ByteDance’s Volcengine has officially launched the Doubao-Seed-2.1 family of large language models and opened their APIs, with all models offering a 256k context window. The new lineup includes two deep reasoning models — Pro and Turbo — with upgrades in complex code delivery, long-chain agent execution, and multimodal vision capabilities.
In official benchmarks, the flagship Doubao-Seed-2.1-pro approaches or surpasses GPT-5.5 on several coding and development tests. On SciCode (scientific computing) and NL2Repo-Bench (library-level code generation), the model scored 59.8% and 47.0%, respectively, edging out GPT-5.5’s 58.4% and 45.1%. On the software engineering benchmark SWE-Pro, it achieved 57.5%, beating Gemini-3.1-Pro’s 54.2% and coming close to GPT-5.5’s 58.6%.
API pricing is set at 6 yuan per million tokens for input and 30 yuan per million tokens for output on the Pro version. The Turbo version is priced at half that: 3 yuan per million tokens for input and 15 yuan per million tokens for output.
Volcengine also introduced Seed-Evolving, a weekly rolling update service. Seed-Evolving lets developers access the latest model — updated at least once a week — through a single Model ID, shifting the delivery model from fixed versions to continuous online evolution.