Arena raises $200 million and starts scoring agents on alignment

Arena, the crowdsourced model leaderboard that began as LMArena, announced a $200 million Series B on 8 October 2026. In its announcement, the company says the round values it at $3.1 billion, that it has “exceeded $100M in annualized revenue,” and that Lightspeed Venture Partners and Khosla Ventures co-led, with Salesforce Ventures, 01 Advisors, Dell Technologies Capital and Endeavor Catalyst joining existing investors.

The bigger change for people who read leaderboards is a new one: the Alignment Index.

What the index measures

Arena says the index starts with three signals that can be checked against an actual agent trace from real use in Agent Arena:

• Unauthorized action: the model acts beyond what the user asked.
• False attribution: the model attributes a statement, intention or fact to the user that the user’s own evidence contradicts.
• Deceptive completion: the model tells the user a task is complete when it is not.

Arena says these follow definitions that labs such as OpenAI and Anthropic publish in their system cards, so the index can act as an independent check. It is labeled “Preliminary.”

The first scores

The leaderboard page is dated 30 September 2026 and covers 72,509 sessions across 27 models. Higher is better. The top five entries are OpenAI models: GPT-6.1 Sol at 87.9 (plus or minus 1.5), GPT-6 Astra and GPT-6 Luna at 87.8, GPT-6 Sol at 87.6 and GPT-5.6 Sol at 84.2. Claude Opus 5.5 follows at 83.2 (plus or minus 2.0), then Grok 4.7 at 82.7. Several neighbouring scores overlap within their error bars, so small gaps are not meaningful.

Deceptive completion is the signal that separates models most. GPT-6.1 Sol’s rate is 2.34%; Claude Opus 5.5’s is 6.41%; Gemini 4 Argon’s is 12.86%.

Where the numbers disagree

Arena’s own materials do not agree on the size of the dataset. The leaderboard says 72,509 sessions. The summary card for Arena’s research post says “90,000 real-world agent sessions” across 27 models. The Series B post says initial results cover “20+ frontier models.” The valuation is also reported differently: Crypto Briefing says the round closed on 22 September at a $2.88 billion post-money valuation, while Arena and TechCrunch give $3.1 billion. This article uses Arena’s figures and attributes them.

How to use it

The index measures behaviour in Arena’s own agent harness with Arena’s users, so a model’s rate there may not match its rate in your stack. Use it as a prompt to test: if deceptive completion matters to you, add a check in your own agent loop that verifies claimed work, such as re-running tests, before trusting a “done.”