Peter tests Anthropic’s Fable 5.2 across 100+ generations to see how it compares with Opus in practice—from detailed 3D worlds and playable games to website UI, data visualization, and code-generated animation.
He explores where Fable 5.2 improves scene coherence, camera movement, and visual explanations, and why website UI shows less obvious progress. Highlights include a $2,618 ancient Rome project, Tower Bridge rendered in Blender, and 100 different ways to draw his dog. He also compares generation costs and shows where Opus still stands out.
00:00 Fable 5.2: First impressions 01:09 3D realism and scene coherence 04:50 Pricing and cost comparisons 07:20 Realism and world understanding 10:33 Camera work and depth of field 11:53 Artistic understanding and Mondrian 13:48 Tower Bridge with Blender 14:45 The $2,618 ancient Rome project 19:16 Games that feel more playable 27:05 Website UI: Limited progress 31:20 Clearer data visualizations 33:43 Explaining complex ideas visually 38:08 Storytelling: Fable vs. Opus 42:16 Dynamic cartoons and camera movement 44:48 Crochet, stop-motion, and visual styles 47:40 Pushing SVG further 48:37 Creative diversity: 100 ways to draw a dog 50:32 Final verdict
Peter tests Anthropic’s Fable 5.2 across 100+ generations to see how it compares with Opus in practice—from detailed 3D worlds and playable games to website UI, data visualization, and code-generated animation.
He explores where Fable 5.2 improves scene coherence, camera movement, and visual explanations, and why website UI shows less obvious progress. Highlights include a $2,618 ancient Rome project, Tower Bridge rendered in Blender, and 100 different ways to draw his dog. He also compares generation costs and shows where Opus still stands out.
00:00 Fable 5.2: First impressions 01:09 3D realism and scene coherence 04:50 Pricing and cost comparisons 07:20 Realism and world understanding 10:33 Camera work and depth of field 11:53 Artistic understanding and Mondrian 13:48 Tower Bridge with Blender 14:45 The $2,618 ancient Rome project 19:16 Games that feel more playable 27:05 Website UI: Limited progress 31:20 Clearer data visualizations 33:43 Explaining complex ideas visually 38:08 Storytelling: Fable vs. Opus 42:16 Dynamic cartoons and camera movement 44:48 Crochet, stop-motion, and visual styles 47:40 Pushing SVG further 48:37 Creative diversity: 100 ways to draw a dog 50:32 Final verdict
#arenaai #anthropic #llmAnthropic’s IPO + Gemini 4 Neon Debuts at #2Arena2026-11-20 | Try the latest models at arena.ai/agent
Anthropic's IPO leads this week's news, alongside Gemini 4 Neon (High) debuting at #2 on the Agent Arena Index. Plus Fable 5.2 versus GPT-6.2, Astra versus Sol, Google's shrinking release gaps, Qwen4-Max, Microsoft & image editing, Amazon leading Search Arena, and Meta Ember reshaping the frontier.
Follow Arena on X: https://x.com/arena
NEWS: November 13–November 20, 2026
- The linked Agent Arena Index snapshot shows Claude Fable 5.2 (Max) leading with 1,448, followed by Claude Opus 5.5 (Max) at 1,336 and Gemini 4 Neon (High) at 1,333. This is a November 19 retrieved snapshot, not the November 15 date of the original leaderboard link. - Anthropic prices its Class A IPO at $83 per share in its November 18 announcement. "Before the IPO: 10 months of progress" connects the announcement to the earlier model evolution through Claude Opus 4.5 versus Claude Opus 5.5 (Max), followed by a gallery of nine model configurations from November 2025–September 2026. November 20 update: Anthropic confirmed the IPO closed at the same $83 offering price; its shares began trading November 19 under ANT. - Fable 5.2 versus GPT-6.2: Chocolate Factory Garden from First Impressions pairs Claude Fable 5.2 (Max) on the left with GPT-6.2 Astra (Max) on the right. The comparison footage is followed by the Astra-versus-Sol demonstration and its dated results. - GPT-6.2 Astra (Max) enters at #6 with 1,291 points, while GPT-6.2 Sol (Max) enters at #11 with 1,241. Sol joins the cost frontier at $0.90 median cost per task. These are November 18 overall Agent Arena Index results. - Gemini 4 Neon (High) debuts at #2 overall with 1,336 points on the Agent Arena Index in Arena's November 18 announcement, with a visual demonstration from First Impressions. - Arena's analysis shows Google's Gemini 4 release gaps shrinking from 28 to 15 to 6 days: Argon → Carbon → Hydrogen → Neon. - Qwen4-Max (xHigh) is the highest-ranked Chinese model in the November 15 Agent Arena announcement: #14 overall, +7.5% net improvement across 14.7k sessions, under the pre-Index scoring method. - MAI-Image-2.7 leaves preview, reaching #3 in Multi-Image Edit and the cost frontier at $0.035 per image in the November 17 announcement. Official editing examples and Arena's Pareto chart show the result. - Amazon-Web-Search-Agent with AgentCore Web Search takes #1 in Search Arena in the November 18 announcement. - Muse Ember (Max) by Meta reshapes the Code Arena in the November 13 announcement: WebDev frontier. Arena's November 16 analysis reports a 1,608 score at about $0.03 per answer, versus 1,438 for the previous best at that price.
#arenai #ainewsGemini 4 Neon | First ImpressionsArena2026-11-19 | ttps://arena.ai/agent
How close does Gemini 4 Neon feel to the frontier when you look at what it actually builds? Peter Gostev compares its outputs with Opus, Fable, Astra, Sol and other Google models across 3D scenes, browser games, interface design and data visualization.
He also considers generation costs and where introductory pricing affects the comparison.
His first impression: Neon shows creative flair and some promising dynamic visualizations, but quality varies across prompts, and he still sees a gap to the strongest models.
Rankings and prices shown reflect the time of recording.
00:00 Introduction and Arena ranking 00:29 Istanbul: first comparisons 01:38 White House and 3D detail 06:38 Creativity, realism and complex scenes 11:57 Live Aid and cinematic movement 13:38 Games: controls and visual quality 15:40 UI and design comparisons 17:02 Data visualization 18:03 UI coherence and design quality 21:59 Dynamic data and consistency 23:36 Pricing and model comparisons 27:30 Final verdict: how close to the frontier?
