Arena adds price and context columns to text leaderboards

• Arena's leaderboard update adds two implementation-relevant columns to its text leaderboard: token pricing as input/output cost per 1M tokens, and each model's maximum context window.
• The linked leaderboard page turns that into a practical comparison surface, letting engineers weigh rank, price, context length, and license in one table instead of checking vendor docs separately.
• Arena is also pointing users to its methodology explainer on how prompt votes are filtered, categorized, and converted into what it calls "research-grade data."
• In follow-up discussion, Arena's cofounder thread said the team now runs more validation against abuse and exposes category-level views such as coding, math, instruction following, and creative writing.

Arena has added price and context columns directly to its text leaderboard. According to the announcement, price is shown as input and output cost per 1M tokens, while context shows the maximum context window.

That matters because the leaderboard is now doing more than rank ordering models by Arena score. The leaderboard page shows those new fields alongside model score, vote count, and license, so teams can compare quality against hard deployment constraints like token budget and long-context support in one place. Arena frames it as a way to compare models "based on what matters for your use case" in the launch post.

Arena paired the UI change with public material on its evaluation pipeline. In the linked explainer, it says user prompts are tagged by category, low-quality or suspicious activity is filtered out, and duplicate or manipulative votes are removed before they affect rankings.

The follow-up discussion thread adds a little more operational detail. Arena says it tracks many categories beyond the overall score, including "Creative Writing," "Instruction Following," occupational domains, and coding, as shown in

. The same thread says a provider once "switched an endpoint against the policy," and that validation against abuse is "a lot better now since that incident." That does not settle broader benchmark skepticism, but it clarifies that Arena is trying to position the leaderboard as a fresh, multi-category, user-driven benchmark rather than a static test set.