LMArena · Week of Nov 9–15, 2026
This week so far
Arena data: coding models now beat April's leader at about 1/20th the cost
Arena published an analysis on Nov 13 of every model listed on its Code Arena WebDev board from December 2025 through Oct 23, 2026, covering 105 models from 24 labs across 44 weekly leaderboards. It says top scores are now climbing about three times faster, and that matching April's best coding model cost roughly 20 times less by October, using per-answer cost data from April onward. The posts were Arena's best-performing of the day on X and were reshared in several languages, mostly as neutral summaries. There was no visible pushback on the method.
Arena weekly recap: Claude Fable 5.2 leads Agent Arena as Gemini 4 Hydrogen adds speed
Arena's Nov 13 weekly recap says Anthropic's Claude Fable 5.2 (Max) took first place in Agent Arena, citing a net improvement of about 30.5% across roughly 7.8k real-world sessions. The recap also highlights Google DeepMind's Gemini 4 Hydrogen, which Arena added to Agent Arena and Code Arena WebDev on Nov 12, as a speed-focused option. The changelog shows the same day's additions of GPT-6.1 Sol, GPT-6.1 Terra, Claude Sonnet 5.5 and Claude Opus 5.5 to Agent Arena. Engagement was moderate and neutral.
Arena ranks Gemini 4 Hydrogen in Agent and Code Arena, Inkling 2 Small tops US open models
Arena posted day-one placements for Google DeepMind's Gemini 4 Hydrogen (High): #5 overall in Agent Arena across 6,000+ real sessions, #4 on knowledge-work tasks, and the fastest model in the Agent Arena top 10. In Code Arena WebDev it debuted at #10 with 1703 points, which Arena framed as a cost-efficient option. Arena also said Thinking Machines' newly open-weighted Inkling 2 Small is the top US open model in Code Arena WebDev at 1647 points and in Agent Arena. Replies on X were mostly neutral, with some users questioning Hydrogen's lower Text Arena score.
Arena publishes Interaction-as-a-Judge research and opens two webdev datasets
Arena released research on using computer-use agents that click through generated apps and games as judges, instead of grading code or screenshots alone. In its write-up, interaction-based judging caught more broken requirements than code-only judging (63.0% to 71.1%), lifted agent self-repair from 55.4% to 75.5% of requirements fixed, and an interaction-only reward used for RL on a 9B Qwen model gained the equivalent of 50+ Arena points. Arena is also open-sourcing ArenaWebdevPreferences and ArenaWebdevCorrectness, built from real Code Arena prompts, agent traces and human labels. The work was done with Carnegie Mellon collaborators and announced in a thread on X, where reaction was modest but positive.
Arena adds Claude Fable 5.2 and Muse Ember to Agent Arena, opens ERNIE 5.5 Flash battles
Arena's leaderboard changelog for Nov 11 lists Claude Fable 5.2 (Max) and Muse Ember (Max) as new Agent Arena entries. On X, Arena said Fable 5.2 took first place with about a 30.4% net improvement over roughly 4.7k community agent sessions, around 1.6x the runner-up Opus 5.5, while Meta's Muse Ember was framed as the cost-efficiency pick at about two cents per task. The same day Arena put Baidu's ERNIE 5.5 Flash Preview into Battle Mode and later said it was trending toward second among open models on Code Arena: WebDev. Reaction on X and Bluesky was mostly people citing the new rankings and price gaps, plus usual caveats about reading leaderboard scores with care.
Arena adds U2 Flash to WebDev board and opens MiniMax H3-Image battles
Arena's leaderboard changelog for Nov 10 lists U2 Flash as a new entry on the Code Arena WebDev leaderboard. Separately, Arena put MiniMax's H3-Image model into Battle Mode for anonymous voting, saying scores will follow once enough votes come in. The image-model announcement on X drew around a hundred likes and reposts. These continue Arena's near-daily cadence of adding new frontier and open models to its public rankings.
Arena posts same-day Code Arena scores for Fable 5.2 and Muse Ember, adds Inkling 2 Small
Arena's leaderboard changelog for Nov 9 lists four additions: Claude Fable 5.2 (xHigh) and Muse Ember (Max) on Code Arena WebDev, Inkling 2 Small on Agent Arena, and Kimi K3.1 Pro on Text Arena. The WebDev board shows Fable 5.2 at 1832 under an AutoEval tag with no votes yet, above Opus 5.5 (Max) at 1818. Muse Ember is listed as a preliminary entry at 1608 with $0.10/$0.50 pricing. Arena's own X posts calling Fable 5.2 the new WebDev #1 and Ember the top small model were its most-engaged posts of the day, and prediction-market accounts on X tracked the Text Arena standings.
No Director, VP, or CTO postings this week.