Previous edition: p50 latency YoY: April 2025 vs April 2026
p50 latency year over year: October 2025 to October 2026
p50 latency year over year: October 2025 to October 2026
Same provider tag, ≥50k requests in both months. Lower bars = faster requests.
Oct 2025Oct 2026Faster YoY
Oct 2025Oct 2026
YoY
Nebius fell 90% year over year. OpenAI rose as its traffic moved to larger models.`successful` flag semantics likely changed between 2025 and 2026; latency YoY is robust because it's quantiles over wall-clock duration and not affected by that.
Did AI providers get faster over the past year? Comparing October 2025 with October 2026, Nebius median latency fell 90%, from 2.8s to 288ms. OpenAI rose from 556ms to 2.3s as its traffic shifted to larger reasoning models.
Why it mattersYear over year latency changes mix two effects: provider infrastructure and the models customers run. A rise usually means heavier models, not a slower provider.
Period
Oct 2025 to Oct 2026
Updated
November 11, 2026
ID
latency-yoy-october-2026
§ 01
Key findings
- 01Nebius: 2.8s to 288ms (-90%).
- 02Vertex (Gemini): 2.8s to 816ms (-71%).
- 03Vertex (Claude): 6.3s to 2.9s (-54%).
- 04Anthropic: 4.2s to 4.0s (-5%).
- 05Bedrock: 3.9s to 4.3s (+10%).
- 06xAI: 4.2s to 5.0s (+19%).
- 07OpenAI: 556ms to 2.3s (+306%).
§ 01b
Questions this answers
- Have LLM providers gotten faster in the last year?
- Which provider improved latency the most?
- Why did some providers get slower?
§ 02
Data
| Provider | Oct 2025 p50(milliseconds) | Oct 2026 p50(milliseconds) | YoY delta(percent) |
|---|---|---|---|
| Nebius | 2.77 s | 288 ms | -89.61% |
| Vertex (Gemini) | 2.77 s | 816 ms | -70.56% |
| Vertex (Claude) | 6.28 s | 2.87 s | -54.35% |
| Anthropic | 4.21 s | 4.02 s | -4.68% |
| Bedrock | 3.94 s | 4.32 s | 9.86% |
| xAI | 4.19 s | 4.97 s | 18.53% |
| OpenAI | 556 ms | 2.25 s | 305.81% |
§ 03
Cite as
APA
Click to copy
BibTeX
Click to copy
ID: latency-yoy-october-2026·Updated November 11, 2026·Period Oct 2025 to Oct 2026