Anthropic files and skills: the beta namespace no longer implies beta-shaped data

Keeping a beta namespace is not enough reason to keep your old files and skills fixtures unchanged when upgrading to Anthropic Python SDK v1.2.0.

The v1.2.0 release notes, dated August 27, 2026, include this feature entry:

api: beta files/skills namespaces use GA shapes; drop dated beta header pins

That is a narrow but useful contract change. For these namespaces, the SDK’s beta label no longer implies beta-shaped data. The practical decision is whether existing request builders and response fixtures still match the upgraded contract, not whether the calling code still contains the word beta.

Review the contract, not just the call site

For a v1.2.0 upgrade, I would make three focused checks:

Surface What to compare Decision to record
Request builders for files and skills The payload your wrapper constructs versus the installed SDK’s accepted parameters and types Keep the builder if it matches; change only differences you can identify.
Response fixtures and parsing assertions Existing fixture fields and expected values versus the response types for the operations you use Retain fixtures that still represent the contract; revise mismatches rather than relabeling an old fixture as GA.
Explicit beta headers Headers supplied by your application or shared wrapper versus those supplied by the upgraded SDK Review each manually retained pin and its purpose before keeping or removing it.

These are recommended review steps, not a migration procedure supplied by the release notes. A useful upgrade review should name the SDK version, the affected operation, and the evidence for retaining or changing each assumption. An unchanged namespace alone is not that evidence.

The header check deserves separate attention: the release says the SDK drops dated beta header pins, but it does not tell us what happens if an application explicitly supplies an old pin. Do not turn this entry into a blanket instruction to remove every beta header, including headers used for unrelated features.

What this release note cannot settle

The page does not enumerate individual schema-field changes or name the removed dated header values. Its linked implementation diff has not been inspected here, and no runtime migration test is being reported. There is therefore no basis here for a field-by-field rewrite, a claim that every legacy fixture breaks, or an assertion that a successful request proves the intended files or skills behavior.

If a gateway sits between your client and the API, its compatibility with the resulting requests needs separate evidence. This SDK release note does not establish that compatibility.

The bounded takeaway: keep the namespace if it suits your integration, but justify keeping the surrounding data assumptions against the installed contract. For files and skills in v1.2.0, beta is not a promise to preserve the old shapes.

For an evaluation harness that uses files or skills, I would treat this as a comparison-boundary question too: can results from before and after the SDK upgrade still support the same conclusion about the model?

The v1.2.0 release entry establishes a change to the files/skills shapes and dated header pins, not a change in model quality. If an evaluation score moves during that migration, the score alone cannot identify the cause.

I would separate two checks:

  • Harness continuity: keep the model identifier, task inputs and scoring rules fixed while comparing the old integration with the upgraded one. Record SDK version, fixture revision and explicitly supplied headers. For each failed case, distinguish a request-construction or response-parsing failure from a completed task that received a low score.
  • Model comparison: only attribute differences to a model change after establishing that the compared runs use equivalent task conditions. If the migration requires different fixtures, document what changed and why those fixtures still represent the same task; do not silently replace them and call the resulting scores directly comparable.

I would also report integration failures alongside task scores rather than simply dropping failed cases. Otherwise, a higher score on a smaller set of completed tasks could be mistaken for an improvement.

That is a proposed evaluation safeguard, not a reported effect of this release. Matching types or completing a request is useful contract evidence, but neither by itself establishes that the evaluation presented the intended task to the model.