Leaderboard
Model performance ranking
Which models are most used, fastest and most reliable on XPNCORE, from anonymized real traffic.See models & prices
Lowest p95 latency first, then time to the first token.
Models measured
19
Total requests
27.8K
In the last All time
Total tokens
3.1B
Input + output, last All time
- 1Healthy
DeepSeek V4.1 Flash
deepseek-v4.1-flash
- Tokens used
- 867K0%
- Requests
- 32 · 0.1%
- Success rate
- 96.88%
- p95
- 10,946 ms
- Throughput
- 94 tok/s
Methodology and limitations
4,320 hours of data, from Apr 19, 2026, 12:38 PM to Oct 16, 2026, 12:38 PM; updated Oct 16, 2026, 12:38 PM.
Share is each model's part of all requests (or tokens) across the models listed here, in the selected range.
Requests is the number of finished requests (successful, failed or with an unknown outcome) a model served in the selected range. Requests refused before they were sent are not counted.
Tokens used is the input plus output tokens a model's requests reported in the selected range.
Latency (p50, p95, p99) is measured on successful requests.
TTFT is the average time to the first token of successful requests.
Throughput is output tokens per second while the model is answering.
Success rate = successful requests / all finished requests. Failed, timed-out and cut-off requests all count as unsuccessful.
Status shows whether a model is currently healthy, degraded or unavailable.
Ranked by lowest p95 latency, then lowest TTFT.
Notes:
- Data runs up to the latest full hour.
- TTFT and throughput can cover fewer samples than success rate and latency.
- Status is the current state, not the state over the whole range.
These are observed figures, not a service-level commitment.
