byesu

LLM API latency and success rate, measured weekly

Week of 2026-11-02 → 2026-11-08 (UTC). 33 text models measured on real production traffic through the byesu gateway. Median success rate across models: 99.41%. Fastest median time to first token: claude-opus-4-7 (1.14 s). Highest mean output speed: claude-haiku-5-5 (166.1 tokens/s).

Updated 2026-11-13 20:45 UTC · next update next Monday

ModelProviderSuccess rateFirst token p50First token p95Total p50Total p95Output speed
gpt-5.6-lunaOpenAI95.22%4.1 s49.0 s8.0 s62.0 s68.5 tok/s
gpt-5.6-solOpenAI86.80%3.7 s18.6 s7.0 s46.0 s64.5 tok/s
gpt-5.6-terraOpenAI97.39%4.3 s27.7 s9.0 s58.0 s52.4 tok/s
gpt-6-astraOpenAI96.80%4.3 s19.5 s11.0 s69.0 s52.8 tok/s
gpt-6-lunaOpenAI98.48%7.8 s46.1 s11.0 s69.8 s100.7 tok/s
gpt-6-solOpenAI98.70%5.4 s44.5 s8.0 s65.2 s90.3 tok/s
gpt-6.1-solOpenAI98.03%4.6 s21.9 s10.0 s50.0 s57.3 tok/s
claude-fable-5-1Anthropic99.60%1.7 s3.5 s3.0 s62.9 s108.7 tok/s
claude-haiku-4-5Anthropic99.53%2.0 s3.9 s3.0 s13.0 s115.0 tok/s
claude-haiku-5-5Anthropic99.70%1.7 s3.9 s4.0 s30.0 s166.1 tok/s
claude-opus-4-5Anthropic95.32%————25.8 tok/s
claude-opus-4-6Anthropic99.98%1.6 s3.4 s4.0 s47.0 s44.9 tok/s
claude-opus-4-7Anthropic99.98%1.1 s2.1 s3.0 s45.8 s35.6 tok/s
claude-opus-4-8Anthropic99.89%1.5 s3.1 s3.0 s43.2 s56.0 tok/s
claude-opus-5Anthropic99.41%1.7 s3.4 s4.0 s121.9 s56.8 tok/s
claude-opus-5-5Anthropic99.57%1.9 s4.5 s7.0 s94.0 s79.7 tok/s
claude-sonnet-4-6Anthropic99.77%1.4 s2.1 s6.0 s99.7 s35.7 tok/s
claude-sonnet-5Anthropic99.72%1.5 s3.2 s6.0 s66.0 s50.8 tok/s
claude-sonnet-5-5Anthropic99.26%1.9 s4.3 s5.0 s50.0 s86.2 tok/s
gemini-3-flashGoogle99.99%3.6 s12.2 s5.0 s11.0 s56.5 tok/s
gemini-3.1-pro-lowGoogle99.78%——28.0 s73.0 s65.3 tok/s
gemini-3.6-flashGoogle100.00%3.6 s5.4 s6.0 s13.9 s133.0 tok/s
gemini-3.6-flash-highGoogle99.95%——6.0 s15.0 s110.1 tok/s
gemini-3.8-flashGoogle99.93%7.0 s33.3 s7.0 s26.0 s35.2 tok/s
gemini-3.8-flash-highGoogle99.92%5.5 s19.2 s10.0 s30.0 s62.0 tok/s
gemini-4-argonGoogle96.22%12.1 s71.4 s19.0 s100.0 s75.7 tok/s
grok-4.5xAI99.03%2.5 s18.7 s16.0 s58.0 s68.0 tok/s
grok-4.6xAI96.92%——4.0 s10.0 s33.7 tok/s
grok-4.7xAI97.93%3.2 s22.9 s6.0 s78.6 s70.2 tok/s
kimi-k3Moonshot AI99.75%5.5 s27.2 s6.0 s44.0 s50.0 tok/s
glm-5.3Zhipu AI98.45%——12.0 s71.0 s72.6 tok/s
glm-5.3-flashZhipu AI95.22%5.9 s38.1 s13.0 s99.8 s43.3 tok/s
deepseek-v4.1-flashDeepSeek98.57%3.6 s8.0 s5.0 s43.0 s154.1 tok/s

Download: CSV (last 4 weeks) · JSON. Live hourly status: byesu.com/status. To measure any endpoint from your own machine, use our open-source script llm-latency.

