OpenArt Ranked #1 on Physion's Anime Benchmark. Four Companies Have Said the Same Thing This Year.

In July, Invideo claimed it was the world's #1 AI video agent. In August, Runway was #1. Two weeks later, Dashverse's Frameo was #1 in a different track. On September 10, OpenArt announced it had ranked first overall on the Anime Track of the same benchmark — Physion Labs' ARC 1.0.

That is four different #1s from the same evaluation in four months. Not because the agents are converging on a clear hierarchy, but because the benchmark keeps splitting into new tracks, new participant sets, and new measurement conditions. The headline each company takes home says nothing about a competitive ranking. It says only that one company won a specific test, on a specific day, under a specific set of constraints that happened to favor its orchestration choices.

OpenArt's own post carries the most telling admission: "What makes the difference is the orchestration layer." That's not a product moat. It's a positioning claim — one that every other #1 winner has made in identical language.

The Benchmark Rotation

Physion-Arc 1.0 launched in August 2026 as an evaluation of minute-long, multi-scene video agents. It tested narrative coherence, cinematic language, and production quality across 16 metrics. The original cohort — Runway, Luma, MiniMax, Kling AI, Utopai Studios, and TapNow — produced 600 videos from 100 screenplays. Runway won.

Invideo, which wasn't in the first round, asked to be added. It was evaluated separately against a seven-agent set, generated 700 videos, and posted a mean score of 72.4 — "never below #2 on any metric." Invideo claimed #1.

Then came the Movie track. Frameo, an agent built by Dashverse for AI short dramas, "scored 88.7 out of 100" on 100 prompts across 800 videos — nearly 20 points ahead of the runner-up. Physion's own founder noted the margin was large enough to surprise the team, prompting an extra validation round. Frameo claimed #1.

Now the Anime track. OpenArt's Director agent was evaluated across text-only, single-reference, and multi-reference tasks. "100 prompts and 800 videos" scored across 19 metrics by human evaluators. OpenArt led all eight human-preference metrics and 11 of 19 total metrics. OpenArt claimed #1.

The pattern is clear. Each track has a different agent mix. Different prompts. Different evaluation weights. Different participant counts. A benchmark that produces a new champion every few weeks isn't measuring a hierarchy — it's measuring a moving target.

When every company can be #1, the ranking is a press release, not a ranking.

OpenArt's Actual Numbers

Setting the benchmark aside, OpenArt is a company worth understanding on its own terms. It's private, founded in 2022 by former Googlers Coco Mao and John Qiao, and raised $30 million in a Series A in January 2026 led by Canaan Partners, on top of a $5 million seed round.

The estimated financials, reported by Sacra and echoed by Canaan's investment thesis, paint an unusually efficient operator:

Metric	Value
Estimated ARR	$70M
Monthly active users	8 million
Employees	~20
Revenue per employee	$3.5M
Status	Profitable

The revenue per employee figure is remarkable if accurate. For context, that ratio would place OpenArt above nearly every publicly traded software company. But these are estimates, not audited disclosures. A private company with 20 people and a credit-based subscription model at "Essential: $7/month for 4,000 credits" up to "Infinite: $28/month for 24,000 credits" can generate real revenue — but it can also inflate ARR through annual prepayments, promotional pricing, or one-time enterprise deals that don't reflect recurring unit economics. The $35M in total funding against $70M in estimated ARR suggests a company that crossed payback quickly, which is plausible, but the reader should treat these figures as directional, not confirmed.

OpenArt's business model is straightforward: a consumer and SMB creative platform that wraps open-source AI models behind a simple, template-driven interface. "About 15 pre-configured, no-prompt workflows" including sketch-to-image, upscaling, face replacement — targeting hobbyists, tabletop RPG artists, and small businesses. The video features, including Director, extend the same model: take a high-level idea, produce a multi-shot video up to five minutes long, with character consistency across scenes.

The heavy reliance on SEO-optimized landing pages for discovery is both the company's strength and its fragility. OpenArt ranks as a "top 10 search result for 'AI art generator'" — which means organic traffic is the primary acquisition engine. That's efficient today, but it creates a single point of failure. A Google algorithm change, a shift in search intent toward a new product category, or a well-funded competitor buying the same keywords at scale would immediately pressure the cost of customer acquisition.