#Gemini4 #GoogleGemini #AI #arenaaiClaude Opus 4.5 to 5.5: Same Prompt, Different ResultsArena2026-11-19 | arena.ai/agent
Watch Claude tackle the same animated cyclist prompt across nine outputs, from Opus 4.5 to Opus 5.5 (Max). See how the artwork and motion change across models released from November 2025 to September 2026.
The challenge: animate a cyclist in pure SVG and CSS, with coordinated pedaling, spinning wheels, and a moving chain. We show each result in sequence, then compare the first and last side by side and bring all nine together.
MODELS AND SETTINGS SHOWN: • Claude Opus 4.5 • Claude Opus 4.5 (Thinking 32K) • Claude Opus 4.6 • Claude Opus 4.6 (Thinking) • Claude Sonnet 4.6 • Claude Opus 4.8 • Claude Opus 4.8 (Thinking) • Claude Fable 5.1 (Max) • Claude Opus 5.5 (Max)
#arenai #anthropicGPT-6.2 Astra vs Sol | First ImpressionsArena2026-11-19 | arena.ai/agent
Is Astra's extra detail worth the cost, or is Sol the better choice for everyday work?
Peter Gostev compares GPT-6.2 Astra and GPT-6.2 Sol across 3D scenes, games, slides, data visualizations, websites, interactive apps and SVG illustrations. See where Astra adds depth and precision, where Sol's motion and design choices stand out, and how the costs shown in the recorded runs compare with the results.
Peter shares which model he prefers for daily use and when he reaches for Astra for more demanding tasks. Which result surprised you most? Leave your prompt ideas in the comments, and subscribe for more hands-on model comparisons.
00:00 Astra vs Sol: what matters 00:30 3D detail and cost 04:20 Motion, crowds and camera work 08:10 Games and gameplay 11:22 Slides and information design 17:00 Data visualization 18:05 HTML visuals and animation 21:22 Sol for daily use and Ultrafast 22:30 Websites and UI design 25:12 Complex apps and interactive visuals 28:05 SVG illustration 29:43 Pushing Astra with a larger city 33:40 Final verdictIntroducing Agent Arena IndexArena2026-11-18 | arena.ai/leaderboardGPT 6.2 vs Claude Fable 5.2: Astra & Sol First ImpressionsArena2026-11-17 | GPT 6.2 Astra and Sol meet Claude Fable 5.2 in a hands-on visual coding comparison.
Peter Gostev compares code-generated 3D worlds, HTML animations, websites, games and data visualizations, with Claude Opus 5.5 appearing in selected comparisons. We look at visual detail, creative range, playable controls and the trade-offs behind the costs shown on screen.
Across these examples, Peter finds strengths in Astra's precision, Fable's creativity and Sol's value, alongside repetitive layouts, control bugs and results that still need work.
NOTE: The costs in the video were based on ‘Astra 6’ prices and the actual costs per generation are 35% lower for Astra 6.2 (vs Astra 6).
00:00 Meet the models and the pricing context 01:08 3D worlds: architecture, water and detail 05:46 Creative range vs. a consistent style 11:48 HTML stories and animations 16:45 Clay, crochet and emerging visual styles 21:04 Website UI: structure and repetitive layouts 27:33 Games: graphics, controls and bugs 34:39 Data visualization 36:37 Explaining concepts with animation 38:41 Pushing SVG detail: the peacock test 40:10 100 ways to draw a dog 41:29 Final thoughts: Astra, Fable and Sol
#GPT62 #Claude #arenaaiCan AI agents push CS research forward?Arena2026-11-17 | Can AI agents rebuild a research paper’s core contribution—and outperform the original implementation? FrontierCS 2 challenges them to beat published methods—without seeing the paper or browsing the web.
Hanchen Li and Qiuyang Mang, both PhD Candidates at UC Berkeley, explain how the benchmark works, what the evaluation results reveal about cost and capability, and how agents exploit weaknesses in research evaluations.
From sparse-attention video generation to reinforcement learning and graph algorithms, this Arena Tech Talk explores the infrastructure and human research judgment needed to evaluate AI research agents reliably.
00:00 Introducing FrontierCS 2 03:23 How benchmark tasks are built and scored 04:43 Case study: sparse-attention video generation 08:09 Results: cost versus performance 09:29 Three challenges in automating research benchmarks 10:59 Reward hacking and evaluation isolation 12:50 RAGEN: bypassing learning with search 15:18 Graph algorithms: exploiting cached results 17:21 Scaling evaluation infrastructure 21:32 Simplifying the multi-sandbox workflow 23:05 Why human research judgment still matters 25:03 Key takeaways
#arenaai #AIResearch #AIAgentsQwen 4 | First ImpressionsArena2026-11-16 | arena.ai/agent
How close does Qwen4-Max get to the frontier? Peter Gostev, Arena's AI Capability Lead, tests it on 3D scenes, games, creative prompts, website design and interactive explanations, comparing outputs with models including Opus, GPT, Kimi and Fable.
Peter looks at visual detail, usability and how well each model handles more demanding prompts. His first impression: Qwen4-Max can produce competent, straightforward websites, but these tests reveal a substantial gap in more ambitious creative work.
NOTE: Qwen4-Max's pricing and Arena ranking were unknown at the time of recording.
00:00 How close is Qwen4-Max to the frontier? 00:29 Golden Gate Bridge: the first 3D test 01:59 Voxel Rome and more 3D comparisons 05:35 Turning paintings into 3D worlds 09:09 Games: Qwen vs. Opus and Fable 10:18 Creative scenes and camera work 12:32 The crochet mushroom test 13:36 Website design: a smaller gap 15:47 Qwen vs. GPT: web design and usability 18:26 More tests: aircraft, hiking and typography 19:33 Interactive explanations: lightning and Enigma 21:27 Phi Phi Islands: Qwen vs. Fable 22:08 Pricing unknowns and final thoughts
A new Agent Arena leader, Gemini Hydrogen’s fast execution, efficient small models, high-resolution image generation, and ERNIE’s arrival.
Follow Arena on X: https://x.com/arena
NEWS: November 6–November 13, 2026 • Claude Fable 5.2 (Max) takes #1 in Agent Arena with +30.4% net improvement across 4.7k real-world sessions. • Gemini 4 Hydrogen (High) launches as the fastest model in the Agent Arena top 10, with a typical step of 7.3 seconds. • Muse Ember (Max) leads small models in Code Arena: WebDev, with 1608 points and a $0.03 median cost per task. • Inkling 2 Small improves Agent Arena net improvement by about 18 percentage points over Inkling-Small. Open weights coming soon. • MiniMax H3-Image arrives in Arena Battle Mode, with scores pending. MiniMax’s model card specifies native generation at approximately 16 million pixels per image (16 megapixels). • ERNIE 5.5 Flash Preview arrives in Arena Battle Mode. Baidu specifies a 1M-token context window; Code Arena scores are pending.
