HomeAI models
ByteDance Seed: Seed 1.6
Bytedance
•
Seed 1.6
•
Released 
June 2025

Bytedance
Seed 1.6

Adaptive-thinking multimodal MoE with a 256K window, for mixed-difficulty reasoning over text and images.

Modality:
Text
Image
PDF
model ID
bytedance-seed/seed-1.6
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

N/A
tok/s

Intelligence Index *

N/A
/ 100

Context Window *

262144
tokens

Input price

1.5
Anytoken

Output price

12
Anytoken
Seed 1.6: Adaptive Reasoning Across Text, Images and 256K Context Seed 1.6 is a general-purpose multimodal model from ByteDance's Seed team, built as a sparse Mixture-of-Experts with 230B total and 23B active parameters. It pairs Adaptive Chain-of-Thought, which sets thinking depth by question difficulty, with a 256K context window and image input. In the lineup it is the generation before Seed 1.8 and the full-size sibling of Seed 1.6 Flash. It suits mixed-difficulty workloads such as document and screenshot analysis, long-context reasoning and tool-using agents. Access Seed 1.6 via AnyAPI.ai

Performance

Adaptive Thinking Reserves Long Reasoning for Prompts That Need It

Seed 1.6 spends reasoning effort selectively. ByteDance's Adaptive CoT starts deep thinking based on difficulty; a secondary analysis reports it triggers about 37% of the time on MMLU-style questions and 90–100% on hard math and logic. In ByteDance's Seed1.8 model card, Seed 1.6 scores 61.0 on MMMU-Pro in NoThink mode. Interpretation: long reasoning chains are spent mainly where problems need them. Production consequence: mixed traffic can share one endpoint, but latency and token use vary per request, so plan for variable output length.

Technical Specifications

What the model supports

Seed 1.6 takes text and image input (some listings also show video) and returns text only. The context window is 256K tokens, listed as 262,144 on OpenRouter, but output is capped at 32,768 tokens, and reasoning tokens count toward that cap. For long-document or hard-reasoning requests, the output ceiling, not the context window, is the practical constraint. Size max-token settings so thinking does not crowd out the final answer.
Verified Specifications — 
Seed 1.6
*
Input modalities
Text
Image
PDF
output modalities
Text
Context window
262144
 tokens
Maximum output tokens
32768
Reasoning
Yes
Knowledge cutoff
June 2025
Pricing (standard)
1.5
 AnyTokens in
 / 
12
 AnyTokens out

Comparison

Seed 1.6 or Seed 1.8: Is Staying on the Older Generation Worth It?

Seed 1.8 is Seed 1.6's direct successor from the same ByteDance Seed team, released in December 2025. Both are multimodal reasoning models with 256K context, and at least one aggregator lists them at comparable rates. ByteDance's Seed1.8 model card reports MMMU-Pro of 65.4 in NoThink mode against 61.0 for Seed 1.6, and 81.3 on MathVision, a 9.5-point gain over Seed 1.6. The decision is whether a validated Seed 1.6 integration justifies delaying a move to the newer, agent-focused model.

Dimension
Seed 1.6
Seed-2.0-Lite
Context window *
262144
tokens
262144
tokens
Output speed *
N/A
tok/s
222.15
tok/s
Intelligence Index *
N/A
8.7
Input pricing
1.5
AnyToken
1.5
AnyToken
Output pricing
12
AnyToken
12
AnyToken
Knowledge cutoff *
June 2025
February 14, 2026

Choose Seed 1.6 when you have prompts, evaluations and output formats already tuned against it and a migration would need re-validation, or when you are testing a lower-risk fallback. Choose Seed 1.8 for new builds, stronger multimodal reasoning, and agent-style work such as search, GUI operation and coding, which its model card emphasizes. Seed 1.6's availability on some hosts is ending, so Seed 1.8 is the safer default for anything long-lived.

