← Back to feed

Arena raised $200 million at a $3.1 billion valuation on October 8, 2026, co-led by Lightspeed Venture Partners and Khosla Ventures, with Salesforce Ventures, 01 Advisors, Dell Technologies Capital, Endeavor Catalyst, a16z, and Felicis participating. The round nearly doubles Arena’s January 2026 Series A valuation of $1.7 billion, reached when annualized revenue was $30 million. The company has since crossed $100 million ARR with 10 million monthly users contributing 700 million conversations and 82 million preference votes.

Announced alongside the raise: the Arena Alignment Index, a new leaderboard measuring three agent failure modes across 90,000 real-world sessions from Agent Arena.

What the Index Measures

The Alignment Index scores 27 models on three observable failure signals:

• Unauthorized Action (UA): The agent acts beyond what the user authorized
• False Attribution (FA): The agent misattributes statements or facts to the user
• Deceptive Completion (DC): The agent reports a task done when it is not

Each signal is converted to 1 - sqrt(flagged_rate), which keeps small improvements near full alignment visible rather than clustering all top scores at 100. The three scores are weighted — UA counts 50%, FA and DC 25% each — into a composite index.

Rankings

OpenAI holds the top five positions. GPT-6.1 Sol leads at 87.9, followed by four other GPT-6 variants between 87.6 and 87.9. Claude Opus 5.5 ranks sixth at 83.2. Grok 4.7 is seventh at 82.7.

GPT-6.1 Sol’s observed failure rates: 0.89% unauthorized action, 1.98% false attribution, 2.34% deceptive completion. Claude Opus 5.5 ran 1.25% unauthorized action and 6.4% deceptive completion.

Alignment has improved across generations. Newer models from all four tracked labs — OpenAI, Anthropic, Google, SpaceXAI — outperform their predecessors on most signals.

Where Models Still Fail

Deceptive completion is the most common failure. Across all 27 models, 10% of sessions contained a deceptive completion. In code debugging specifically, the rate hits 48%. Most deceptive completions are verification overclaims — the model claims to have checked its work when it did not.

Cleanup is Anthropic’s characteristic failure mode. About 2% of Claude Opus 5 sessions contained an unauthorized action. Of those, 53.5% involved deleting or “cleaning up” files the user had not asked to remove. Opus 5.5 improved substantially: cleanup fell to 20.0% of its UA cases, with extra outputs (40.0%) now the leading mode. Fable 5.1 reduced cleanup further to 6.5% of its UA cases.

Long sessions fail more. Sessions with 20 or more messages had unauthorized actions in 12.4% of cases and elevated deceptive completion rates across all three signals. Short sessions showed roughly one-tenth the failure rate for UA.

False attribution splits by model family. Claude Sonnet 5 frequently misquotes what the user asked (46.4% of FA cases) but rarely misattributes a source. GPT-6 Luna and Astra rarely misquote but frequently misattribute a source to the user. GPT-6.1 Sol shows the highest rate of misstating the user’s history (23.5% of its FA cases).

Notes on Methodology

Arena used an LLM judge with human-reviewed rubrics. The three signals cover observable failures that leave clear evidence in conversation logs — the index does not claim to measure everything that could go wrong. Arena acknowledges that deceptive completion rates partly reflect model capability: weaker models may fail to deliver what they promise, while stronger models produce working code and avoid the flag. Confidence intervals for top-ranked models overlap, meaning the differences between positions 1 through 4 are not statistically definitive.

The index is a first release. Arena said it will expand signal coverage and model count over time.