When an AI model returns after a restriction, “available again” leaves a practical question unanswered: available under which conditions? A familiar model name can sit behind a different safety gate, and different requests can receive help from different systems.
Follow the request through three layers: the model’s capability, the classifier’s decision and the account’s access policy. Fable 5’s return changed the deployed service, so a familiar model name alone could not describe the resulting workflow.
What changed in the June resolution
Anthropic’s redeployment account says export controls were lifted on June 30 and global Fable 5 access resumed July 1. It described a revised classifier, routing blocked requests to Opus 4.8, and wider government evaluation and information-sharing commitments. It also reported more false positives on benign coding work. These are the company’s descriptions of its controls and testing, not guarantees about every future interaction.
The timeline describes the resolution of the June incident. Temporary subscription allowances mentioned in the announcement are historical, so they should not be used as current purchase guidance.
Three layers in one interaction
Take a fictional programming question: a developer asks why a service crashes when it receives a certain input. The underlying model may be capable of analyzing that code. A safety classifier separately assesses whether the request crosses a sensitive boundary. The product then decides what response path is available.
Those layers can disagree in ways that are visible to the user. A routine debugging request might be flagged. A more suspicious request might reach a less capable answering system. A technically complete answer might still be unavailable under the account’s access policy.
This explains why model benchmarks alone do not describe the entire product experience. A benchmark may measure whether the model can solve a problem. A user needs to know whether the deployed service will accept that problem, which system will handle it, and how any restriction is communicated.
Routing preserves a path, but changes the task
Redirecting a request is different from ending the interaction. It can preserve some assistance, but it also changes the capability available to complete the work. The user should not have to guess whether a different system produced the answer.
Imagine the developer’s crash report is redirected. The second model may still identify a common input-validation error. It might struggle with a more complicated chain of dependencies. Whether that is acceptable depends on the task, not on a general assumption that any answer is equivalent.
For an organization evaluating the service, useful questions include whether routing is disclosed, whether the result remains adequate, and whether the workflow can record what happened. Those checks are especially valuable when the same integration handles both ordinary maintenance and security-sensitive work.
A successful test has a boundary
Anthropic reported that its new classifier blocked the particular technique in the Amazon report in more than 99 percent of tested cases. That is a scoped result. It is not a measured probability that any possible attack will fail, and it is not a promise that benign requests will pass.
A simple comparison helps. A door tested against one forced-entry method can perform well in that test without proving that every entrance to the building is secure. The result remains useful when the reader keeps the tested method, conditions, and outcome in view.
The same care applies to false positives. Counting refusals alone misses legitimate work that was interrupted. Counting completed tasks alone misses harmful work that should have been blocked. An evaluation needs both useful completion and the consequences of mistakes.
The return involved institutions as well as software
The company also described earlier government access to relevant future models, joint evaluation work, and faster sharing of serious misuse findings. These arrangements affect how powerful systems are assessed and released. They do not establish that government evaluators independently endorse every company claim.
The June executive order explicitly limits what its voluntary framework authorizes. Reading that limitation alongside the product changes avoids turning one negotiated resolution into a universal claim about mandatory licensing.
For a task you rely on, record whether the request was accepted, which system handled it and whether the result met the task’s requirements. Compare those observations with the conditions actually offered, rather than treating a restored menu entry as the whole resolution. The companion on the earlier shutdown explains why availability became a separate business dependency.
