# higress vs litellm-python

Operated by the Busbar team; Busbar is also an entrant. Gateway overhead against a local mock is not end-to-end model latency.  [Methodology and limitations](https://onthebench.ai/gateways/method) .

Requested cell: openai → openai. Metric: added_latency_p99_us.

Comparability: comparable. Ratios require an identical protocol pair, hardware, architecture, metric definition, clean harness commit, recorded mock digest, and metric-specific workload. Missing evidence never matches missing evidence. These are observed gateway measurements against a local mock, not a claim about model latency or statistical significance.

- Load-generator binary digests were not recorded; instrument equivalence is based on the identical clean harness source commit and mock digest.
- Measurements were made at different times on independently provisioned machines.

Ratio (B ÷ A): 10.097560975609756. No statistical significance is asserted.

[Comparison JSON](https://onthebench.ai/api/v1/compare?a=higress&b=litellm-python&ingress=openai&egress=openai&metric=added_latency_p99_us)  ·  [All Higress evidence](https://onthebench.ai/gateways/higress)  ·  [All LiteLLM · Python evidence](https://onthebench.ai/gateways/litellm-python)

## higress

Build higress-registry.cn-hangzhou.cr.aliyuncs.com/higress/all-in-one:2.2.5@sha256:9846006cb3cf2dafa942deab0e1d004ed30e219fbc9580438f9e3c58e7096422. Measured 2026-11-10T17:39:42Z.  [Run evidence](https://onthebench.ai/gateways/runs/run-higress-558a1502d4a7ea1f2b74c745739723084c953a3ae5dcf16760a2694a92c21270)

| Ingress → upstream | Metric | Value | State / reason | p99 bound (ms) |
| --- | --- | --- | --- | --- |
| openai → openai | added_latency_p50_us | 718 | valid; gateway_c1_samples=7996; direct_c1_samples=262593 | — |
| openai → openai | added_latency_p99_us | 943 | valid; gateway_c1_samples=7996; direct_c1_samples=262593 | — |
| openai → openai | gateway_c1_p99_us | 971 | valid; gateway_c1_samples=7996 | — |
| openai → openai | direct_c1_p99_us | 28 | valid; direct_c1_samples=262593 | — |
| openai → openai | cpu_us_per_request | 709.3196158019531 | valid | — |
| openai → openai | rps_per_cpu_second | 1409.801699716714 | valid | — |
| openai → openai | cost_window_conc | 8 | valid | — |
| openai → openai | cost_core_utilisation | 0.7330563381485045 | valid | — |
| openai → openai | cost_window_ok | 24883 | valid | — |
| openai → openai | cost_window_rps | 4146 | valid | — |
| openai → openai | cost_threads | 142 | valid | — |
| openai → openai | cost_nonvol_ctxt_per_request | 0.008399308765020295 | valid | — |
| openai → openai | cost_majflt | 0 | valid | — |
| openai → openai | added_ttft_p50_us | 881 | valid; ttft_gw_samples=100; ttft_direct_samples=100 | — |
| openai → openai | added_ttft_p99_us | 1103 | valid; ttft_gw_samples=100; ttft_direct_samples=100 | — |
| openai → openai | added_gap_p50_us | 2 | valid; gap_gw_samples=63; gap_direct_samples=63 | — |
| openai → openai | added_gap_p99_us | below_resolution | below_resolution; gap_gw_samples=63; gap_direct_samples=63 | — |
| openai → openai | streams_sustained_fps | 52911.767627714544 | valid | — |
| openai → openai | streams_sustained | 1688 | valid | — |
| openai → openai | idle_rss_mib | 1163.541015625 | valid | — |
| openai → openai | steady_state_rss_mib | 1170.94140625 | valid | — |
| openai → openai | recovered_rss_mib | 1171.9921875 | valid | — |
| openai → openai | peak_rss_mib | 1181.8125 | valid | — |
| openai → openai | peak_rss_hwm_mib | 1287.14453125 | valid | — |
| openai → openai | time_to_plateau_s | 66 | valid | — |
| openai → openai | growth_rate_mib_per_min | 0.8963186649029332 | valid | — |
| openai → openai | load_s | 300 | valid | — |
| openai → openai | shape | not_measured | not_measured | — |
| openai → openai | idle_shape | not_measured | not_measured | — |
| openai → openai | idle_static | 1 | valid | — |
| openai → openai | idle_growth_rate_mib_per_min | 0.6061112080903336 | valid | — |
| openai → openai | throughput_rps | 2688 req/s | concurrency 2; samples 16131 | 1 |
| openai → openai | throughput_rps | 5685 req/s | concurrency 8; samples 34120 | 5 |
| openai → openai | throughput_rps | 6095 req/s | concurrency 32; samples 36576 | 10 |
| openai → openai | throughput_rps | 6138 req/s | concurrency 64; samples 36832 | 50 |
| openai → openai | throughput_rps | 6138 req/s | concurrency 64; samples 36832 | 100 |
| openai → openai | throughput_rps | 6138 req/s | concurrency 64; samples 36832 | unbounded |

