OpenRouter: Seed2.1 Turbo Live Provider Performance and Calling Configuration Observations

One-sentence takeaway

Within the observation window for a single upstream Provider on OpenRouter, the Turbo page shows approximately 2.24 seconds of P50 latency, 48 tok/s P50 throughput, and 100% uptime over three days; these are gateway observations, not a fixed SLA or a model-capability benchmark.

Use cases

• Suitable tasks: Evaluating Turbo for low-latency/high-throughput routing, OpenAI-compatible integration, reasoning-token recording, and provider health monitoring.

• Unsuitable tasks: Extrapolating one gateway's P50 latency to latency in all regions or through direct Ark access, or treating uptime as a guarantee of business availability.

• Applicable model version: OpenRouter slug bytedance-seed/seed-2-1-turbo; the upstream model snapshot is not disclosed on the page.

• Applicable clients, Agents, or APIs: OpenRouter Chat Completions/Responses/Anthropic-compatible interfaces; the page states that there is currently only one upstream Provider.

• Recommended reasoning tier and parameters: Use the reasoning parameter as specified on the gateway page, and save the returned reasoning_details; set the specific effort and max_tokens values per task.

Test environment

• Observer: OpenRouter's public model page.

• Provider: The page shows one upstream Provider. OpenRouter forwards requests directly and does not select among multiple Providers.

• Statistical window: The page shows uptime over the past three days, with the current window running from Aug 15 22:00 to Aug 18 22:00; latency and throughput are page-level P50 observations.

• Request task: The page does not disclose the prompt set, request distribution, region, or concurrency settings used for latency statistics.

Input/configuration

• Model slug: bytedance-seed/seed-2-1-turbo.

• Context and output: The page shows a 262,144-token context window and a maximum 262,144-token output; this differs from the 256K input wording on the official Ark page, so verify against the actual endpoint response before use.

• Pricing: The page shows a weighted average of $0.50 per million input tokens and $2.50 per million output tokens; caching and discounts may make the actual payment lower than the list price.

• Reasoning: The page states that the reasoning parameter and reasoning_details are supported, and that complete reasoning details should be retained when continuing a conversation; this is not a guarantee of a native Volcengine Ark field.

Results

• P50 latency: 2.24 seconds.

• P50 throughput: 48 tok/s.

• Uptime (past three days): 100.00%.

• Availability (past three days): 99.58%; the page also shows 99.35% over the past 24 hours.

• Weighted average price: $0.50 per million input tokens, $2.50 per million output tokens.

• Traffic observations: The page reports 15.3M prompt tokens, 3.77M reasoning tokens, and 1.27M completion tokens; these are activity volumes on the OpenRouter page, not quality scores for a single task.

Conclusion

OpenRouter data supports including Turbo in low-latency, cost-sensitive Agent routing trials, and highlights the need to record reasoning tokens and gateway health during evaluation. It does not show that Turbo outperforms other models on coding or vision tasks, nor does it replace verification of latency, pricing, and data compliance for direct Ark access.

Limitations

• The statistics are provider/gateway observations; the page does not disclose the prompt, token-length distribution, concurrency, region, P95/P99, or complete sampling method.

• The health of a single Provider cannot represent the experience of Ark, other aggregators, or all users.

• Page values update continuously; this entry records only the window observed on 2026-08-18.

• OpenRouter's displayed 262,144 context/max output and the 256K input wording on the official Volcengine Ark page may reflect differences in units or interface layers and must not be combined without verification.

Reproduction steps

1. Using the same OpenRouter slug, fixed prompt-token buckets, concurrency, and region, record TTFT, total latency, throughput, reasoning tokens, and errors separately.

2. Compare direct Volcengine Ark access with at least one reference model, keeping the input, tools, stream, max_tokens, and timeout consistent.

3. Calculate P50/P95/P99, availability, failed retries, and per-task cost separately; do not use page-level P50 as a substitute for tail latency.

4. Correlate gateway statistics with task completion rates, independent tests, and human-review results, and regularly save the collection time window.

Source excerpts or observations (for compliance short quotes only)

• The page emphasizes that current requests are forwarded directly to one Provider, with no selection among multiple Providers.

• The page states that complete reasoning_details should be retained when continuing a conversation; this is a reusable calling-configuration note.