Model drift means that an AI model's behaviour and performance change noticeably over time, sometimes even within the same model name. In July 2026 this became concretely visible: High Learning Rate documented silent changes in OpenAI's Codex agents and Guardian review prompts, GetReadyForAgents reported a reduced context window for GPT-5.6 Sol Codex, and a Dev.to analysis described a quality drop in GPT-5.5 Codex. For anyone working with sensitive or high-trust information, this means you cannot rely on the model name alone.
You only notice drift through your own measurements: golden datasets per workflow, regression tests and logging which model version and configuration were active. Without those measurements a deterioration can work through into decisions, explanations and compliance unnoticed. It is therefore about making things visible, not trusting a single version number.
What happened in July 2026 with Codex and GPT around model drift?
The direct trigger is a series of observations around OpenAI's Codex line. In the post "Your Agent Changed Under the Model Name" High Learning Rate describes, on 13 July 2026, how OpenAI made several silent changes to its agent stack: rolling back a context-window increase after unexpectedly high costs, adjustments to reasoning experiments and restoring a regression in the behaviour of Guardian review prompts and tools.
The pattern: the behaviour changed without the model name changing. That makes it hard to trace without your own measurements.
Two other observations reinforce this picture. On 14 July 2026 GetReadyForAgents reported that users saw an effective context-window reduction for GPT-5.6 Sol Codex, from around 353,000 to around 258,000 tokens, without a prominent announcement. A day earlier, on 5 July 2026, an author on Dev.to published an analysis of possible reasoning-token clustering in GPT-5.5 Codex, with developer measurements suggesting a drop of roughly eight to ten percentage points on complex, multi-step coding tasks compared with GPT-5 Codex. At that point OpenAI had not yet given an official diagnosis.
Our editorial assessment: these three signals come from different sources and measure different things (agent behaviour, capacity, task quality). Together they show that drift in 2026 is not one phenomenon, but several layers that each need to be measured separately.
What is interpretation drift and why is stable accuracy not enough?
A performance graph that stays flat is no guarantee that the model keeps reasoning the same way. That is the core of a peer-reviewed study in Scientific Reports (Nature portfolio) of 10 March 2026, which introduces the concept of "interpretation drift". The authors show that models under label noise can substantially change their internal logic and explanation rules, while their F1 score and predictive performance remain largely stable.
For work in which explanation and accountability matter, that is important. Two models can produce the same outcome, but on the basis of different reasoning. For auditability this means you must look not only at accuracy, but also at the consistency of the type of decision and its justification. Our assessment: precisely in the case of disagreement between multiple models as a verification signal, this distinction between performance drift and interpretation drift is practically usable.
How do you detect model drift per workflow?
A worked-out framework comes from Divinci.ai, which on 26 May 2026 described a regression-testing approach for custom LLMs. They distinguish four kinds of drift: quality drift, coverage drift, judge drift and production drift, and use a baseline manifest plus replay of production traces to determine whether score differences come from the model, the evaluator or the production environment.
Translated into a practical checklist for high-trust workflows:
- Establish a golden dataset per workflow with representative, controlled examples and expected outcomes.
- Log which model version and configuration were active for each run, so that a change is traceable if the name stays the same.
- Separate the evaluator from the model under test to spot judge drift when the assessor itself shifts.
- Replay production traces to distinguish runtime drift from model changes.
- Set thresholds to block releases in the event of demonstrable quality degradation.
- Measure not only accuracy, but also explanation consistency to catch interpretation drift.
More background on measurement methods can be found in the topic hub on AI verification and in our overview of layered hallucination detection.
What does model drift mean for high-trust and privacy-sensitive processes?
For long legal files, compliance analyses or incident reconstructions, drift directly affects reliability. A smaller context window can mean that part of a long file falls out of view; a shifted reasoning profile can produce different errors than the week before. Our assessment: the risk lies not in the change itself, but in the fact that the change goes unnoticed.
A verification layer fits here. IamVera.ai is not a language model and not a chatbot, but a privacy-focused verification layer that can route a task through selected independent models and expose verification steps, corrections, disagreements and sources for inspection. It does not make outputs correct on its own and does not remove the need for review, but it can help make visible, per workflow, which model versions and configurations were running and how outcomes relate to baseline behaviour. The pre-processing and anonymisation via the Semantic Privacy Shield take place on EU infrastructure; the workflow is designed to send only anonymised content onward, and when a privacy check fails nothing is sent onward. The professional final judgement always remains with the user.
The common thread: treat model drift as a verifiable risk layer, not as a statistical effect that you take for granted.
Sources and references
Sources: The article relies on High Learning Rate, GetReadyForAgents and a Dev.to analysis of Codex changes, a Scientific Reports study on interpretation drift and a regression-testing framework from Divinci.ai.