This long X post argues that Seed1.8 has significantly improved reasoning efficiency and multimodal capabilities, but information extraction can consume tokens excessively and tends to drift after more than ten turns; its coding is closer to a usable 0-to-1 prototyping tool than a top-tier engineering model.
Suitable tasks: Early trials requiring multimodal understanding, long-chain reasoning, low-cost search, or “vibe coding” prototypes.
Unsuitable tasks: Continuous planning over more than ten turns, complex system-level engineering, low-budget information extraction, and tasks with strict spatial-reasoning requirements.
Applicable model version: The post discusses ByteDance Doubao-Seed-1.8; the provider/snapshot and complete runtime environment were not disclosed.
Applicable client, agent, or API: Not disclosed; the content is an English retelling on X of a Zhihu article.
Recommended reasoning tier and parameters: The post mentions token-efficiency differences between medium and high, but does not disclose reusable sampling parameters; the tier and budget must be fixed independently.
Source format: An X post summarizing a Chinese Zhihu evaluation; the original link points to https://zhuanlan.zhihu.com/p/198510997881.
Tasks observed: Long-chain reasoning, 10K-text information extraction, coding/vibe coding, multi-turn conversations, and 2D/3D spatial understanding.
Configuration: The post did not disclose the prompts, system prompt, tool schema, number of repetitions, model snapshot, or scoring script.
Not disclosed. The post provides only task-level descriptions and token/turn observations, with no complete input or API configuration that can be copied directly.
Reasoning efficiency: The post says the medium version can achieve roughly the level of intelligence Seed1.6 reached with about 15K tokens using approximately 5K tokens, and mentions an entry cost of around ¥2; the billing region and specific experiment table were not disclosed.
Relative capability: The post says Gemini 3 Pro and GPT-5.2 can still achieve higher scores with about 60% of the tokens, suggesting that Seed1.8’s advantage is more in sustained search and detailed verification than in the strongest intelligence per token.
Information extraction: On source text of about 10K tokens, the post observes that CoT may restate or annotate the entire text, bringing token costs to roughly 2×; lowering the reasoning budget causes a marked drop in accuracy, while turning reasoning off makes the task nearly unusable; Gemini 3 Pro uses about 4K tokens for a similar task (all as relayed by the post).
Multi-turn consistency: The post considers it “basically usable” compared with Seed1.6, but says goal tracking and reasoning drift after about 10+ turns.
Coding and spatial understanding: Usable for 0-to-1 vibe coding, but still weak in system-level thinking; gains in 2D/3D spatial capabilities are limited.
The post provides a clear risk checklist for model selection: in budget-constrained extraction tasks, guard against CoT expansion; for multi-turn agents, set up stage-based summaries and resets; and for coding tasks, schedule human review. Multimodal long-chain tasks and prototyping can be prioritized for initial trials.
This is a personal/community retelling, not a controlled benchmark; all figures lack the original prompts and logs.
During this Tabbit visit, the Zhihu article displayed “a wasteland where no knowledge exists” and redirected to the homepage, so its body could not be verified; users must manually open the Zhihu page and take over access if they need to check the original charts.
The X post does not clearly specify Doubao-Seed-1.8’s endpoint, reasoning tier, pricing region, or the versions of the comparison models; “5K/15K” and “2×” cannot be treated as general performance guarantees.
“Vibe coding is usable” does not mean the result has passed security, regression, or system-design acceptance.
Use a fixed Seed1.8 endpoint and prepare a 10K-text extraction task, a 12–15-turn multi-turn conversation, a small coding task, and a spatial-image task.
Run each task separately with no/low/medium/high thinking, recording input/output/reasoning tokens, accuracy, turn count, and latency.
For the multi-turn task, save a goal summary every five turns and inject it again, comparing whether this reduces drift; add testing, static checks, and a security review to the coding task.
Save the results side by side with the target models under the same input budget, distinguishing data relayed from the original post from your own measured data.
The X post summarizes the risk as “after ~10+ turns, reasoning drifts”; this short quotation represents only a community observation, and the original Zhihu charts cannot currently be verified.
Doubao Seed 1.8
