Sacra estimates that Arena hit $100M in annualized revenue in June 2026, up from $30M in 2025.

Arena launched its first paid product in September 2025 and crossed $30M in annualized revenue within four months, reaching a $100M run rate eight months after commercial launch. That ramp suggests the early revenue base was concentrated among a small number of high-ACV customers, frontier model labs and large enterprises, rather than a broad self-serve base.

Revenue is consumption-based rather than pure seat-based SaaS, so it scales with the volume of evaluation campaigns, battles, votes, and prompts a customer runs through the platform. That structure makes revenue less predictable than classic subscription software but ties Arena's growth to how actively labs are testing and releasing models.

The public platform that feeds Arena's commercial product grew alongside revenue: monthly visitors doubled from 5M in January 2026 to 10M in June 2026, with 700M total conversations and 82M total votes accumulated on the platform. Agent Mode, launched roughly a month before the June 2026 milestone, was already generating 5M turns per month and growing around 10% week over week, pointing to a new source of evaluation data and commercial demand.

Arena raised a $150M Series A in January 2026 led by Felicis and UC Investments at a $1.7B post-money valuation. At the time of the raise, Arena had reached over $30M in annualized revenue within four months of launching its commercial product.

Before the Series A, Arena raised a $100M seed round in May 2025 led by Andreessen Horowitz and UC Investments, shortly after incorporating in April 2025. Additional participants across rounds include The House Fund, LDVP, Kleiner Perkins, Lightspeed Venture Partners, and Laude Ventures.

Arena has raised $250M in total primary equity across its seed and Series A rounds.

Arena is a platform where users enter a prompt, receive responses from two anonymous AI models side by side, and vote on which output is better. The model identities stay hidden until after the vote, reducing brand bias in the judgment. Those votes accumulate into public leaderboards that rank models across text, code, search, image generation, video, vision, and document tasks.

The core interaction is Battle Mode: a user submits a prompt, Arena routes it to two models simultaneously, and the user reviews both outputs before voting. After the vote, the model names are revealed, and the user can continue the conversation or start a new one. Instead of relying on post-hoc surveys, Arena collects preference data during live usage.

Arena applies the same blind-comparison mechanic across modalities. In Code Arena, users prompt models to build apps or websites and preview both live outputs before voting, so judgments are based on actual execution rather than a quick read of code. Image Arena and Video Arena use the same setup for generated media, while Search Arena compares models on current-information tasks. Each modality contributes to its own leaderboard segment, making the rankings more task-specific.

On top of that evaluation layer, Arena offers Max and Agent Mode. Max is a routing product that selects which underlying model should handle a given prompt, using millions of community votes as training signal, and now covers text, search, vision, image generation, image editing, and front-end coding. Agent Mode handles multi-step goals, with an agent using tools such as web search, code sandboxes, file handling, and image generation.

Agent Mode uses a different evaluation architecture from simple pairwise voting. Arena tracks full task traces and extracts signals from confirmed success, user corrections, tool hallucinations, and bash recovery. In one recent seven-day window, Arena recorded over 160,000 Agent Mode tasks and more than 2 million tool calls, turning the product into a live benchmark for agentic AI.

The commercial layer, AI Evaluations, lets frontier labs and enterprises run structured evaluation campaigns against Arena's user base. Labs can test pre-release model checkpoints against real-world prompts before shipping, and enterprises can benchmark models against their specific use cases, using the same platform that generates public leaderboard data to power private, paid evaluation runs.

Arena is a B2B business built on a free public product. The public platform, where millions of users compare AI models for free, serves as both the data engine and top of funnel for the commercial service. Users get access to frontier models without separate subscriptions, while Arena collects in-the-wild human preference data at scale.

Its commercial product, AI Evaluations, sells that preference signal to labs and enterprises running model assessments. Pricing is consumption-based: customers pay based on the volume of battles, votes, and prompts consumed in their evaluation campaigns rather than a fixed seat count. Revenue therefore scales with testing intensity, linking demand to model development and release cycles.

The cost structure is less software-like than a typical SaaS business. Arena pays for access to third-party model APIs across many providers, bears compute costs for serving chats, image and video generation, agent sandboxes, and file workflows, and invests in the statistical infrastructure required to turn raw votes into credible rankings. Expensive modalities like video and agents carry usage limits tied to those costs.

The core loop is straightforward: more users generate more votes, which improves leaderboard credibility, attracts more model providers and enterprises, funds more product development, and draws more users. In contrast to B2B data businesses where the public product and paid product are separate, Arena's free layer is the production system for its commercial layer.

Neutrality is a structural part of the business model rather than a branding claim. The commercial product depends on Arena being treated as a trusted referee. If labs viewed rankings as pay-to-play, the evaluation service would lose value. Its methodology, applied equally to publicly released models regardless of commercial relationship, is the basis for the paid product.

Arena competes across three layers: public benchmarking, enterprise evaluation tooling, and model routing infrastructure. No single rival spans all three, but each layer creates pressure on Arena's authority, distribution, or enterprise budget capture.

Artificial Analysis is the clearest direct competitor for mindshare among model buyers. Rather than relying on crowdsourced pairwise voting, it runs its own intelligence, price, speed, and latency evaluations and publishes a cross-model comparison surface. Its differentiation is that it combines capability scores with cost and performance data in a single operator-focused view, which appeals to enterprises optimizing for budget and SLA rather than pure preference quality.

