OpenRouter: Real-Time Latency, Throughput, and Provider Activity for Seed 2.0 Lite

One-sentence takeaway

The current OpenRouter page shows Lite at approximately 1.90 seconds P50 latency, 44 tok/s P50 throughput, and 99.98% three-day availability. This makes it a candidate for a low-latency default route, but these are real-time statistics from a single gateway, not a benchmark of the model's capabilities.

Use cases

• Suitable tasks: High-concurrency text/vision/tool-calling workloads, low-latency default routing for Agents, gateway health monitoring, and reasoning-token cost estimation.

• Unsuitable tasks: Treating OpenRouter's P50 or pricing as an Ark direct-connection SLA, or using it to prove Lite's answer quality.

• Applicable model version: OpenRouter slug bytedance-seed/seed-2.0-lite; the upstream model snapshot is not disclosed.

• Applicable clients, Agents, or APIs: OpenRouter's OpenAI-compatible API and Responses/Anthropic-compatible endpoints; the page currently shows one upstream Provider.

• Recommended reasoning tier and parameters: The page supports reasoning, max_tokens, temperature, top_p, tools, tool_choice, and response_format; fix and record them during evaluation.

Test environment

• Observer: OpenRouter's public model page, not a controlled laboratory benchmark.

• Provider: The page shows one Provider, with OpenRouter forwarding requests directly.

• Statistical window: The page currently shows three-day data from Aug 15 23:00 to Aug 18 23:00, and lists availability for the past 24 hours.

• Request distribution: The page does not disclose prompts, input lengths, concurrency, region, P95/P99, or sample count.

Input/configuration

• Model slug: bytedance-seed/seed-2.0-lite.

• Context/output: The page shows a 262,144-token context window and 131,072 maximum output.

• Weighted-average pricing: $0.2547 per million input tokens and $2.008 per million output tokens; list pricing is $0.25/$2.

• Invocation: The page provides an OpenAI-compatible quick start and recommends retaining reasoning details to continue multi-turn thinking.

Results data

• P50 latency: 1.90 seconds.

• P50 throughput: 44 tok/s.

• Uptime (past 3 days): 100.00%.

• Availability (past 3 days): 99.98%; the past 24 hours also show 99.98%.

• Page activity volume: 581M prompt tokens, 14M reasoning tokens, and 2.99M completion tokens; this reflects gateway traffic, not a quality score.

• Public application signals: The page lists applications such as Claude Code, Hermes Agent, and OpenClaw as generating traffic, but does not disclose their success rates or task outcomes.

Conclusion

Lite's real-time OpenRouter data supports positioning it as a low-latency, high-throughput, low-price candidate. It is suitable as a high-coverage default route, with high-accuracy or complex engineering tasks upgraded to Pro. Quality and tail latency still need to be retested with the target provider, real tasks, and the same harness.

Limitations

• A single provider/gateway and a dynamic three-day window cannot represent all regions, direct Ark connections, or future pricing.

• The statistical sampling method, input-length distribution, concurrency, and error classification are undisclosed; P50 does not represent tail latency.

• The 262K/131K figures are displayed on the OpenRouter page and may differ from the context and output limits of other providers.

Reproduction steps

1. Fix the model slug, input-token buckets, concurrency, stream setting, max_tokens, reasoning, and tool configuration.

2. Collect TTFT, total latency, throughput, errors, retries, reasoning tokens, and pricing separately for OpenRouter and direct Ark connections.

3. Calculate P50/P95/P99, availability, and cost per completed task, and correlate them with actual task success rates and human review.

4. Save the collection time window and provider snapshot each time to avoid treating dynamic page figures as permanent specifications.

Source excerpts or observations (for compliant short quotations only)

• The page calls Lite a default candidate for production workloads and displays latency, throughput, pricing, and context together.

• The page says that one upstream Provider forwards requests directly; therefore, gateway data is better treated as an access observation than as a cross-vendor comparison.