Overview
Seed 2.0 is ByteDance's second-generation foundation model family, released February 14, 2026. It powers the Doubao app — China's leading AI chatbot with 155 million weekly active users (per QuestMobile, Dec 2025) — and is available via the Volcano Engine API. The release introduces four model variants (Pro, Lite, Mini, and Code), each systematically optimized for large-scale production deployment. Compared to Seed 1.8, Seed 2.0 strengthens three areas: multimodal understanding (particularly video and complex document parsing), execution of multi-step instructions, and cost efficiency across different usage scales. The model is positioned as a cost-efficient alternative to Western frontier models, delivering competitive benchmark performance at roughly 3.7–10x lower API pricing than GPT-5.2 and Claude Opus 4.5.
What's New
Four Model Tiers for Every Workload
Seed 2.0 ships with four distinct variants rather than a single flagship model, each optimized for a specific use case:
- Seed 2.0 Pro — Frontier reasoning: 98.3 on AIME 2025, 3020 Codeforces rating. Best for complex research, deep reasoning, and multi-step AI agent workflows. Input: $0.47/M tokens, Output: $2.37/M tokens.
- Seed 2.0 Lite — Production workhorse balancing performance and cost. Scores 93 on AIME 2025 and 2233 Codeforces. Best for enterprise chatbots, document pipelines, and general API workloads. Input: $0.09/M, Output: $0.53/M.
- Seed 2.0 Mini — Optimized for inference throughput and deployment density. Handles high-concurrency batch jobs at the lowest cost in the family. Input: $0.03/M, Output: $0.31/M.
- Seed 2.0 Code — Tuned for software development workflows: code generation, debugging, and automated PR reviews. Performance benchmarks for Code versus Pro are detailed in the official Seed 2.0 model card.
Comprehensive Multimodal Upgrade
Seed 2.0 delivers the largest multimodal improvement in the series to date:
- Video understanding — Processes hour-long videos and answers questions about temporal sequences, motion, and content. Scores 89.5 on VideoMME (long-video) and 75.2 on MotionBench. Surpasses human-level performance on EgoTempo (as reported by the Seed team in the official model card).
- Visual mathematical reasoning — Reaches SOTA on MathVision (88.8) and MathVista (89.8), improving significantly over Seed 1.8's 81.3 on MathVision.
- Document and chart understanding — Top-tier on ChartQAPro (71.2) and OmniDocBench 1.5 for handling complex, unstructured document layouts common in enterprise workflows.
- Streaming video QA — Strong performance on real-time video analysis benchmarks (OVBench: 69.2, LiveSports-3K: 78), enabling fitness coaching, live monitoring, and environment-awareness applications.
On the LMSYS Chatbot Arena, Seed 2.0 ranks 6th overall (Text Arena) and 3rd (Vision Arena) as of Feb 16, 2026 (per Seed team; arena rankings change frequently).
Enhanced LLM and Agent Capabilities
Beyond benchmarks, Seed 2.0 specifically targets the gap between solving competition math and completing real-world long-horizon tasks:
- Long-tail knowledge — Seed 2.0 Pro exceeds GPT-5.2 on SuperGPQA, matching Gemini 3 Pro on STEM benchmarks including FrontierSci. Achieves gold medals on IMO, CMO, and all five ICPC programming contests tested.
- Instruction following — Improved consistency and controllability on MultiChallenge (68.3), COLLIE (93.9), and MARS-Bench (85.6), building the foundation for reliable multi-constraint agent workflows.
- Agentic search and deep research — Scores 77.3 on BrowseComp, 74.7 on WideSearch, and 90.4 on tau2-Bench (Retail), demonstrating production-ready agentic performance across research, search, and tool-use tasks.
- Science discovery — Leads on AInstein Bench, and produces hypothesis-driven, step-by-step experimental plans for scientific research tasks that domain experts rate as exceeding typical LLM output.
Significant Cost Reduction vs. Competitors
Seed 2.0's pricing is the headline change for API users:
| Model | Input $/1M | Output $/1M | vs. GPT-5.2 |
|---|---|---|---|
| Seed 2.0 Pro | $0.47 | $2.37 | ~3.7x / ~5.9x cheaper |
| Seed 2.0 Lite | $0.09 | $0.53 | ~19x / ~26x cheaper |
| Seed 2.0 Mini | $0.03 | $0.31 | ~58x / ~45x cheaper |
| GPT-5.2 (reference) | $1.75 | $14.00 | — |
| Claude Opus 4.5 (ref) | $5.00 | $25.00 | — |
In agentic workloads consuming 1 billion output tokens per month, switching from GPT-5.2 to Seed 2.0 Pro reduces API costs from ~$14,000 to ~$2,370 — an 83% reduction.