#arenai #ainewsLetting AI Agents Play the Games They Evaluate — Interaction as a JudgeArena2026-11-12 | Today’s AI-generated apps range from complex websites to full games, and frontier models rarely make obvious mistakes in their code anymore. Judging them from code or static screenshots alone has become hard. A human can usually spot a bug just by playing the game or clicking around the site, and computer-use agents can now do the same.
Wayne Chi, a PhD candidate at Carnegie Mellon University and research intern at Arena, presents interaction as a judge. A computer-use agent navigates the generated application, and the resulting interaction traces are used for downstream evaluation, either by an LLM judge or a trained reward model. The talk covers both preference evaluation (using real Code Arena battles) and correctness evaluation (using requirements mined from real user complaints in Agent Arena).
Highlights: → A 9B reward model trained on interaction traces outperforms frontier LLM judges on preference evaluation, and even with no code at all, it performs comparably to a code-only reward model → On broken requirements, interaction as a judge beats code agent as a judge by about 8% → Performance scales with both computer-use method quality and model capability → For inference-time scaling, it outperforms code agent as a judge from the first iteration and overclaims less → Used for RL training, it produces less code bloat and more successful builds, and improves generation quality by close to 7% over the base model on Arena’s internal auto-rubric
00:00 Introduction to interaction as a judge 00:43 Why code and screenshots fall short: the Minecraft clone example 04:40 Using computer-use agents in place of humans 05:19 Agenda 06:05 Two tasks: preferences vs. correctness 07:48 Mining requirements from real user complaints in Agent Arena 10:27 Requirement verification vs. requirement discovery 11:24 Three computer-use methods: iterative, one-shot, heuristic 14:24 Computer-use methods in action 16:04 Experimental setup 17:58 The judge: LLM judge and reward model 21:17 Results: preference evaluation 26:29 Why code-free judging unlocks new data 27:15 Interaction quality matters and scales 29:30 Q&A: games, harnesses, and RL environments 33:58 Results: correctness evaluation 36:43 Inference-time scaling 40:16 Reinforcement learning with interaction as a judge 46:43 Takeaways and future work
#aievaluation #llm #arenaai #gamedevelopmentFable 5.2 is in the ArenaArena2026-11-09 | arena.aiMeta’s Muse Ember is in the ArenaArena2026-11-09 | arena.aiGemini Omni 1.2 Flash Debuts at #1 in Video GenerationArena2026-11-06 | arena.ai/agent
Gemini Omni 1.2 Flash debuts at #1 in both Text-to-Video and Image-to-Video. This week also covers faster image generation, new open weights, and the cost of agent tasks.
Follow Arena on X: https://x.com/arena
NEWS: October 30–November 6, 2026 • Gemini Omni 1.2 Flash leads both video-generation categories in Arena’s November 5 debut announcement: 1,541 points in Text-to-Video and 1,502 in Image-to-Video. Watch the official Google demo. • GPT-Image-2.6 Sunburst and Flare arrive with image generation that OpenAI says is 50% faster. See matched official examples and follow-up edits. • Kimi K3.1 Pro (Max) releases its weights, ranking #1 among open models in Agent Arena with +5.1% net improvement in Arena's November 4 announcement. • A day later, Tencent Hy4 leads Chinese models in Agent Arena with +5.4% net improvement and a $0.57 median task cost. The same November 5 chart shows Kimi at +5.3%; this is a new snapshot, not a comparison against Kimi's earlier +5.1% score. Hy4's weights were still pending, so this does not displace Kimi's #1 open-model result. • Arena’s four-model cost analysis shows GPT-6 Astra’s cost relative to Gemini 4 Carbon rising from 3.4× at the median to 6.8× at the 95th percentile. The chart also includes Claude Opus 5.5 and GPT-6.1 Sol, using Carbon’s introductory pricing.
#arenai #ainewsHow Krea’s Small Team Trains Its Own Image Models | Victor PerezArena2026-11-05 | How does a small team build its own creative AI models? Peter Gostev, AI Capability Lead at Arena, talks with Victor Perez, co-founder and CEO of Krea, about the work behind Krea’s models and its platform for image and video generation.
They go inside data curation, creative direction, reinforcement learning and LoRAs for consistent people, products and styles. Victor also explores the trade-offs between quality, speed and cost, the challenge of teaching AI taste, and Krea’s plans for specialist models and an open-source release.
00:00 Meet Victor Perez 00:45 How Krea’s small team builds image models 04:50 Compute, taste and creative direction 07:44 Multi-image editing and training data 10:12 LLMs as judges and reinforcement learning 15:43 LoRAs for consistent people, products and styles 22:14 Training LoRAs at the right checkpoint 27:31 What changed in AI image editing 31:33 Balancing quality, speed and cost 35:31 Where faster image models come from 38:31 Control as the next frontier 41:03 Style diversity and specialist models 45:19 Why Krea builds models and a platform 47:56 Better interfaces for creative agents 49:43 Can AI understand media and aesthetics? 52:44 Taste, culture and the search for novelty 59:45 Krea’s roadmap: expert models and open source
#arenaai #aiimages #machinelearningIs AI Building AI Yet? Nathan Benaich on the State of AI Report and Recursive Self-ImprovementArena2026-11-04 | Nathan Benaich — founder and general partner of Air Street Capital and author of the State of AI Report — breaks down this year's ~250-slide report: how close we really are to AI building AI, where AI is (and isn't) changing biology and robotics, and why the gap between AI-native companies and everyone else keeps widening.
We cover early evidence of recursive self-improvement, including an Anthropic experiment where human evaluators preferred the AI-driven research fork over the human one, the first AI-discovered drugs moving through clinical trials, why serious AI-for-biology efforts need their own wet labs, how a single robotics model can now deploy across every customer, and Wayve's bet on end-to-end driving.
On the business side, Nathan shares data showing AI-native companies growing roughly 3x faster than AI-enabled ones, and a Ramp dataset where the top 1% of firms spend $7,000 per employee per month on AI versus a median of $12.50.
