Arena said it raised a $200 million Series B at a $3.1 billion valuation, with Lightspeed Venture Partners and Khosla Ventures co-leading the round. TechCrunch reports that the financing follows a $150 million Series A at a $1.7 billion post-money valuation in January. The company said its annualized run-rate revenue reached $100 million in June.
The funding announcement arrived with a preview of the Arena Alignment Index, which the company says compares 27 models across 90,000 real-world agent sessions. It tracks three kinds of failure: acting beyond a user's request, misattributing statements or facts to the user, and claiming a task is complete when evidence shows it is not.
How Arena built the index
Arena says it created rubrics for each signal, refined them through human review, then used an LLM judge to label eligible sessions. The company adjusted the resulting rates for conversation length and weighted unauthorized actions at 50% of the final score, with false attribution and deceptive completion at 25% each.
In Arena's current table, GPT-6.1 Sol leads with an index score of 87.9, followed by Claude Opus 5.5 at 83.2 and Grok 4.7 at 82.7. Arena also reports that deceptive completion appeared in 10% of sessions on average and 48% of code-debugging sessions. Those are Arena's preliminary results, not a broad measure of every safety risk: the company says its three signals cover only a small part of alignment.
Comments