Doubao-Seed-2.1-pro

doubao-seed-2-1-pro-260915

Get API keyAPI reference

A new-generation flagship model moving toward production-grade intelligence, with fully upgraded Coding, Agent, and multimodal capabilities. With stronger autonomous planning, long-chain execution, and dynamic repair capabilities, it is equipped to handle real-world complex enterprise tasks. It features a 1M ultra-long context and 256K maximum output, with outstanding performance in image, video, and spatial understanding.

Context window1.0M

Provider豆包

Released2026/10/20

Playground

Pricing

The same model is available through multiple service channels — choose based on latency, reliability and cost.

Prices in $ / 1M tokens

To pick a channel, add a provider field to the request body, for example "provider": { "channel": "direct" }. Valid values are direct / stable / economical; omit it to use the default channel.

Direct

Direct upstream connection — best when you need native behavior and the full context window.

Input	Output	Cache read	Cache write
0.90/M	4.48/M	0.18/M	<0.01/M

Overview

Input

Text Image

Output

Text

Doubao-Seed-2.1-pro API: The Doubao flagship with 1024K context, fully upgraded across Coding, Agent, and multimodal

Doubao-Seed-2.1-pro is the flagship tier of ByteDance Doubao's Seed2.1 series. The Seed2.1 series was released on June 23, 2026 at the Volcano Engine FORCE conference, with the first version 260628 positioned as "a new-generation flagship built for the Coding and Agent era"; after about three months of iteration, it was upgraded to version 260915 on September 16, 2026 and made fully available on Volcano Ark, while also being integrated into Doubao App, Doubao Work, and TRAE.

The official description of version 260915 is "a new generation of large models moving toward production-grade intelligence," with continued strengthening in three directions — Coding engineering delivery, long-horizon Agent task execution, and multimodal understanding — emphasizing autonomous planning, long-chain execution, and dynamic repair capabilities. The biggest change in specifications is context: version 260628 had only a 256K context window, while version 260915 expands it to 1024K, with a maximum answer and maximum chain of thought of 256K each. The official team also states that this version optimizes token usage efficiency, that Agent task delivery is more stable, and that multimodal understanding of 3D and professional graphics/text has also been enhanced.

Two things to know in advance: deep thinking is enabled by default, and the thinking level can be adjusted among four tiers — minimal, low, medium, and high — with high as the default; although the maximum answer is 256K, only 4K is output by default, so long-output tasks must explicitly set max_tokens.

SeaWhale AI provides Doubao-Seed-2.1-pro through an OpenAI-compatible interface, supporting image and video input, deep thinking, tool calling, and structured output.

Get an API key · Model ID: doubao-seed-2-1-pro-260915

---

Why choose Doubao-Seed-2.1-pro

• 1024K context — 4x that of version 260628 (256K), while also supporting a 256K maximum answer and a 256K chain of thought
• Leading visual perception — official benchmark BabyVision 73.7, higher than GPT-5.5's 55.9 and Gemini 3.1 Pro's 54.4
• Outstanding spatial reasoning — ERQA 72.0, higher than Gemini 3.1 Pro (70.8), GPT-5.5 (64.5), and Claude Opus 4.7 (52.5)
• Strong long-video understanding — VideoMME 89.2 and TOMATO 79.5, both higher than Gemini 3.1 Pro in the official comparison
• Agent task delivery — Agent Startup Bench 68.8, on par with GPT-5.5 (68.1) and higher than Claude Opus 4.7 (62.3)
• Complete engineering interface — GUI task handling, tool calling, structured output, implicit and explicit context caching

---

Core capabilities

01 Long-horizon Agent execution

The core goal of Seed2.1 is to finish long tasks in real production scenarios. The official team says Seed2.1 Pro achieved the highest score on GDPVal, scored 68.8 on Agent Startup Bench, and used reinforcement learning to reduce the average number of execution steps for GUI and non-GUI tasks by 16%. Version 260915 further improves the stability of Agent task delivery; combined with 1024K context, it suits research, office, and engineering tasks spanning several hours.