0:00 How the State of AI Report started 3:40 This year's biggest surprises 5:22 Recursive self-improvement: how real is it? 8:31 What evidence would prove AI can do research? 11:15 AI in maths vs. open-ended science 13:01 Progress in AI for biology and drug discovery 19:31 How LLM ideas spill over into biology 22:57 Why AI labs need their own wet labs 26:16 Culture and philosophy in pharma 29:19 Robotics: how real is the progress? 32:03 Generalist models vs. real-world deployments 33:41 How transferable are robotics models? 36:10 Wayve vs. Waymo: different bets on self-driving 38:46 The business of AI: revenue growth data 41:53 The widening AI adoption gap 42:44 Codex usage and who delegates real work 46:14 Could persistent agents drive adoption? 47:00 Grading last year's predictions 50:23 Predictions for next year: cyber, model weights, continual learningGemini 4 Carbon Takes #1 | GPT-6.1 Terra Joins the FrontierArena2026-10-30 | Gemini 4 Carbon leads Text Arena, Kimi K3.1 builds in 3D, and GPT-6.1 Terra arrives on the Agent Arena Pareto frontier.
Follow Arena on X: https://x.com/arena
NEWS: October 23–29, 2026 - Gemini 4 Carbon (High): #1 Text Arena, #2 Agent Arena, and #4 Code Arena: WebDev. All three results are preliminary. - Kimi K3.1 Pro: Babylon and a giant Pacific octopus from Arena's 3D showcase. At its checked ranking, Kimi K3.1 Pro (Max) would rank #1 among open models in Agent Arena once its weights are released. The result is preliminary and weights are pending release. - GPT-6.1 Terra (Max): #8 in Agent Arena with +10.6% net improvement at $0.62 median cost per task across nearly 9,000 sessions, in Arena's October 29 announcement. See its arrival extend the frontier in our animated excerpt. - MAI-Image-2.7-Preview: #3 in Single Image Edit at its October 27 debut, with 1,470 points — 45 above MAI-Image-2.6 in that announcement snapshot. - Arena connectors: bring services such as Gmail, Google Drive, Google Calendar, and Linear into Agent Mode. Free for signed-in users. - Happy Halloween
#arenaai #ainewsThe Unit Economics of AI: What It Really Takes to Deploy LLMs at Scale | Ido Rosen of Amazon SDOArena2026-10-29 | What does it actually take to deploy AI in the real world — not on benchmarks, but in production, at scale, on a budget?
Peter Gostev sits down with Ido Rosen from Amazon to talk about the practical side of AI: why unit economics should come first, why "LLM as judge" can mislead you, what a decade in quant trading teaches you about evaluation, and what happens to frontier-lab pricing when open-weights models are only months behind.
0:00 Ido's day-to-day at Amazon SDO 1:09 Agents everywhere, or one feature at a time? 2:50 Why we rarely wait on LLMs (yet) 4:57 Is there a product or business-model gap? 6:53 Unit economics: the head, torso & tail of a catalog 8:21 Are small models getting better? 10:06 What actually works: post-training, fine-tuning & hacks 12:08 Team skills and "simplify and invent" 14:03 Detour: evals and why LLM-as-judge can mislead 16:44 Lessons from a decade in quant trading 17:43 Arena, auctions & choosing the best model 20:43 The skills gap in deploying LLMs 21:34 Open source vs frontier models 22:04 Inference premiums & the labs' free cash flow 24:40 What if open weights stay 6 months behind? 26:31 Does intelligence plateau? A lesson from nature 28:48 Pareto frontiers and why labs will specialize 31:17 A "spiky" world of superhuman niches 33:49 Asymmetric returns: radiology to recommendations 35:39 How big can the AI market really get? 36:50 Wrap-up
Google's new model, Gemini 4 Carbon, put through a long run of 3D generations, creative prompts, visual explainers and UI tasks, compared against Gemini 4 Argon, Claude Opus 5.5, GPT-6.1 Sol, Muse Wonderful, Kimi and DeepSeek.
- Carbon is a clear step up from Gemini 4 Argon, with more detail and a nicer look - Claude Opus 5.5 puts in more detail and flourish, but in the first test it cost $27.50 vs $2.70 for Carbon - Carbon is more creative and dynamic than GPT-6.1 Sol, but Sol is more complete and precise - Carbon beats Muse Wonderful and Kimi in these tests - Verdict: a good improvement, and solidly in the top three
0:00 Intro: Gemini 4 Carbon 0:15 First test (White House): Carbon vs Argon 1:28 Carbon vs Claude Opus 5.5 2:42 Carbon vs GPT-6.1 Sol 3:06 Carbon vs Muse Wonderful and Kimi 3:48 Why test with 3D generations 4:45 Playable 3D game: Carbon, Argon, Sol, Opus 6:28 Where Carbon falls short: gaps in detail 7:50 Stonehenge 8:30 Golden Gate Bridge 10:41 Golden Gate Bridge: GPT-6.1 Sol comparison 11:53 Carbon's creativity 15:33 Monet paintings transition 18:37 Van Gogh paintings world 21:47 Why creativity matters in a model 23:30 Another 3D prompt: Carbon vs Opus, Sol, Muse, DeepSeek 26:27 London through the ages 29:56 Istanbul 31:44 Harbin ice city 33:20 New test: visual explainers (how lightning works) 35:34 John Snow and the cholera map 38:20 UI work 39:45 Golden Gate Bridge completion animation 40:43 More UI generations 41:55 Ski resort site and ramen ordering 44:20 Verdict
#Gemini #AI #LLM #arenaaiSelf-Improving AI Agents: Exo and the Bitter Lesson | Alex KrentselArena2026-10-28 | Alex Krentsel (Senior Researcher at Google and PhD candidate at UC Berkeley) presents Exo, an architecture for agentic systems that can safely edit their own harness at runtime. Borrowing the exokernel idea from operating systems, Exo splits an agent into a protected mechanism layer (state, history, secrets, a "guardian") and a fully evolvable policy layer.
Alex walks through the design, why safety is the hard part of self-improvement, and results on benchmarks and open-world RuneScape tasks, where Exo was 31-62% cheaper than Codex and Pi.
Joint work with Sylvia Ratnasamy, Ion Stoica, Martín Casado, and Ankur Goyal.
