How fast each agent's model responds — tail latency, not just averages.

p95 latency by agent

95th-percentile response time · lower is better

Codex CLI · gpt-5-codex

6.8k ms

Claude Code · claude-opus-4-8

4.2k ms

support-copilot · gpt-4o

3.4k ms

Cursor · claude-sonnet-4-6

2.6k ms

data-pipeline-agent · claude-haiku-4-5

1.1k ms

Latency by agent

percentiles in ms · tput = output tokens / sec

AgentModelCallsp50p95p99AvgOut tok/s

Codex CLIgpt-5-codex5402.4k6.8k11.2k3.2k31

Claude Codeclaude-opus-4-81.2k1.8k4.2k7.8k2.1k42

support-copilotgpt-4o3.1k1.1k3.4k6.1k1.9k58

Cursorclaude-sonnet-4-69809002.6k4.8k1.4k71

data-pipeline-agentclaude-haiku-4-55.4k4001.1k2.2k620120

Slowest responses

individual outlier calls · the tail you actually feel

WhenAgentModelLatencyTokensSession

2026-06-02 19:28:00Codex CLIgpt-5-codex11.2k ms28.4kee71aa90

2026-06-03 13:01:00Claude Codeclaude-opus-4-87.8k ms19.8k1a1f867b

2026-06-03 13:50:00support-copilotgpt-4o6.1k ms14.2kab9920fd

Read percentiles, not averages. A healthy average hides the tail — p95 and p99 expose the slow calls users actually wait on. Codex CLI's 6.8s p95 and 11.2s p99 are the cost of a higher-latency model under load.

This is call latency — not task success. Latency measures how fast a model responds, not whether the agent accomplished the goal. Reliability — success rate, goal completion, and output quality — is a separate dimension we do not yet measure.

Success rate · projectedGoal completion · projectedOutput quality · projectedgap — not yet built

Fields: performance.agents (name · model · calls · p50 · p95 · p99 · avg_ms · tput) · performance.slowest (ts · agent · model · duration_ms · tokens · session) — from app/data/c16-fleet.ts (qcontrol /api/performance shape).