Image models

ModelProviderSuccess rateTime per request p50p95
gpt-image-2OpenAI99.15%29.0 s122.0 s
gpt-image-2.5OpenAI98.44%23.0 s80.0 s
gemini-3.0-pro-image-ultraGoogle97.27%37.0 s115.8 s
gemini-3.1-flash-image-ultraGoogle98.02%14.0 s37.6 s
gemini-nano-banana-2.1-ultraGoogle97.02%——

Last 4 weeks

Weeks starting 2026-10-12 to 2026-11-02 (UTC). Each line runs from the oldest week to the latest; the range next to it is the lowest and highest weekly value. Success-rate lines share one scale (90% to 100%, lower if a week fell below).

ModelSuccess rateRangeFirst token p50Range
gpt-5.6-luna83.22% – 95.22%4.1 s – 7.6 s
gpt-5.6-sol86.80% – 96.64%2.4 s – 4.4 s
gpt-5.6-terra94.80% – 98.40%2.1 s – 5.1 s
gpt-6-astra94.76% – 96.80%2.4 s – 5.0 s
gpt-6-luna88.76% – 99.33%7.8 s – 14.6 s
gpt-6-sol88.88% – 98.70%3.9 s – 7.5 s
gpt-6.1-sol96.19% – 98.03%3.1 s – 5.6 s
claude-fable-5-199.26% – 99.83%1.7 s – 3.2 s
claude-haiku-4-597.45% – 99.53%2.0 s – 2.9 s
claude-haiku-5-599.70% – 99.84%1.7 s – 3.3 s
claude-opus-4-595.32% – 99.65%1.4 s – 1.4 s
claude-opus-4-699.34% – 99.98%1.5 s – 4.2 s
claude-opus-4-799.09% – 99.98%1.1 s – 1.7 s
claude-opus-4-899.57% – 99.92%1.5 s – 2.7 s
claude-opus-598.58% – 99.82%1.7 s – 3.0 s
claude-opus-5-599.10% – 99.57%1.9 s – 3.1 s
claude-sonnet-4-699.68% – 99.99%1.3 s – 2.4 s
claude-sonnet-599.70% – 99.82%1.5 s – 2.4 s
claude-sonnet-5-599.12% – 99.27%1.9 s – 3.0 s
gemini-3-flash99.66% – 99.99%1.5 s – 3.6 s
gemini-3.1-pro-low77.70% – 99.89%— – —
gemini-3.6-flash99.97% – 100.00%3.6 s – 3.6 s
gemini-3.6-flash-high87.07% – 99.95%— – —
gemini-3.8-flash92.51% – 99.93%5.3 s – 10.3 s
gemini-3.8-flash-high96.56% – 99.92%4.7 s – 6.0 s
gemini-4-argon87.49% – 96.96%10.6 s – 16.3 s
grok-4.598.80% – 99.94%2.5 s – 5.4 s
grok-4.696.92% – 98.76%— – —
grok-4.797.93% – 99.79%2.5 s – 5.0 s
kimi-k398.22% – 99.88%3.5 s – 6.7 s
glm-5.389.91% – 99.49%6.9 s – 6.9 s
glm-5.3-flash79.82% – 95.22%5.9 s – 11.2 s
deepseek-v4.1-flash94.44% – 98.57%3.6 s – 3.6 s

How these numbers are measured

  1. Source. Every request that goes through the byesu gateway updates per-model, per-hour counters. This page sums one full week (Monday 00:00 to Monday 00:00 UTC). It is real customer traffic, not synthetic probes.
  2. Success rate is what the client sees: one request counts once, even if the gateway retried it on another route. Failures caused by the caller (insufficient balance, invalid parameters, client cancellation) are not counted as failures.
  3. Time to first token is measured at the gateway from receiving the request to the first output byte from the model. For reasoning models it includes thinking time. It excludes the network between you and the gateway. It is measured on streaming requests only, while total latency covers all requests, so for models with many short non-streaming calls total latency can be lower than time to first token.
  4. Total latency is the time until the response finished. Output speed is output tokens divided by generation time, weighted across all requests.
  5. p50 and p95 are computed from a 1-in-10 sample of the week's successful requests in the request log, where total latency is recorded in whole seconds. A model with fewer than 100 sampled requests shows a dash. Success rate and output speed come from the hourly counters and cover every request. Prompts differ by model, so this is not a controlled head-to-head benchmark: a model used for long agentic sessions will show higher total latency than one used for short chats.
  6. Models with fewer than 500 requests in the week are left out. We do not publish request volumes.

About byesu

byesu is an OpenAI-compatible and Anthropic-native API gateway: one key for Claude, GPT, Gemini, Grok and other models, pay-as-you-go. Setup guides: docs.byesu.com.