01 / AI · Analysis
OpenAI sets an end date for new fine-tuning jobs
New training and existing model use have separate cut-offs, while Werk emphasised the facts a model needs to receive.

OpenAI announced that it would stop accepting new fine-tuning jobs from 6 Jan 2027, including from existing active customers.
DEV described fine-tuning as training a model on a user's own examples. The restriction narrows access to that way of adapting a model to specific cases.
Werk's official blog argued that instructions cannot replace facts a model cannot see. As training access contracts, that approach puts the information supplied to an agent at the centre of efforts to correct its mistakes.
Conditions for running existing custom models
OpenAI's deprecations document gives 7 May 2026 as the restriction date for organisations that had never run fine-tuning. That measure blocks new training jobs for those organisations, separately from the later cut-off for existing active customers.
The same document ties the end of access to existing fine-tuned models to the retirement of their underlying base models. Stopping new training jobs therefore does not mean that already trained models stop running immediately: creating a model and using it remain separate activities under the schedule.
Reliable rule-following remains a separate task
The τ-bench researchers identified reliable rule-following as an area needing improvement. Their findings also highlighted the need for methods that help agents act consistently.
Werk's blog addressed the information available to the model. Its argument distinguishes facts the model needs to receive from instructions that tell it how to behave. Supplying a missing fact and requiring consistent behaviour are distinct parts of correcting agent errors.
That distinction sets a basis for assessing Werk's approach: the argument concerns what information an agent receives, while its effect requires evidence of fewer errors. Werk has not published an actual error-reduction rate for the approach.