Limitations & Trade-offs

Where Seed 1.6 falls short

1
Superseded and being retired on some hosts. OpenRouter lists Seed 1.6 as going away on November 11, 2026, and ByteDance's own platform moved users to Seed 1.8 and later. A production dependency on this model carries migration risk. Treat it as a bridge or a baseline for regression tests, and confirm current availability before pinning it into a long-lived system. For new projects, Seed 1.8 or Seed 2.x is the better starting point.
2
No independent benchmark coverage. BenchLM shows no sourced benchmark rows for Seed 1.6, and aggregators report no coding or agentic scores. Available figures are ByteDance-reported or from secondary analysis. For coding assistants, autonomous agents and strict instruction following, run your own evaluations before committing. Models with published third-party results are easier to justify for those workloads.
3
Output cap is far below the context window. Input can reach 256K tokens, but a response is limited to 32,768 tokens, and reasoning tokens count toward that limit. Hard problems that trigger deep thinking can leave less room for the final answer, and long generations such as full reports or large code files may truncate. Some listings also apply higher input rates to long prompts, with thresholds that differ between sources.
4
Text-only output and inconsistent video support. Seed 1.6 returns text only, so it cannot generate images, audio or video. Image input is consistently listed, but video input appears on some provider pages and not others, so check the exact route before building video-understanding features. Workloads that need generation of non-text media or confirmed video analysis should use a different model.

Best-Fit Workloads

Where this model earns its place

01

Screenshot and Document Analysis

‍
Seed 1.6 accepts image input alongside text, and ByteDance reports 61.0 on MMMU-Pro in NoThink mode. That suits reviewing screenshots, charts, scanned pages and UI states against written instructions or specifications. Output is text only, so it works for extraction, QA and explanation rather than visual generation. Confirm video support on your specific route before relying on it.

02

Long-Context Reasoning Over Large Documents

‍
The 256K context window lets Seed 1.6 take long contracts, reports or multi-file context in one request, and adaptive thinking applies deeper reasoning where the question requires it. Good for synthesis and cross-reference questions across a large corpus. Keep answers within the 32,768-token output cap, and note that reasoning tokens consume part of it.

03

Mixed-Difficulty Assistants

‍
Adaptive Chain-of-Thought decides whether to think deeply, reportedly engaging far less often on simple knowledge questions than on hard math or logic. A single endpoint can therefore serve both quick lookups and harder analytical queries without manual routing. The trade-off is variable latency and token use per request, which matters for strict response-time targets.

04

Tool-Using and GUI-Oriented Agents

‍
Seed 1.6 supports tools and tool_choice, and ByteDance describes GUI-based interaction as part of the model's capabilities. It can drive function-calling backends and screenshot-driven workflows. Independent agentic benchmarks are absent, so validate reliability on your tasks. For heavier agent work, the newer Seed 1.8, built around agentic evaluation, is the stronger candidate.

Pricing in anytokens via AnyAPI
Input
1.5
₳
Output
12
₳
Cache write
—
₳
Cache read
—
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceTechnical SpecificationsComparisonLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access Seed 1.6 via AnyAPI.ai

Access Seed 1.6 through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Seed 1.6 without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Seed 1.6 and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Seed 1.6 against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Seed 1.6 from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Seed 1.6 for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Seed 1.6 has a 256K-token context window, listed as 262,144 tokens on OpenRouter, and a maximum output of 32,768 tokens. Context and output are separate limits. Reasoning tokens count toward the output cap, so hard prompts that trigger deep thinking leave less room for the visible answer.

Yes. Seed 1.6 accepts text and image input and supports tools with tool_choice, plus structured outputs on the OpenRouter listing. Output is text only. Some providers also list video input, but others do not, so verify video support on your specific route.

Seed 1.6 uses Adaptive Chain-of-Thought (AdaCot), which starts extended reasoning based on question difficulty rather than always thinking. A secondary analysis reports it triggers about 37% of the time on MMLU-style questions and 90–100% on hard math or logic. Some hosts also expose low, medium and high effort levels.

OpenRouter lists Seed 1.6 as going away on November 11, 2026, and a CometAPI notice says the Doubao 1.6 series was discontinued in favor of Seed 1.8. For new builds, Seed 1.8 is the safer default. Keep Seed 1.6 for validated integrations or comparison baselines, and confirm current availability first.

Seed 1.6 sits in the lower-cost tier compared with premium frontier reasoning models, and it is more expensive than its Flash sibling. Some providers raise input rates for very long prompts. Current AnyAPI.ai rates for Seed 1.6 are shown in the AnyTokens pricing component on this page.

* Benchmark data source: Artificial Analysis artificialanalysis.ai