0:00 What makes a harness "bitter lesson aligned"? 0:31 Introducing Exo and how the project started 1:24 Background: from next-token predictors to autonomous agents 5:17 The emerging self-improving layer 6:14 What is an agentic system? Defining harness and policy 8:22 Q&A: domain-specific vs. general agents, thicker vs. thinner harnesses 11:13 Why finding a good policy is hard: agent optimizers and their limits 14:30 Key idea: let the harness edit itself at runtime 17:37 The safety problem: fate sharing and unprotected state 19:02 Exo's design: protected mechanism layer + evolvable policy layer 20:13 Inspiration from operating systems: the exokernel 22:21 What lives in each layer 24:35 Q&A: will the split change as models get more trustworthy? 28:35 An agent as policy + state + sandbox 29:34 How self-improvement works: the guardian 31:03 Q&A: protecting the guardian, comparison with other harness designs 32:50 Secrets handling and self-editing its own improvement process 34:19 Implementation and open source 35:10 Evaluation setup: closed benchmarks and open-world tasks 36:35 Q&A: how does the policy actually evolve? 37:43 RuneScape results: Exo vs. Codex and Pi 41:34 Q&A: cost-awareness and the meta prompt 42:35 Second run: carrying learnings to a new account 44:07 Limitations: where self-improvement doesn't pay off 45:09 Q&A: "dreaming," reward sparsity, and benchmark gaps 49:24 Personalized harnesses and policy "branches" 50:12 Designing systems aligned with the bitter lesson
With the release of Connectors in Agent Mode, you can now bring your everyday tools — including Notion, Linear, Gmail, Google Drive, Google Calendar, GitHub and many more — directly into Agent Mode so the agent can read and execute actions across services in a single task.
With Connectors on Arena.ai, users can take on even more complex, real-world tasks from start to finish.
#arenaai #agenticaiKimi K3.1 Pro is in the Arena | Arena.aiArena2026-10-24 | arena.ai/agent
Moonshot’s Kimi K3.1 Pro has officially made its debut in the Arena.
#arenaai #llm #agenticai #moonshotaiKimi K3.1 Pro First Impressions: Has Moonshot Caught Up to the Frontier?Arena2026-10-23 | arena.ai/agent
Peter Gostev puts Moonshot’s Kimi K3.1 Pro, the follow-up to the open-weight Kimi K3, through 3D scenes, HTML animations, website and UI builds, and research-driven pages. He compares it with Kimi K3, Opus 5.5, Sonnet 5.5, GPT 6.1 Sol, GPT 5.6 Sol, Gemini 4 Argon, DeepSeek 4.1 Flash, and the Muse Wonder models.
Peter’s take: Kimi K3.1 Pro is a huge jump over K3, but in these tests it’s not quite at the frontier, and cost is a big part of the story.
Note on pricing: K3.1 pricing wasn’t available when this was recorded, so the costs shown use Kimi K3 prices. The actual cost is approximately a third lower.
0:00 3D scene: dinosaurs 3:12 3D scene: bears and salmon 7:32 3D scene: Acropolis 9:38 3D scene: Roman town center (voxel) 12:21 3D scene: Tower of Babel 14:55 HTML story: first example 16:43 HTML story: Taj Mahal 17:45 HTML animation: Trojan horse on a vase 19:24 HTML scene: Venice 20:34 Animation: train delivering mail 22:46 Website: ramen shop 27:07 Website: toy store 29:26 Data layout: Grand Prix 32:22 Research UI: alpine walk 35:06 Research UI: data center map 38:19 Research UI: shopping comparison 39:19 Closing thoughts on competition and cost 41:19 Claymation example
#arenaai #moonshotai #opensource #llmDeepSeek Builds 3D Worlds + Muse Wonder Full Hits Top 10Arena2026-10-23 | DeepSeek builds immersive 3D worlds, Muse Wonder Full lands in the top 10, and the frontier keeps moving.
Follow Arena on X: https://x.com/arena
NEWS: October 16–23, 2026: - DeepSeek-V4.1-Flash-20261017 (Max): 3D worlds and the #1 open-model position in both Agent Arena and Code Arena: WebDev. - Muse Wonder Full (Max): #6 in Code Arena: WebDev with 1,726 points, and #7 in Agent Arena. - Krea 3 Ultra: #4 in Multi-Image Edit with 1,435 points. - GPT-Image-2.5 Sunburst: available in Direct Mode through November 2 at 8 a.m. PT. - MiniMax-M3.1-Flash: +106 WebDev points and 50% lower token pricing compared with MiniMax-M3.
#arenaai #ainewsWhat Building an AI Model Actually Looks Like | Pradeep Natarajan of Amazon AGIArena2026-10-21 | Can AI be both accurate and creative? Pradeep Natarajan, Senior Principal Scientist at Amazon AGI joins Peter Gostev on Arena Conversations to talk about what building an AI model actually looks like day to day: the data work, the training recipe, the evals, and why the metrics we choose shape what models become.
- Why Pradeep estimates 90-95% of gains come from getting the fundamentals right, not from a single algorithmic breakthrough - How evals, not just compute, become the real differentiator - Why harness optimization is as important as the model itself - How the metrics we define (like accuracy and style) steer models away from creativity
0:00 What model building looks like day to day 3:33 Compute vs. data as the bottleneck 5:15 The core training recipe: pre-training, SFT, and RLHF 7:15 Big breakthroughs vs. accumulated small insights (PPO to GRPO) 8:28 Could you build a frontier model just from the papers? 11:00 Do models really generalize? 12:47 Why evals are the real moat 16:12 The unsolved problem: evaluating models in real deployments 18:10 Why harnesses matter as much as the model 21:20 How every AI leap opens up new fields 22:25 Recursive self-improvement and human-machine co-intelligence 26:12 From magical to mundane: how fast expectations reset 28:46 Can models do real research? 32:26 Why AI writing feels like it's getting worse 32:59 Metrics vs. creativity: the "Mount Everest" problem 35:36 Will AI writing get better?
#arenaai #amazonai #llmDeepSeek-V4.1-Flash is in the Arena | Arena.aiArena2026-10-20 | arena.ai/agent
DeepSeek-V4.1-Flash-20261017 (Max) showcases improved capabilities in 3D world generation, landing at #1 among open models in Agent Arena.
#arenaai #llm #agenticai #deepseekMuse Wonder FULL from Meta | First ImpressionsArena2026-10-19 | arena.ai/agent
Meta just released Muse Wonder and Muse Wonder Full, and in this video we test both on 3D scenes, animated HTML clips and several website builds. We compare them against Claude Opus 5.5, Sonnet 5.5 and Fable 5.1, GPT-6.1 Sol, GPT-6 Astra and Gemini 4 Argon.
