Doubao-Seed-1.8: Observations on the Puter Developer API and Search Agent Selection

One-sentence takeaway

Puter's third-party model card positions Seed1.8 as a search- and Agent-first model rather than a top coding model, and provides a runnable API integration for agents, but its specifications and pricing represent only the Puter route.

Use cases

• Suitable tasks: Quickly trying search, GUI, financial analysis, customer service, and multimodal Agents; prototyping with the JavaScript, Python, or OpenAI-compatible APIs.

• Unsuitable tasks: Treating Puter's free tier/pricing/context limits as a native Volcengine SLA, or using them to replace a controlled coding benchmark.

• Applicable model version: Puter route name volcengine/doubao-seed-1.8; the underlying snapshot is not disclosed.

• Applicable client, Agent, or API: Puter.js, Puter OpenAI-compatible endpoint.

• Recommended reasoning tier and parameters: The page only says that three adjustable thinking modes are available and gives no specific default value; start with the provider default and save the response model and usage.

Test environment

• Integration methods: Puter.js AI API, Python OpenAI SDK, or cURL.

• Example input: Explain quantum computing in simple terms; the page also provides an API/Playground entry point for the same model.

• Evaluation information: The page's model card summarizes BrowseComp-en, VideoMME, token efficiency, and general benchmark selection conclusions, but does not disclose a unified question set, number of repetitions, or a complete harness.

Input/configuration

from openai import OpenAI

client = OpenAI(
    base_url="https://api.puter.com/puterai/openai/v1/",
    api_key="YOUR_PUTER_AUTH_TOKEN",
)

response = client.chat.completions.create(
    model="volcengine/doubao-seed-1.8",
    messages=[
        {"role": "user", "content": "Explain quantum computing in simple terms"}
    ],
)
print(response.choices[0].message.content)

Results

• The page reports BrowseComp-en 67.6 and VideoMME 87.8, and says Seed1.8 is better suited to search/Agent use cases; this can be cross-checked against the official model card's BrowseComp-en 67.6 and VideoMME 87.8.

• The page claims that Seed1.8 reaches the intelligence level of Seed1.6 at about 5K tokens, compared with approximately 15K tokens for Seed1.6; this is a token-efficiency conclusion relayed by the page and lacks per-question data.

• The Puter page lists 256K context, 32K maximum output, input at $0.13 per million tokens, and output at $1.25 per million tokens; these are specifications for the Puter route.

• The page explicitly notes that, in independent evaluations, Seed1.8 may trail Gemini 3 and GPT-5 on general benchmarks, making it more oriented toward search/agent use than coding.

Conclusion

If the goal is to quickly validate multimodal search or tool Agents, Puter's low-friction API is suitable for a smoke test; for production or code benchmarks, switch to the official Volcengine/BytePlus endpoint, fix the snapshot, thinking tier, tools, and billing region, and retest.

Limitations

• Puter is a proxy provider, so its underlying route, rate limits, context, and pricing may differ from those of the native service.

• The page mixes model background such as “released in December 2025” with current provider specifications, so it should not be treated as a ByteDance official model card.

• The page claims a 32K maximum output, while the BytePlus official API documentation lists max_completion_tokens as supporting up to 64k; this is a clear provider/API boundary and the two figures should not be combined into one number.

• Per-question inputs, sampling parameters, number of repetitions, hardware, and tool-execution logs are not disclosed; the results are better treated as a selection observation than as an independent benchmark.

Reproduction steps

1. Use Puter Playground or the OpenAI SDK above to request the same baseline prompt, and save the model, usage, latency, and complete response.

2. Design one public input each for search, GUI, video, and code; enable different thinking tiers for each.

3. Repeat each task at least 5 times, and measure the success rate, tool-call validity rate, output tokens, latency, and failure type.

4. Rerun the same inputs and budget through the native Volcengine/BytePlus API, and report provider specifications and model capability separately.

Source excerpts or observations (for compliant short quote only)

The page describes Seed1.8 as oriented toward “search and agent workflows” while also providing an OpenAI-compatible API; this article retains that positioning but separates third-party and official specifications.