• Multi-step planning, execution, and dynamic repair after failures
• Long-running research and document delivery tasks
• GUI automation of computer and phone interfaces

02 End-to-end Coding delivery

The official positioning of Seed2.1's coding capability is full-cycle delivery: requirements analysis, feature implementation, bug fixing, environment setup, and verification. In official benchmarks it scores 71.0 on Terminal Bench 2.1, close to Claude Opus 4.7 (71.7) and Gemini 3.1 Pro (70.7). Version 260915 focuses on strengthening multimodal Coding, writing and modifying code with reference to screenshots, design mockups, and 3D scenes.

• Multi-step engineering tasks in terminal environments
• Front-end development and fixes based on UI screenshots and design mockups
• Already integrated into TRAE, and can also be connected to programming tools such as Claude Code through the compatible interface

03 Image and video understanding

Multimodal is the area where Seed2.1 Pro has the clearest advantage over competitors. In official benchmarks: WorldVQA 53.0 (GPT-5.5 is 34.6), CharXiv-RQ 85.4, MMMU-Pro 81.6, MathVision 92.6; on the video side, VideoMME 89.2, OVOBench 80.7, MMLongBench-128K 78.3. Version 260915 also enhances understanding of 3D objects and densely formatted documents.

• Professional graphics/text parsing such as charts, financial report tables, and engineering drawings
• Understanding and retrieval of long videos and streaming video
• Spatial reasoning and 3D scene understanding

04 Adjustable thinking intensity and engineering features

The thinking level offers four tiers — minimal, low, medium, and high — with high as the default; keep the default for complex tasks, and lower it for latency-sensitive scenarios. It also supports function calling, structured output (the json_schema mode is recommended), and explicit caching such as implicit caching, prefix caching, and session caching, reducing the cost of multi-turn agent calls.

• Four thinking intensities switched by task
• Tool calling and JSON structured output
• Context caching reduces the cost of repeated calls with long prefixes

---

Best use cases

Scenario	Description
Long-horizon agents	Research, office, and engineering tasks lasting multiple hours; 1024K context holds the entire process
Multimodal Coding	Writing and fixing code with screenshots, design mockups, and 3D scenes
Professional document parsing	Understanding and extraction of financial report tables, engineering drawings, and scientific charts
Video understanding	Long-video retrieval, content analysis, streaming video Q&A
GUI automation	Reading UI screenshots and planning and executing operation steps
Domestic flagship selection	Compared side by side with domestic flagships such as GLM, Qwen, and DeepSeek

---

Differences between Doubao-Seed-2.1-pro and other versions in the same series

Capability	Doubao-Seed-2.1-pro (260915)	Doubao-Seed-2.1-pro (260628)	Doubao-Seed-2.1-turbo	Doubao-Seed-2.1-lite
Model ID	doubao-seed-2-1-pro-260915	doubao-seed-2-1-pro-260628	doubao-seed-2-1-turbo-260628	doubao-seed-2-1-lite-260915
Release date	September 16, 2026	June 23, 2026	June 23, 2026	September 2026
Context window	1024K tokens	256K tokens	256K tokens	1024K tokens
Max answer	256K tokens	256K tokens	256K tokens	256K tokens
Max chain of thought	256K tokens	256K tokens	256K tokens	256K tokens
Input modalities	Text, images, video	Text, images, video	Text, images, video	Text, images, video, audio
Deep thinking	Enabled by default, four adjustable levels	Enabled by default	Enabled by default	Enabled by default, can be disabled
Positioning	Series flagship, production-grade Agent and Coding	Series debut flagship	Low cost, low latency, performance comparable to Pro	Lightweight and high-frequency, office and multimodal

Specifications come from the Volcano Ark model list and model release announcements; the benchmark scores in this article come from the official Seed Seed2.1 page and correspond to the June debut version — the official team has not yet published a complete benchmark table for version 260915. Specific billing is subject to the real-time price card at the top of the page.

---

FAQ

When was Doubao-Seed-2.1-pro released? The Seed2.1 series was released on June 23, 2026 at the Volcano Engine FORCE conference, with 260628 as the first version. It was upgraded to version 260915 on September 16, 2026, and the API became fully available on Volcano Ark.

What did 260915 upgrade compared with 260628? In terms of specifications, the context window expanded from 256K to 1024K. In terms of capabilities, the official team says this version improves the stability of Agent task delivery and multimodal Coding, enhances 3D and professional graphics/text understanding, and optimizes token usage efficiency, lowering the overall call cost for equivalent tasks.