The benchmark headline doesn't change the business. The business stands or falls on whether $28 per month is enough to retain 8 million users who are consuming AI video at a rate the company can actually afford to serve.

The Cost Claim And What It Actually Means

OpenArt's benchmark post emphasizes economics: "approximately 15% lower cost than the runner-up and 58% lower than the third-ranked agent." The CEO's LinkedIn post adds that Director achieved a score of 77.8 at a cost of $21 per video.

But this is OpenArt's own cost metric for its own inference pipeline, measured under benchmark conditions — not a total cost of ownership calculation for a customer producing content at scale. The $21 per video number is an interesting data point, but it doesn't tell us the cost to produce a 30-second ad, a five-minute YouTube video, or a recurring content pipeline, which is what actual users need. It also doesn't account for the cost of curation, re-generation, editing, or the human time spent steering the agent — which is often the real expense in AI video workflows.

The cost advantage is also track-specific. OpenArt's Director won on human-preference metrics — "particularly strong subjective preference" driven by "cinematic presentation, visual appeal, and overall viewing experience." Character.ai led the objective evaluation, posting stronger instruction-following, consistency, and execution scores. The benchmark's own framing acknowledges the split: subjective taste versus technical accuracy. A tool that looks better isn't necessarily cheaper to operate in production, especially when the downstream use case demands precision over aesthetics.

The 58% cost advantage sounds decisive until you realize it measures one company's inference cost for one minute of anime on a specific day — not the economics of a real creative workflow.

What This Benchmark Actually Tests

Coco Mao's framing of the competition is worth examining closely. He compares it to an orchestra — the conductor matters more than the individual musicians.

This is a clever positioning argument that serves two purposes simultaneously. First, it deflects from the fact that OpenArt doesn't own its own foundation models. Second, it argues that the competitive moat in AI video is workflow intelligence, not model capability — which means every few months when a new model launches, OpenArt's relative position doesn't reset.

But this claim has been made by every other company that claimed a Physion-Arc #1. Invideo said the category will be won on "technical depth and strategy rather than funding size." Dashverse argued that Frameo's dominance came from building for a specific production use case rather than benchmark optimization. Runway claimed victory on "cinematic taste" — a subjective advantage that shifts with every human evaluation panel.

When every winner claims the moat is different from the last winner's moat, none of them have a moat.

The more useful question isn't which orchestration layer is best today. It's whether orchestration is defensible at all. When the underlying models are open-source or available through API to every competitor, the orchestration layer becomes a feature — and features get copied. The question for OpenArt is whether its 20-person team can iterate on workflow improvements faster than Runway, Invideo, and a dozen other players can replicate them.

The Real Risk For OpenArt

OpenArt is a well-operated private company in a space that is becoming dangerously crowded. The estimated $70M ARR, 8M users, and profitability are encouraging if accurate. But the company faces structural risks that a benchmark headline doesn't address.

The commoditization risk is direct. OpenArt wraps models from DALL-E, GPT, Stable Diffusion, Flux, and others behind a user-friendly interface. As those models improve and become cheaper, the value of the wrapping layer diminishes. A competitor with deeper model integration, lower inference costs, or a more compelling distribution channel can replicate OpenArt's interface in weeks.

The model-dependency risk compounds. OpenArt's video features assemble "over 50 AI models" into an end-to-end stack. Each model dependency is a potential point of failure — pricing changes, API deprecations, quality regressions. When Runway or Luma or a foundation model provider ships a native video agent that does what Director does, OpenArt's orchestration advantage evaporates.

The benchmark says OpenArt is the best at orchestrating other people's models. The investment question is whether orchestrating other people's models is a defensible business.

OpenArt's benchmark #1 is a marketing event, not a competitive signal. The company's actual business — credit subscriptions for AI image and video generation from a consumer and SMB audience — faces the same challenges every model-agnostic platform faces: margin pressure from improving commodity models, distribution competition from better-funded players, and the persistent risk that the foundation model providers themselves become the product layer.

The reader who wants to understand OpenArt should look past the #1 headline and focus on whether the $28 monthly subscription can survive in a market where the underlying capability to generate a one-minute video is approaching zero marginal cost. That's the real test — not the latest Physion-Arc track.