ByteDance used its Volcano Engine FORCE conference in Beijing on June 23 to unveil a broad AI lineup, led by a new flagship language model and a preview of its next video generator. Volcano Engine is the cloud arm of TikTok’s parent that sells ByteDance’s models to businesses.
The headline model, Doubao 2.1 Pro, is pitched at coding, autonomous agents and visual understanding. ByteDance says it has crossed a production-level capability threshold and, in the company’s own benchmarks, outperforms Anthropic’s Claude Opus 4.6 across several metrics, including the Terminal Bench, SWE-Pro and SciCode coding tests and the OSWorld and MMMU-Pro agent and multimodal tests.
The sharper pitch is price. Doubao 2.1 Pro is listed at 6 yuan per million input tokens and 30 yuan per million output tokens, with cached inputs as low as 1.2 yuan, and ByteDance says total cost of ownership runs nearly 80% below Claude Opus 4.6. A Turbo version costs half as much for high-volume use.
The aggressive pricing reflects ByteDance’s scale at home: daily calls across the Doubao family have passed 180 trillion tokens, more than 10 times higher than a year ago, and the firm holds about 49.5% of China’s public-cloud model-as-a-service market, according to IDC. More than 1.1 million companies and developers use its Volcano Ark platform.
On video, ByteDance previewed Seedance 2.5, skipping several version numbers to signal a generational jump. It claims the model can generate a single, continuous 30-second clip in one pass, rather than stitching short segments, and can take up to 50 reference inputs such as images, audio and 3D mockups for tighter control, plus localized editing that changes one part of a frame. It is in enterprise beta, with a public launch targeted for early July and no pricing disclosed.
The company also upgraded Seedance 2.0 to native 4K and introduced an image model, Seedream 5.0 Pro, an audio model, Seed-Audio 1.0, and developer tools including an enterprise agent workbench called ArkClaw.
The Price Pitch
The releases sharpen a strategy of matching frontier capability at a steep discount. If Doubao 2.1 Pro performs near Claude Opus 4.6 at a fraction of the cost, it pressures the pricing of Western labs in the enterprise market, echoing the discounting that DeepSeek and other Chinese developers have used to win developers.
The important caveat is that every performance and cost claim here is ByteDance’s own. No independent benchmarks exist yet, the model comparisons are self-reported, and community claims of parity with Anthropic’s models remain untested. For buyers, the real read will come once outside testing and real workloads arrive.
ByteDance’s AI Push
The lineup shows ByteDance running the same model-family choreography as Google and OpenAI, shipping a flagship language model, a video model, an image model and an audio model in a single keynote.
CEO Liang Rubo framed the priorities bluntly, saying “climbing the AI summit is the most important thing for ByteDance right now,” and the company has narrowed its focus toward model capability and its model-as-a-service business.
ByteDance also previewed an AI copyright platform that lets creators build licensed derivative works, with figures like filmmaker Stephen Chow as early partners. Volcano Engine’s president called video generation a path toward world models, a sign ByteDance sees these tools extending well beyond entertainment into manufacturing and autonomous driving