Hugging Face leaderboards are a different kind of rival, less authoritative on closed frontier models but deeply embedded in the open-source ecosystem. Because Hugging Face pairs evaluation with model hosting and community distribution, it has stronger pull with open-model developers who want benchmark legitimacy without depending on Arena's public traffic. OpenCompass serves a similar role for research labs and Chinese model ecosystems, with more configurability and regional benchmark depth than Arena's U.S.-centric prompt mix.

LangSmith, Braintrust, Humanloop, and Weights & Biases Weave compete for enterprise evaluation budget by embedding evals in the development lifecycle. LangSmith combines observability with offline and production evaluations. Braintrust supports CI/CD-integrated testing with remote eval infrastructure. Weave ties tracing and scoring into a reproducible management layer.

None of these products replicates Arena's crowdsourced battle mechanic, but they reduce the need for a third-party public benchmark by helping teams build custom eval suites aligned to their own use cases. If enterprises standardize on integrated eval and observability platforms, they may use public leaderboards only intermittently rather than buy Arena's commercial evaluation service. Patronus AI and Confident AI compete more on assurance than ranking, hallucination detection, safety, governance, and audit trails, targeting the same enterprise evaluation budget in regulated or production-critical deployments where procurement wants guardrails rather than a leaderboard position.

OpenRouter is a meaningful threat because it turns real developer usage into ranking. Its model rankings draw on benchmark scores plus live traffic from millions of users, and it exposes many models through a unified API. If developer usage becomes more persuasive than head-to-head preference votes as a quality signal, OpenRouter's distribution layer can pull attention from Arena's public authority without building a dedicated benchmark product.

A deeper risk is vertical integration from both directions. Frontier labs like OpenAI and Anthropic are expanding native eval and guardrail capabilities, which could let them internalize Arena-style optimization using private telemetry and synthetic judges. If labs reduce pre-release access to Arena or route evaluation spend toward internal stacks, Arena's position as an external checkpoint weakens. Scale AI, Mercor, Surge, and Invisible compete for the same evaluation and post-training budget from a different angle, managed human labeling workforces rather than crowdsourced preference data, but they draw from the same budget pool Arena is targeting.

Arena's expansion logic is vertical integration: collecting human preference data, ranking models, selling evaluation services, and routing production traffic. Each layer creates a new revenue surface on top of the same underlying asset.

Max and the Arena API are the clearest business model expansion. AI Evaluations is episodic, with labs buying evaluation campaigns around model releases, while an API gateway that routes production traffic can sit inside live applications and generate recurring consumption volume. Arena already offers an OpenAI-compatible API with routing modes that select models based on quality, cost, and latency tradeoffs, shifting from telling customers which model is best to operating the infrastructure that chooses and serves the model.

Agent Mode adds a second product surface. Moving from single-turn answer comparison to multi-step workflow evaluation targets a larger budget pool: AI agents, software engineering automation, and enterprise task execution, rather than model marketing alone. With 5M agent turns per month growing 10% week over week as of June 2026, Arena is accumulating agentic trace data that static benchmarks cannot replicate.

Arena began with frontier model labs as its primary commercial customers, but the target for AI Evaluations also includes enterprises deciding which models to deploy, how to route workloads, and how to validate agents before production. That is a larger customer base than the small number of labs releasing frontier models.

Fullstack Code Arena, which added databases, API keys, web search, bash, and deployment to the coding environment, expands Arena's utility for developers and entrepreneurs who want a daily work tool rather than a benchmarking toy. Greater use as a work surface gives Arena more realistic task data and more retained users, increasing the volume and density of prompts that feed commercial evaluation campaigns.

BiomedArena, built in partnership with DataTecnica, provides a template for vertical expansion into high-value domains where generic benchmarks are weak and error costs are high. The same playbook, pairing domain experts with Arena's voting interface and ranking methodology, could extend into legal, finance, education, and cybersecurity, where buyers may pay for task-specific evaluation rather than general Elo scores.

Arena's evaluation quality benefits from diversity of prompts, languages, and user types, which creates an incentive for international expansion. The French government's compar:IA project, which uses the same broad concept of large-scale human preference collection for French-language models, establishes a precedent: governments and regional institutions can build sovereign evaluation stacks, and Arena could power or partner on localized public-interest evaluation efforts rather than compete only as a consumer site.

Localized arenas and region-specific leaderboards would also strengthen Arena's commercial product for non-U.S. enterprises and model builders that need evaluation data matched to their own language and cultural context, rather than the English-language prompt distribution that currently dominates the platform.

Benchmark gaming: As Arena's leaderboard becomes a public quality signal for frontier models, labs have increasing incentive to optimize pre-release checkpoints for Arena-style pairwise outcomes, which could gradually degrade the ecological validity of the rankings that underpin Arena's commercial value.

Lab disintermediation: Frontier labs could internalize Arena-style evaluation using private telemetry, synthetic judges trained on harvested Arena prompts, and proprietary preference models, reducing their dependence on Arena as an external checkpoint and shrinking the pre-release testing budget that currently drives a large share of Arena's revenue.

Neutrality tension: Because the same labs whose models Arena ranks publicly are also paying customers for private evaluation campaigns, any perception that commercial relationships influence leaderboard outcomes could damage Arena's public trust, its enterprise credibility, and the flywheel that connects the two.