The Full model is a big step up from the regular Wonder, but it still isn’t at the frontier on design and front-end work. Try the models yourself on Arena.
0:00 Why Meta’s release was a mistake 2:35 Where Wonder Full lands vs. GPT, Astra, Opus and Fable 3:50 3D test: the White House 5:23 Opus 5.5 builds an entire world 7:59 3D test: island, water and boats 10:01 HTML animation clips (Great Wall, Taj Mahal, Trojan vase) 14:15 Website test: ramen shop 16:38 What separates good models from great ones (understanding the user) 18:08 Website test: toy site and agentic image generation 20:04 Grand Prix race dashboard 24:33 Website test: chocolatier 27:09 Website test: ski resort 28:21 Verdict: Wonder Full should have shipped to everyone 30:24 Open weights and cost caveats 31:21 Try the models on Arena
#arenaai #metamuse #llmMuse Wonder #2 in Text, Muse Spark Open & MiMo V2.6 Pro Tops Open ModelsArena2026-10-16 | Meta’s Muse Wonder debuts at #2 in Text Arena, while new model arrivals, smarter routing, and research track a fast-moving frontier.
Follow Arena on X: https://x.com/arena
NEWS: October 9–16, 2026 👉 Muse Wonder (Max): Text Arena #2 / 1513 points, Code Arena: WebDev #15 / 1651 points, and Agent Arena #12 / +6.4% net improvement in its October 13–14 announcement snapshots. 👉 Muse Spark Open: available in Agent Arena; scores coming soon at the October 15 cutoff. 👉 Max router: frontier performance at 37% lower median cost per task than Claude Fable 5.1 (Max) during Arena’s reported test period. 👉 MiMo V2.6 Pro: #1 among open models in Agent Arena, with +3.8% net improvement after the Bash Reliability update. 👉 Gemini-4 Argon (High): #5 overall in Agent Arena, with +13.2% net improvement in the October 13 snapshot. 👉 Confirmed Success: Arena reports the top five agents reached 94% of tasks confirmed done by users, up from 80% in July. 👉 July’s best is now average: the top-20 average matched Claude Fable 5’s July 1 Confirmed Success baseline in 12 weeks.
#arenaai #ainewsMeta Muse Wonder, Mistral Le Chonk, and What Chinese Labs Do NextArena2026-10-15 | Meta says Muse Wonder is on par with Opus 5.5. Dawid doesn’t buy it. In this rapid fire episode, we cover Meta’s strategy with Muse, Mistral’s new Large 4 “Le Chonk,” and the two trends to watch next in AI.
0:00 Is Muse Wonder on par with Opus 5.5? 0:57 Mistral Large 4 “Le Chonk” and open source in Europe 1:55 What to watch next 2:23 Will Chinese labs release soon? 2:53 The race on cache read pricesDo AI Detectors Actually Work? Arena Conversations With Pangram CEO Max SperoArena2026-10-14 | Can you tell whether a piece of writing came from a human, ChatGPT, Claude or Gemini? Peter Gostev talks with Max Spero, CEO of Pangram Labs and one of TIME’s 100 most influential people in AI, about how AI text detection works—and why it takes more than spotting em dashes.
We go inside Pangram’s training process: synthetic mirrors, model-family fingerprints, AI-assisted writing, image detection and robustness against AI humanizers. Max also explains how the team chooses base models, finds difficult training examples and builds the data and infrastructure to keep up with new AI models.
00:00 Meet Max Spero and why he built Pangram 01:11 Pangram’s approach: deep learning vs perplexity 02:52 Training data and synthetic mirrors 05:27 Identifying ChatGPT, Claude and Gemini 09:05 New models and signals beyond em dashes 12:15 Edited and AI-assisted writing 14:51 AI images, video and deepfakes 17:37 How Pangram trains its image detector 22:33 Model size, scaling and difficult examples 26:48 Humanizers and adversarial robustness 28:22 Choosing the right base model 33:53 Data quality, training pipelines and retraining 38:40 The promise and limits of small classifiers 43:11 What model fingerprints reveal in the wild 47:21 Hiring engineers and researchers at Pangram
#arenaai #aidetector #llmMuse Wonder from Meta | First ImpressionsArena2026-10-13 | arena.ai/agent
Meta says its new Muse Wonder Max is at the level of Opus 5.5 and Astra. We put it through dozens of one-shot prompts (3D scenes, animated HTML pieces and front-end design) and compared it with Astra, Opus 5.5, GPT-6.1 Sol, Claude Sonnet 5.5, Fable, Kimi and Mistral, including what each generation cost.
The short version: it’s better than Kimi on most tasks, but across one-shot generations it looks closer to GPT-5.6 Sol than to the frontier models Meta compared it with.
0:00 Meta’s Muse Wonder Max and what we’re testing 0:44 3D test: octopus 3:59 3D test: White House 6:07 3D test: dinosaurs 7:47 HTML “video” generations (20-second animated clips) 11:16 Front-end test: race / Grand Prix website 15:07 Front-end test: optical illusions website 16:54 Front-end test: aquarium website 18:23 Front-end test: apartment comparison website 20:34 Verdict: where Muse Wonder Max lands 22:53 Arena’s agent leaderboard 23:59 Muse Spark and what to watch next
New image models, stronger coding tools, and research testing what AI can really do.
Follow Arena on X: https://x.com/arena
NEWS: October 2–9, 2026 - Nano Banana 2.1 (gemini-nano-banana-2.1): #4 in Multi-Image Edit, #6 in Text-to-Image, and #6 in Image Edit - New open-source models in Code Arena: MIT-licensed MiMo-V2.6-Flash, MiMo-V2.6-Pro, Mistral Large 4 - Arena’s post-training research: FLUX.2-dev gains 69 Arena points with human-preference and rubric rewards - Introducing Arena’s Alignment Index, with new signals for real-world AI safety: 48% of code-debugging sessions in the launch study were flagged for deceptive completion - Arena announces a $200M Series B at a $3.1B valuation
#arenaai #ainewsArenas Series B: $200M at $3.1B, and a New Way to Measure AI AlignmentArena2026-10-08 | arena.ai/agent
Arena has announced a $200M Series B at a $3.1B valuation, and is releasing the Alignment Index, an independent measure of AI-human alignment built from traces of real-world tasks.
