Seed 2.1 Turbo lands on OpenRouter: 262K coding agent, $0.50/M
ByteDance's cheap long-horizon coder is finally Western-accessible, with Seed-2.0-Code, Seedream 5.0, and Nvidia's open Nemotron Lightning landing too.
Copy markdown
262K coding brain, one OpenRouter call away
Seed 2.1 Turbo is now live on OpenRouter at $0.50/M in and $2.50/M out with a 262K-token context, tuned for code delivery and long-horizon agent workflows. It was previously locked behind ByteDance's Volcano Engine console — now it's a drop-in OpenAI-compatible endpoint you can route agents through today.
The price-to-capability pitch — verify before you switch
ByteDance positions the Seed 2.1 line as comparable to GPT-5.5 on coding delivery, agent long-chain execution, and multimodal understanding, at roughly a fifth of frontier cost. Independent Code Arena testing put the Pro tier near Claude Opus 4.6; Turbo is half Pro's price. Run your own evals before you migrate a pipeline.
Seed-2.0-Code and Seedream 5.0 dropped the same day
Alongside Turbo, ByteDance's dedicated Seed-2.0-Code ($0.50/$3, 262K context) and the Seedream 5.0 Pro image model ($0.045/image) went live on OpenRouter on Aug 12. You can now wire ByteDance's coding and image stack through a single gateway instead of a China-region console.
Elsewhere: Nvidia's Nemotron 3.5 Lightning is a 30B open MoE you can run
Nvidia released Nemotron 3.5 Lightning — 30B total, 3B active — under the permissive OpenMDW-1.1 license with weights, data, and recipes public. Its NVFP4 checkpoint runs on an RTX 5090 or DGX Spark and posts ~86% accuracy while finishing a 10K-task run ~30% faster than Qwen3.6 35B. It works in vLLM, Ollama, and llama.cpp, and the companion NeMo Switchyard router sends each task to the cheapest capable model.