Availability & Access
Seed 2.0 is available through two main channels:
Doubao App (Consumer)
- Web chat is available at doubao.com/chat; access requirements may vary by region — verify on-site
- Seed 2.0 Pro and Code models are live in the Doubao app and TRAE (ByteDance's AI coding IDE)
Volcano Engine API (Developers)
- Full series API is live at console.volcengine.com/ark
- OpenAI SDK-compatible — change the base URL and API key; existing code works without modification
- Important: Volcengine requires account verification including real-name authentication for many services; phone or email sign-up options may exist, but availability and requirements vary by region. International developers may face additional friction and should verify current requirements before signing up.
System Requirements & Limitations
- No local deployment option — cloud-only inference via Doubao app or Volcano Engine API
- API rate limits apply per tier; agentic workflows with high output volumes deplete limits faster
- Seed 2.0 Code and Pro share pricing — choose based on task type rather than benchmark scores alone
Pricing & Plans
| Model | Input $/1M | Output $/1M | Best For |
|---|---|---|---|
| Seed 2.0 Pro | $0.47 | $2.37 | Frontier reasoning, research, complex agents |
| Seed 2.0 Lite | $0.09 | $0.53 | General production, chatbots, document pipelines |
| Seed 2.0 Mini | $0.03 | $0.31 | High-throughput batch, content moderation |
| Seed 2.0 Code | ~$0.47 | ~$2.37 | Software development, code generation |
| Doubao App | Free | — | Consumer chat via doubao.com/chat |
Reported pricing based on published API rates (source: Volcengine console / third-party reporting); denominated in USD for reference. Volcengine bills domestic accounts in CNY. Enterprise terms may be available — check Volcengine sales or contract terms for details.
Pricing note: API pricing is subject to change. Verify current rates at console.volcengine.com/ark before committing to production workloads.
Best For
- Developers and teams running token-intensive API workloads (1M+ tokens/day) who need frontier-level performance at 3.7–10x lower cost than GPT-5.2
- Video analysis pipelines requiring hour-long video understanding, temporal reasoning, and motion detection at scale
- Enterprise teams needing versatile model coverage — from the cheapest batch classification (Mini) to frontier reasoning (Pro) — under a single API contract
- Startups spending $5K–$50K/month on GPT-5.2 or Claude Opus 4.6 who want to materially extend AI budget runway
- Products built for Chinese-speaking markets, leveraging Doubao's dominant position as China's #1 AI chatbot
- Teams evaluating the most capable AI chatbot APIs for multimodal document and video workflows
FAQ
What is Doubao Seed 2.0?
Seed 2.0 is ByteDance's second-generation foundation model family, released February 14, 2026. It powers the Doubao app and is accessible via the Volcano Engine API. The release introduces four model variants (Pro, Lite, Mini, Code) designed for different cost and capability trade-offs. It is the successor to Seed 1.8, with improvements in multimodal understanding, instruction following, and agent task execution. See the best AI chatbots comparison to see how it stacks up against ChatGPT, Claude, and others.
How does Seed 2.0 pricing compare to GPT-5.2 and Claude?
Seed 2.0 Pro costs ~$0.47/M input and ~$2.37/M output — approximately 3.7x cheaper on input and 5.9x cheaper on output than GPT-5.2 ($1.75/$14.00). Compared to Claude Opus 4.5 ($5.00/$25.00), the savings are roughly 10x on both input and output. For monthly workloads consuming 1 billion output tokens, this represents a cost reduction from ~$14,000 (GPT-5.2) or ~$25,000 (Claude Opus 4.5) to ~$2,370 (Seed 2.0 Pro).
Can I use the Seed 2.0 API outside China?
Direct access to the Volcano Engine API requires Chinese phone number and real-name identity verification, creating significant barriers for international developers. The Doubao consumer app (doubao.com/chat) is accessible globally without registration. For API access, third-party gateways that provide global access to Chinese AI models are available, though not officially supported by ByteDance.
What is the difference between Seed 2.0 Pro, Lite, Mini, and Code?
Pro delivers the highest benchmark performance (98.3 AIME 2025, 3020 Codeforces) and is suited for deep reasoning, research, and complex agent workflows. Lite is the production workhorse — roughly 5x cheaper than Pro while retaining 93 AIME 2025 performance. Mini prioritizes throughput and cost above all else ($0.03/M input) for batch processing and classification tasks. Code shares Pro's core capabilities but is specifically tuned for software development workflows including code generation, debugging, and automated PR reviews.
How does Seed 2.0 compare to Seed 1.8?
Seed 2.0 makes three major improvements over Seed 1.8: (1) multimodal understanding — SOTA gains on MathVision (81.3 → 88.8), VideoMME (87.8 → 89.5), and document parsing; (2) agent capabilities — significant improvements on BrowseComp (67.6 → 77.3) and tau2-Bench tool-use benchmarks; (3) model coverage — introducing four distinct variants versus Seed 1.8's single general-purpose model, offering better cost flexibility. If you're comparing more broadly with other AI assistants, see the best AI assistant guide for a full overview.