0:00 Series B: $200M at a $3.1B valuation 0:33 Arena's growth: sessions, votes, and Agent Arena 0:58 Introducing the Alignment Index 1:32 The three safety signalsMistral Large 4 (Le Chonk) | First ImpressionsArena2026-10-07 | Try the models yourself: arena.ai/agent
Mistral Large 4 is here: the first French model in a while, a 1T-parameter open-source model (49B active) with weights due at the end of the month. We put it through a series of real-world tests and compare it with leading proprietary and open models, including GPT-6.1 Sol, Astra, Claude Fable, Claude Opus 5.5, Claude Sonnet 5.5, Kimi K3, GLM 5.3 Flash, DeepSeek Flash and Muse Spark, so you can calibrate where it actually stands.
0:00 Mistral Large 4 is here 0:21 Specs: 1T total / 49B active parameters, and training compute 1:20 Pricing vs. open models like Kimi K3 and GLM 5.3 Flash 3:12 Where it ranks on WebDev Arena 3:48 Website design test: watchmaker site 6:50 Website design test: homes to explore 10:24 Website design test: optical art exhibition 15:13 Does open source mean cheaper? 16:48 Mona Lisa animation and style tests 18:50 White-collar work: Boeing 787 supply chain 22:34 Cost vs. performance: DeepSeek Flash, Muse Spark and the Pareto frontier 25:02 3D generation tests 27:40 Verdict: why it matters for Europe
#arenaai #opensource #llmSonnet 5.5 vs GPT-6.1 Sol: The $8 FrontierArena2026-10-02 | arena.ai
This Week in the Arena: Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 4 Argon, and the latest AI rankings.
Follow Arena on X: https://x.com/arena
NEWS: September 25–October 2, 2026
• Claude Sonnet 5.5 vs GPT-6.1 Sol: original Tower of Babel demos from Arena’s first impressions. • Claude Opus 5.5 (High): #2 in Agent Arena, +12.15% net improvement at $1.31 median cost per task; 40–56% lower cost than Opus 5 (High/Max). • Gemini 4 Argon (High): #1 in Text Arena with 1525 points at its September 30 announcement. • MiMo-V2.6-Pro: #5 among open-source models in Agent Arena, with +3.17% net improvement across 8.1K+ sessions. • RIP em dashes: Claude Opus 5.5 uses 95% fewer em dashes than Opus 5 in Arena’s high-reasoning Text Arena sample (0.8 vs 15.2 per 1,000 words). • One prompt, 19 takes on Ancient Rome: Arena’s model progression, featuring Claude Opus 5.5 and Claude Sonnet 5.5 demos. • AI judge self-bias: across 12 models, judges chose their own answers 58% of the time on average; people chose those same answers 34% of the time. • Steerability: Arena’s updated methodology now credits getting it right the first time and evaluates every judgeable assistant turn.
#ArenaAI #AI #TechNewsGemini 4 Hits #1 on Text Arena, But Is It Benchmaxxing? Or is Google Back?Arena2026-10-01 | arena.ai/agent
Google has released Gemini 4, its first big new model in a while, landing at #1 on Text Arena and on the Pareto frontier.
Peter responds to accusations that Google is benchmaxxing, explains how the staged rollout appears to be working, and discusses why focusing on Gmail, Sheets, Docs and Search may matter more to Google than leading the coding race.
0:00 Is Google back? 0:48 Is Google benchmaxxing? 1:42 #1 on Text Arena 2:11 Who should be using Gemini 4? 3:05 Why it's not fully available yet 4:12 Is this a requirement for all US labs? 5:05 Google's product strategy vs. the coding race 6:59 What Gemini 4.1 should fix
#arenaai #llm #agenticai #googlegeminiGemini 4 Argon is in the Arena | Arena.aiArena2026-10-01 | arena.ai/agent
Gemini 4 Argon has officially made its debut in the Arena.
#arenaai #llm #agenticai #googlegeminiGemini 4 Argon | First ImpressionsArena2026-09-30 | Try the models yourself: arena.ai/agent
Google just dropped Gemini 4 Argon — the first Gemini version-number change in a while. Peter runs it head-to-head against other frontier models on a handful of generations for a quick vibe check.
The headline: Gemini 4 Argon makes the Pareto frontier at roughly a third cheaper than GPT-6.1 Sol and around 70% cheaper than Opus 5.5. Gemini 4 makes some strong generations, but results were inconsistent across prompts showing that stability is something to watch.
0:00 Gemini 4 Argon is here 0:12 Agent Arena ranking & the Pareto frontier 0:57 Example generations: Golden Gate Bridge 2:21 Sanctuary islands 3:15 Game generation 4:19 Tower of Babel 5:27 White House 6:53 Why consistency matters more than any single generation 7:59 Back to the leaderboard: cost vs. performance
#arenaai #llm #gemini4Sonnet 5.5 vs GPT-6.1 Sol | First ImpressionsArena2026-09-30 | Try the models yourself: arena.ai/agent
As the first efficiency-optimized models to follow Fable and Astra and Fable, can Claude Sonnet 5.5 and GPT-6.1 Sol still deliver strong results at a fraction of the price?
We run them side by side on multiple prompts, tracking what each generation actually costs, compared with Opus, Astra, and Fable. Sol follows instructions more literally, while Sonnet interprets prompts more broadly and brings more dynamism and storytelling, especially in the HTML animations.
The takeaway: Sonnet 5.5 or GPT-6.1 Sol can be a sensible daily driver, and you only need to move up when something isn't working. Also, skip max reasoning on Anthropic models.
00:00 Intro: GPT-6.1 Sol vs Sonnet 5.5 00:48 Tower of Babel 3D test and costs 04:19 Why Max for Sol and High for Sonnet 05:47 Painting transitions: literal vs. interpretive 07:46 City flyover test 09:33 Pricing vs. Astra and Fable 12:37 Why smaller models are catching up 14:14 Webpage and data-viz UI tests 16:55 Reasoning-level advice: Anthropic vs. OpenAI 19:22 HTML storytelling animations: Trojan horse 21:31 Sisyphus 23:18 First manned balloon flight over Paris 24:56 Educational uses: Golden Gate Bridge 26:29 The tortoise and the hare 28:42 Conclusion: good, cheap daily drivers
#arenaai #llm #anthropic #openaiRebuilding Ancient Rome — 10 Months of AI ProgressArena2026-09-29 | arena.ai/agent
Watch the progress of frontier models in bringing Ancient Rome to life on Arena, from Q4 2025 to today with Claude Sonnet 5.5.