## litellm-python

Build ghcr.io/berriai/litellm:v1.108.3@sha256:e2bce769eb29b100d53ce90cbf2d6bba00dce82cb894c79d092ef6377523d7e5. Measured 2026-11-10T17:39:41Z.  [Run evidence](https://onthebench.ai/gateways/runs/run-litellm-python-10bb57507b26d310b59f12956034785b615cd5338cdaf5cee0e587c350c53acf)

| Ingress → upstream | Metric | Value | State / reason | p99 bound (ms) |
| --- | --- | --- | --- | --- |
| openai → openai | added_latency_p50_us | 8512 | valid; gateway_c1_samples=701; direct_c1_samples=267297 | — |
| openai → openai | added_latency_p99_us | 9522 | valid; gateway_c1_samples=701; direct_c1_samples=267297 | — |
| openai → openai | gateway_c1_p99_us | 9549 | valid; gateway_c1_samples=701 | — |
| openai → openai | direct_c1_p99_us | 27 | valid; direct_c1_samples=267297 | — |
| openai → openai | cpu_us_per_request | 8644.793152639088 | valid | — |
| openai → openai | rps_per_cpu_second | 115.67656765676568 | valid | — |
| openai → openai | cost_window_conc | 8 | valid | — |
| openai → openai | cost_core_utilisation | 0.24990558455671374 | valid | — |
| openai → openai | cost_window_ok | 701 | valid | — |
| openai → openai | cost_window_rps | 116 | valid | — |
| openai → openai | cost_threads | 32 | valid | — |
| openai → openai | cost_nonvol_ctxt_per_request | 0.028530670470756064 | valid | — |
| openai → openai | cost_majflt | 0 | valid | — |
| openai → openai | added_ttft_p50_us | 11154 | valid; ttft_gw_samples=100; ttft_direct_samples=100 | — |
| openai → openai | added_ttft_p99_us | 11974 | valid; ttft_gw_samples=100; ttft_direct_samples=100 | — |
| openai → openai | added_gap_p50_us | below_resolution | below_resolution; gap_gw_samples=63; gap_direct_samples=63 | — |
| openai → openai | added_gap_p99_us | 9 | valid; gap_gw_samples=63; gap_direct_samples=63 | — |
| openai → openai | streams_sustained_fps | 1207.9948436569616 | valid | — |
| openai → openai | streams_sustained | 62 | valid | — |
| openai → openai | idle_rss_mib | 592.95703125 | valid | — |
| openai → openai | steady_state_rss_mib | 628.60546875 | valid | — |
| openai → openai | recovered_rss_mib | 628.60546875 | valid | — |
| openai → openai | peak_rss_mib | 628.60546875 | valid | — |
| openai → openai | peak_rss_hwm_mib | 628.60546875 | valid | — |
| openai → openai | time_to_plateau_s | 61 | valid | — |
| openai → openai | growth_rate_mib_per_min | 0 | valid | — |
| openai → openai | load_s | 302 | valid | — |
| openai → openai | shape | not_measured | not_measured | — |
| openai → openai | idle_shape | not_measured | not_measured | — |
| openai → openai | idle_static | 1 | valid | — |
| openai → openai | idle_growth_rate_mib_per_min | 0 | valid | — |
| openai → openai | throughput_rps | bound_not_met req/s | concurrency not recorded; every cleanly-served rung that reported a tail had it at or above 1ms, so this gateway carried no measurable throughput under that bound | 1 |
| openai → openai | throughput_rps | bound_not_met req/s | concurrency not recorded; every cleanly-served rung that reported a tail had it at or above 5ms, so this gateway carried no measurable throughput under that bound | 5 |
| openai → openai | throughput_rps | 117 req/s | concurrency 1; samples 708 | 10 |
| openai → openai | throughput_rps | 119 req/s | concurrency 2; samples 718 | 50 |
| openai → openai | throughput_rps | 119 req/s | concurrency 2; samples 718 | 100 |
| openai → openai | throughput_rps | 119 req/s | concurrency 2; samples 718 | unbounded |
