--- title: "ByteDance's Doubao 2.0 Cuts Token Prices by 10x as Seedance 2.0 Goes Live" date: 2026-10-03 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-model-bytedance-doubao2 language: en --- # ByteDance's Doubao 2.0 Cuts Token Prices by 10x as Seedance 2.0 Goes Live > On 14 February 2026 ByteDance's Volcano Engine launched Doubao-Seed-2.0, with video model Seedance 2.0 and image model Seedream 5.0 Lite arriving days earlier — and token prices roughly one order of magnitude below overseas peers. One company dropped a language model, a video model, and an image model in the same week — and then priced the language one like a commodity. The move turns a research showcase into a line item cheap enough to ignore. For enterprises watching their AI bill, that changes the math overnight. ## The release that wasn't one model On **14 February 2026**, ByteDance's cloud arm, Volcano Engine (火山引擎), launched the **Doubao-Seed-2.0** family — the first major version upgrade since Doubao debuted in May 2024. Two days earlier it had already shipped the video model **Seedance 2.0** (12 February) and the image model **Seedream 5.0 Lite** (13 February). Three core models, one coordinated push into multimodal agent tooling. Doubao 2.0 is positioned as a multimodal Agent model, not a chatbot. ByteDance says it was rebuilt around how enterprises actually call models in production: reading messy charts and documents, following multi-step instructions, and chaining tool calls. ## What Doubao 2.0 actually does The family comes in four variants — **Pro, Lite, Mini, and a Code model** — all supporting text, image, and video input. ByteDance claims strong showings across 19 benchmarks with 12 first places, citing maths and visual reasoning (MathVista, MathVision), document understanding (ChartQAPro, OmniDocBench 1.5), and long-context tests (DUDE, MMLongBench). On video, it says Seed 2.0 Pro exceeds human scores on EgoTempo (71.8 vs a 63.2 human baseline) for sensing change, motion, and rhythm. Agent capability is the selling point. On HLE-Text ("Humanity's Last Exam"), ByteDance reports **54.2 for Pro**; a secondary test put Pro at 52.4, ahead of Zhipu GLM-5.0 and MiniMax 2.5 but behind Alibaba's Qwen-3-Thinking-Max. In science, it claims SuperGPQA ahead of GPT-5.2 and a HealthBench first. These are company-reported; treat them as ByteDance's claims. ## The price is the story Where Doubao 2.0 genuinely breaks from the pack is cost. Volcano Engine's published RMB rates (per million tokens, ≤32K input) are: - **Pro: RMB 3.2 input / RMB 16 output** (≈ US$0.44 / US$2.21; HK$3.45 / HK$17.20) - **Lite: RMB 0.6 input / RMB 3.6 output** (≈ US$0.08 / US$0.50; HK$0.65 / HK$3.90) - **Mini: RMB 0.2 input / RMB 2 output** (≈ US$0.03 / US$0.28; HK$0.22 / HK$2.20) ByteDance says token pricing dropped by roughly **one order of magnitude** versus overseas peers. A local comparison: the Lite input price of RMB 0.6 per million tokens is about **one-tenth of Zhipu's GLM-5** and roughly **1/35th of Claude Sonnet 4.5's** list rate. For agent workloads — which burn tens of times more tokens than a chat — that gap compounds fast. ## Seedance 2.0: video that takes direction Seedance 2.0 is built to behave like a director, not a slot machine. It accepts **text, video, audio, and image** inputs together, so a creator can pin a look with one image, set motion with a video, and set pace with a few seconds of audio. ByteDance says physical-law adherence and character consistency improved, making it usable for film, advertising, and marketing rather than just novelty clips. The model reached users through Jianying (CapCut) and Doubao within days of launch. ## Seedream 5.0 Lite: images that check the news Seedream 5.0 Lite unifies understanding and generation in one architecture. Its notable addition is **real-time retrieval augmentation** — it can pull fresh information from the web to answer time-sensitive creative briefs, useful for news posters and trend-driven assets. ByteDance says it infers intent from short, vague text or image prompts and keeps subject consistency and text-image alignment tight. According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the 2026 Chinese model race has shifted from raw parameter scale to cost-per-task and multimodal reliability — exactly where ByteDance is now aiming its prices. ## Honest limitations The 14 February 2026 launch date, the Doubao-Seed-2.0 family structure (Pro/Lite/Mini/Code), the 12–13 February dates for Seedance 2.0 and Seedream 5.0 Lite, the 256K context, and the RMB price table come from Volcano Engine-adjacent coverage (Science and Technology Daily, Sichuan Online, Eastmoney, 53AI) and model aggregators (TokenMix). The **benchmark claims** (19 benchmarks / 12 firsts, EgoTempo 71.8, HLE-Text 54.2, SuperGPQA, HealthBench) are **ByteDance-reported**; we found no independent third-party audit at writing. RMB-to-USD/HKD use ≈7.24 and ≈7.8. The expert quote is a qualitative observation attributed to BrainNet (脑机网) with no quantifiable claim. Some catalog sources list Seedream/Seedance "5.0 / 2.0" with different internal dates; we used the launch-window dates corroborated by multiple outlets. ## What readers can do now - **If you run high-volume agents:** prototype on Doubao-Seed-2.0-Lite or Mini and measure your bill against the RMB 0.6 / RMB 0.2 per-million input rates — the 10× drop matters most at scale. - **If you make video or image ads:** test Seedance 2.0's multi-modal input (image+video+audio) and Seedream 5.0 Lite's live-retrieval for time-sensitive creatives. - **If you procurement-compare:** wait for independent agentic benchmarks before trusting the HLE / SuperGPQA numbers; the price advantage is real, the score lead is vendor-reported. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-model-bytedance-doubao2 Free to quote with attribution and a link to the original.