Featured models: - Claude Opus 4.5, 4.6, 4.7, 4.8, 5 and 5.5 - Claude Fable 5 and 5.1 - Claude Sonnet 5.5 - GPT-5.2, 5.3 Codex, 5.4, 5.5, 5.6 Sol, 6 Astra, and 6 Sol - Gemini 3.1 - Kimi K3 - Qwen 3.8 Max
See example prompt: https://x.com/arena/status/2104992778371625189
#arenaai #llm #agenticaiClaude Code vs Codex vs Pi: Which Coding Agent Wins?Arena2026-09-28 | arena.ai/agent
Could your coding agent cost twice as much for almost the same results?
Melissa Pan, a PhD candidate at UC Berkeley's Sky Computing Lab, explores the hidden “harness tax”: how the system around an AI model affects its cost and performance. Comparing Claude Code, Codex, and Pi, she shares three findings: harness choice affected cost more than accuracy in these benchmarks, a simple four-tool harness remained competitive, and models performed well outside their own provider’s tools.
She walks through the results and explains what it means for building useful coding agents on a realistic budget.
00:00 The hidden harness tax 00:41 Finding 1: Cost changes more than accuracy 00:54 Comparing cost and task success 01:33 Pi vs. Codex vs. Claude Code 02:06 Finding 2: Four tools can be enough 02:51 Finding 3: Models work across harnesses 03:16 Why harness research still matters 03:45 Making your AI budget go further
#CodingAgents #AgenticAI #AIResearchGPT-6 Sol vs. Opus 5.5: New Models, New RankingsArena2026-09-25 | arena.ai
This Week in the Arena: GPT-6 Sol, Claude Opus 5.5, and the latest AI rankings—all in under a minute.
Follow Arena on X: https://x.com/arena
NEWS: September 21–24, 2026
- GPT-6 Sol vs. Claude Opus 5.5 - Opus 5.5 (Max): #1 in Code Arena: WebDev with 1818 points - GPT-6 Luna (Max): 1593 points, up 74 over GPT-5.6 Luna (xHigh) - Qwen Image 2.1: #1 among open models in Image Edit and Text-to-Image - Grok 4.7 and MiMo V2.6 Pro: fresh Code Arena results - Praise and complaints: what real user feedback reveals about coding agents - Academic Partnerships Program: research funding up to $50,000, with applications due October 30 - Redesigned Arena leaderboard: a new hub for model releases, rankings, capabilities, and news
#ArenaAI #AI #TechNewsClaude Opus 5.5 | First impressions and comparison with GPT-6 SolArena2026-09-24 | arena.ai/agent
After hours of testing, and hundreds of dollars in tokens, Peter doesn’t think comparing Claude Opus 5.5 to GPT-6 Sol even makes sense anymore. They're just fundamentally different models playing different roles in your stack.
Peter also shares his theory on what this says about Anthropic's RL strategy — and why Opus 5.5 might be making Anthropic's own flagship, Fable, hard to justify.
0:00 – Intro: Opus 5.5 0:16 – Why this comparison doesn't really make sense 1:00 – Golden Gate Bridge: Opus vs Sol side by side 2:02 – The shocking numbers: tokens, cost & lines of code 3:29 – Stonehenge: 7 hours and $61 later 4:35 – The $90 castle vs the $1.50 castle 5:14 – The octopus that looks like real footage 6:04 – More generations (6 hours for a bear?!) 7:41 – Is Opus actually competing with Fable? 8:12 – London over 2,000 years: Londinium to The Shard 10:48 – Tower of Babel + a theory on Anthropic's RL advantage 14:20 – Manhattan & Istanbul: which model for which use case 15:53 – HTML animated shorts: Claude's secret superpower 16:18 – Truman Show: the edge of the world 17:34 – The painted Rickroll 18:23 – Freddie Mercury at Wembley 19:05 – Distracted boyfriend & Usain Bolt 19:38 – Icarus: like an animated children's book 20:22 – Fall of the Berlin Wall 20:46 – The Trojan Horse vase (with the full prompt) 23:19 – Final verdict: two totally different categories
#ClaudeOpus55 #GPT6 #Anthropic #AI #AICoding #arenaaiClaude Opus 5.5 is in the Arena | Arena.aiArena2026-09-24 | arena.ai/agent
Claude Opus 5.5 has officially made its debut in the Arena.
#arenaai #llm #agenticai #anthropicGPT-6 Sol | First impressionsArena2026-09-22 | arena.ai/agent
GPT-6 Sol just launched — and the headline isn't raw quality, it's efficiency. In this video, Peter runs GPT-6 Sol head-to-head against GPT-5.6 Sol on the same prompts (both on max reasoning) and compare not just the generations, but total tokens used and wall-clock time.
The results: comparable or better outputs in a fraction of the time, often at half the token cost. Peter suggests this points to a bigger trend of smarter, cheaper, faster models — and explains why you can stop cranking reasoning to max.
0:00 Intro: GPT-6 Sol launch & making sense of the model names 0:36 First head-to-head: same prompt, tokens & generation time compared 2:29 3D balloon test: 53 min / 30M tokens vs 10 min / 4M tokens 3:26 Stonehenge comparison 4:01 Octopus test & my theory on why validation burns time, not tokens 5:18 Why cost matters: agent-mode rankings and subscription burn 8:02 Monet painting world: creativity test 9:32 Van Gogh generation & a curious token discrepancy 10:01 The trend: smarter models that don't validate in circles 11:59 Raindrop macro world: the quality uplift 12:54 How GPT-6 Sol feels vs the brittle 5.6 14:21 Advice: try lower reasoning settings (max is rarely worth it) 16:53 Humpback whale finale & closing thoughts
#GPT6 #AI #LLM #arenaaiGame development might be the toughest frontier for AI coding agentsArena2026-09-17 | Try top models for free on Arena:
Read the full article: https://x.com/arena/status/2098471986548625487One year of coding models in 30 secondsArena2026-09-15 | Try top models for free on Arena: arena.aiGPT-Image-2.5 (sunburst) vs Nano-BananaArena2026-09-10 | Try the latest models for free: arena.aiTop models return to ArenaArena2026-09-09 | Try frontier models for free: arena.ai