DeepSeek V4 Pro 0813 changed behavior between 2026-10-21 and 2026-10-29
Same 22 probes, same settings, one week apart: its identity answer changed (self → claims:anthropic, self → claims:openai).
BenchGecko sends the same 22 probes to the same flagship models every week and compares each run with the previous one.
Between 2026-10-21 and 2026-10-29, DeepSeek V4 Pro 0813 changed: its identity answer changed (self → claims:anthropic, self → claims:openai).
| Probe | Before | After |
|---|---|---|
| identity-en-1 | self | claims:anthropic |
| identity-zh-1 | self | claims:openai |
Labs and providers update models, system prompts and safety filters behind the same API name. Single flips happen even at temperature 0; this was flagged because the change crossed the index's fixed thresholds.
Full history: https://benchgecko.ai/gecko-tests/model-drift-index · JSON: https://benchgecko.ai/api/lab/model-drift
Published automatically when a BenchGecko test or price tracker crossed a fixed threshold. Every number and quote comes from the stored run, linked above. Spotted an error? hello@benchgecko.ai