LIVE BENCH / GCP / MODEL RAMP

LLM inference serving

Ramp a release candidate without losing time-to-first-token.

GCP REFERENCEllm_inference_servinghuman-loaded / externally operable

LIVE PROJECTION / RELEASE / SPLIT / VERTEX AIBENCH READY / TICK 000scroll topology

healthy flowfailure propagationlive signal fieldfailed / unreachable

Availability—

TTFT—

Overflow0 /s

Achieved RPS—

Public list-price estimate£5,435

SLONOT TESTED

Live controlsHUMAN / OPEN

Scenario

Peak load180 req/sMonthly budget15,200 GBPNew model traffic20%

Human pins0 active

Root causeNOT TESTED

Model split80 / 20stable / release candidate

No active failure or fault evidence is present; no root cause identified.

Availability zones0 failed

Fault injection0 active

NO ACTIVE FAULTSSelect a node, or use a region-boundary control, to inject latency (ms) or packet loss (%).

Reference and tools

GCP referenceOfficial spec ↗

Vertex AI endpoint traffic splits map deployed models to percentages that must total 100; private endpoints do not support traffic splitting.

Tool surface SITE TOOLSREGISTERING

Registering 37 tools.

Model boundary

Simulation values are inspectable model assumptions. Pricing is a public list-price estimate, not a bill. One bench, two operators, no silent overwrite.

Snapshot / 2026-08-27