay 31xAI's Grok-Imagine-Video-1.5-Preview (720p) has reached the #1 position on the Arena Image-to-Video leaderboard with an Elo score of 1,473. The model unseated previous leaders from ByteDance and Alibaba, marking a significant jump in human-preferred video generation quality.
ArenaMay 28Arena.ai introduced seven domain-specific categories to its Code Arena: WebDev leaderboard after analyzing 250,000 user prompts. The new views reveal that aggregate scores hide significant performance gaps, with specific models excelling at aesthetic design while others dominate logical simulations.
ArenaMay 26Arena.ai officially ranked Microsoft's MAI-Image-2.5 model at #2 in its Image Edit leaderboard with a score of 1401, advancing the Pareto frontier for generative quality. The model outperformed high-fidelity offerings from xAI and OpenAI by 10 points in blind human-preference testing.
ArenaMay 26Alibaba's Qwen3.7-Max debuted at #4 on the Arena.ai frontend coding leaderboard, establishing it as the highest-ranked model from a Chinese lab. The results place the model on par with Anthropic's Claude Opus 4.6 for agentic web development tasks at a significantly lower price point.
ArenaMay 21HiDream-01-Image debuted at #27 overall on the Arena.ai Text-to-Image leaderboard, securing the #4 spot among open-source models. The ranking validates the performance of its unified transformer architecture against proprietary systems from OpenAI and Google.
ArenaMay 21Arena.ai released a three-year analysis of the price-performance Pareto frontier, revealing that frontier-level intelligence now costs roughly $0.10 per million tokens. The data shows the performance gap between budget and flagship models has nearly collapsed, shifting the market toward high-efficiency reasoning.
ArenaMay 19Gemini 3.5 Flash has entered the Arena.ai leaderboards with a ninth-place ranking in both the overall Text and Frontend Coding categories. The model establishes a new price-performance frontier by delivering a 70-point jump in coding capability over its predecessor.
ArenaMay 18Alibaba's Qwen3.7 Max and Plus preview models have debuted on the Arena.ai leaderboards, ranking #13 in text and #16 in vision. The results establish Alibaba as a top-six global AI lab with specific strengths in math, software engineering, and expert-level reasoning.
ArenaMay 15Arena.ai's latest Text Arena data reveals that the performance gap between top US and Chinese AI models has shrunk from 278 to just 29 Elo points in three years. This real-world evidence confirms that Chinese labs have reached near-parity with frontier US systems despite hardware restrictions.
ArenaMay 15Anthropic has surpassed OpenAI in business customer adoption with a 34.4% market share according to fintech data from Ramp. Arena.ai notes that its community-driven leaderboards predicted this shift six months in advance, with Anthropic taking the top spot in human preference rankings in late 2025.
ArenaMay 12Arena.ai released its latest Text Arena rankings based on over 6 million community votes, placing Anthropic's Claude Opus 4.7 Thinking at the top of the leaderboard. The data reveals that while overall scores are tightening, models are developing specialized strengths in areas like creative writing, math, and expert-level reasoning.
ArenaMay 9Arena.ai added OpenAI's GPT-5.5 Instant to its blind evaluation leaderboards, revealing the model's performance across text, vision, and specialized professional categories. The results show the model excels in multi-turn dialogue but lags behind high-tier variants in raw reasoning and document analysis.
Frequently asked questions