1985 status

                api           net   first   answer       speed    1k tok   failed   cache  source
coreweave ok    responses    61ms   213ms   1066ms   223 tok/s    4693ms        -       -  bench
crusoe    ok    responses   143ms   373ms    812ms   342 tok/s    3294ms        -       -  bench
nebius    off
together  ok    responses   240ms   592ms    802ms   249 tok/s    4604ms        -       -  bench
fireworks ok    responses   156ms  1456ms   2676ms   124 tok/s    9513ms        -       -  bench
baseten   ok    responses   268ms  2032ms   2027ms   220 tok/s    6572ms        -       -  bench

net: network round trip to the provider plus its least server time. first: time to the first token of any kind, reasoning included. answer: time to the first answer token, after reasoning.
speed: output tokens per second, reasoning included. 1k tok: predicted time for a 1000 token reply. cache: share of a known prefix the provider served from its prompt cache.
glm-5.3 reasons for about 160 tokens before answering.
live: fitted to traffic of the last few minutes. bench: the last benchmark, 2026-10-21T21:12:32.947Z.