What are the context and output limits? The context and maximum input are both 1024K tokens, the maximum answer is 256K tokens, and the maximum chain of thought is 256K tokens. The default answer length is only 4K, so explicitly set max_tokens when you need long output.

How do I set the thinking intensity? Deep thinking is enabled by default, and the thinking level supports four tiers: minimal, low, medium, and high, with high as the default. Keep the default for complex coding and long-horizon Agent tasks; for latency-sensitive tasks such as conversation, classification, and extraction, it is recommended to set it to low or minimal.

Which tools and features are supported? It supports image and video input, function calling, structured output (the json_schema mode is recommended), GUI task handling, and explicit caching such as implicit caching, prefix caching, and session caching. Note that Pro does not support audio input; if you need to directly understand speech, choose Doubao-Seed-2.1-lite.

Does Doubao-Seed-2.1-pro lag anywhere? Yes. In official benchmarks, it still lags behind top international models on long-horizon coding and debugging: NL2Repo-Bench 47.0 is lower than Claude Opus 4.7's 58.2, and SWE-Atlas Debugging 35.2 is lower than GPT-5.5's 44.7; on the white-collar office benchmark xDailyBench it scores 61.0, also lower than GPT-5.5 (73.0) and Claude Opus 4.7 (69.0); on the knowledge-oriented SuperGPQA it scores 70.8, lower than Gemini 3.1 Pro's 76.6. The official team also acknowledges that there is still room for improvement on the most challenging open-ended tasks and frontier research questions. In addition, third-party hands-on feedback indicates that its front-end aesthetics for open-ended web and canvas self-drawing are not stable enough.

Should I choose 2.1-pro or 2.1-turbo / 2.1-lite? Choose Pro when you need 1024K ultra-long context and the most complex Agent and multimodal Coding tasks; choose Turbo for high-volume online calls where cost and latency matter more and 256K context is sufficient; choose Lite for high-frequency office work, audio input, or batch processing.

---

Why use the Doubao-Seed-2.1-pro API with SeaWhale AI

• No need to activate a Volcano Engine account — call directly with a single key, no separate application for Ark model access
• OpenAI-compatible interface — the multimodal request format matches OpenAI; just change the model name in your existing code
• Compare old and new versions under the same account — share one key with Doubao-Seed-2.1-lite, 2.0-pro, and other domestic flagships
• Pay-as-you-go — billed by actual token usage, no minimum spend

API

API integration

Model IDUse this value as the model in inference requests

doubao-seed-2-1-pro-260915

API KeyBearer token used to authenticate inference requests

Manage API keys

Base URLOpenAI compatible · /chat/completions

https://api.haijingai.com/v2/

provider OptionalSelects a service channel; omit it and the system picks the default

"provider": { "channel": "direct" }

doubao-seed-2-1-pro-260915 usage examples

SeaWhale AI is compatible with the OpenAI API protocol, so you can call it with the OpenAI SDK or plain HTTP requests. Streaming is enabled by default.

cURL

curl https://api.haijingai.com/v2/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API_KEY>" \
  -d '{
    "model": "doubao-seed-2-1-pro-260915",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Hello!"}
    ],
    "provider": { "channel": "direct" },
    "stream": true
  }'
# provider is optional — remove this line to use the default channel

Python

from openai import OpenAI

client = OpenAI(
    base_url="https://api.haijingai.com/v2",
    api_key="<API_KEY>",
)

stream = client.chat.completions.create(
    model="doubao-seed-2-1-pro-260915",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello!"},
    ],
    stream=True,
    # Optional: pick a service channel; omit to use the default
    extra_body={"provider": {"channel": "direct"}},
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

Node.js

import OpenAI from 'openai'

const client = new OpenAI({
  baseURL: 'https://api.haijingai.com/v2',
  apiKey: '<API_KEY>',
})

const stream = await client.chat.completions.create({
  model: 'doubao-seed-2-1-pro-260915',
  messages: [
    { role: 'system', content: 'You are a helpful assistant.' },
    { role: 'user', content: 'Hello!' },
  ],
  stream: true,
  // Optional: pick a service channel; omit to use the default
  // @ts-expect-error provider is a SeaWhale AI extension, not in the OpenAI SDK types
  provider: { channel: 'direct' },
})

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? '')
}