Observation period — public launch soon. The record below is real and updated daily.

Significant Latency anomaly·drift·2026-09-11

GPT-5.6 Sol took 5.3× its baseline time to respond (1.7 s → 9.2 s).

Model: openai/gpt-5.6-sol
pinned openai/gpt-5.6-sol-20260709; alias resolved to openai/gpt-5.6-sol-20260709 at 2026-09-11 09:00:01 UTC (matches the pin)
Prompt: D7 (drift)
Change: 9,156 ms against a baseline of 1,734 ms (5.3×; z = 4.16, σ = 0.4 in log units, default (D34 n=1))
Baseline: 1 prior day with data; 1 matched the baseline value (2026-09-10)
Detected: 2026-09-11 · run 20260911T090002Z-da0dd3 (2026-09-11 09:00:02 UTC) · scorer v1
Severity: significant — rule: a length or latency anomaly is significant
Event ID: 2026-09-11:D7:openai/gpt-5.6-sol:anomaly:latency_ms
Permalink: https://modeldrift.watch/e/2026-09-11/D7/openai/gpt-5.6-sol/anomaly/latency_ms/

Prompt D7sent verbatim

Return only valid JSON matching {"city": string, "population": integer, "country": string} for Osaka. No markdown, no code fences, no explanation.

Why this prompt Scored by parse attempt on raw string.

Scored correct if the response returns bare JSON matching the requested schema.

Before baseline

2026-09-10

Scored: answered · format ok
Received: 2026-09-10 22:11:13 UTC
Run: 20260910T221001Z-bb5bd1
Served by: OpenAI · finish stop
Size: 55 characters · 50 tokens out incl. hidden reasoning · 1.7 s

{"city":"Osaka","population":2752412,"country":"Japan"}

End of response · 1 lines · 55 characters · sha256 7926685c90ce7ded3fdf8ca089edefa4fb752118b7ad2b773d7488a0d699115f

After this run

2026-09-11

Scored: answered · format ok
Received: 2026-09-11 09:01:43 UTC
Run: 20260911T090002Z-da0dd3
Served by: Azure · finish stop
Size: 55 characters · 59 tokens out incl. hidden reasoning · 9.2 s

{"city":"Osaka","population":2752412,"country":"Japan"}

End of response · 1 lines · 55 characters · sha256 7926685c90ce7ded3fdf8ca089edefa4fb752118b7ad2b773d7488a0d699115f

Reading this record

Response text is shown exactly as the API returned it, with Markdown left unrendered. Line numbers and highlights are added by this site; highlights come from the same patterns the scorer uses, so they show what the verdict rests on. Each model is asked once per day (k = 1): a single change can be sampling noise, which is why every event links the full responses rather than a summary. How scoring works.