Anthropic's Python SDK v2.7.0 adds claude-fable-5-2 and claude-mythos-5-2

Anthropic published anthropic-sdk-python v2.7.0 on November 9, and the headline feature is easy to miss between the webhook plumbing and the Bedrock fixes: two new model IDs, claude-fable-5-2 and claude-mythos-5-2, now appear in the API surface. That is the first time these names show up in a public release note, and it is the kind of detail that usually leaks through a changelog before a launch blog post ever lands.

The rest of the release reads like a client library growing into its enterprise shoes. There is parent organization filtering and grouping in the usage and cost reports, a new sandbox_egress_restricted error on Admin API write methods, and a user_profile.program_status_updated webhook event. None of these are glamorous. All of them point at the same thing: Anthropic is building out the administrative layer that large customers need before they can run Claude at scale inside a regulated organization.

Two new model IDs, no announcement

The changelog line is four words long: “add claude-fable-5-2 and claude-mythos-5-2.” No context, no capability notes, no deprecation of anything. Tessera has not confirmed what these models are, what they cost, or when they become callable outside the SDK’s type definitions. {/* TODO: verify whether claude-fable-5-2 and claude-mythos-5-2 are generally available, and what their pricing and context windows are — searched the release page and found only the changelog line */}

What the naming does tell us is that Anthropic is running at least two model lines in parallel under a versioned scheme. The 5-2 suffix matches the pattern of a dot release inside a major family, the way claude-3-5-sonnet did before it. Whether “fable” and “mythos” are capability tiers, modality variants, or something else entirely is an open question. The fact that both landed in the same SDK bump suggests they ship together, or at least that the API team wanted the client ready before either goes public.

For anyone building on the SDK, the practical effect is immediate: the type stubs now accept those strings. If you pin your dependency and call a model name that the SDK does not know, you get a validation error before the request ever leaves your process. So a model that exists in the API but not in your installed SDK is, for most intents, a model you cannot use. The release is the unlock, even if the announcement is not.

The admin API is the real story

The more consequential changes are the ones nobody will tweet about. Parent organization filtering and grouping in the usage and cost reports means a company with subsidiaries, or an agency managing several client orgs, can now pull a single consolidated view instead of stitching together per-org exports. That is a billing and governance feature, and it is the sort of thing that procurement teams ask about before they sign.

Then there is sandbox_egress_restricted. This is a new error class on Admin API write methods, and its name is doing a lot of work. Egress restriction means network traffic leaving a sandbox is being blocked or limited. Anthropic is now surfacing that condition as a typed error rather than letting it manifest as a generic failure. Read that as a signal about where the platform is heading: more sandboxed execution, tighter network boundaries, and an expectation that developers will handle the restricted case explicitly. If you are building agents that call out to external services, this error is the one you will learn to catch.

The user_profile.program_status_updated webhook event, accepted on organization webhook endpoints, rounds out the picture. Anthropic is giving organizations a way to react when a user’s program status changes, which implies there is a program with statuses worth tracking. That could be a beta access program, a usage tier, or something tied to the sandbox controls. The changelog does not say. {/* TODO: confirm what “program_status_updated” refers to — beta enrollment, usage tier, or sandbox access — no public documentation found */}

Bedrock and Vertex get the fixes they were owed

The bug-fix list is where the multi-cloud users live, and it is a long-overdue cleanup. The parse and count_tokens methods were missing from beta.messages.with_raw_response and with_streaming_response on both Bedrock and Vertex, which meant anyone using raw responses on those platforms had a hole in their beta surface. The SDK now also sends the X-Stainless-Async header on those paths and streams raw requests that name no stream class.

Two Bedrock fixes stand out. Signing requests that have no Connection header no longer raises a KeyError, which is the kind of bug that only bites in a specific proxy or gateway configuration and then bites hard. And the Bedrock Mantle clients now serialize list query parameters as a[]=1&a[]=2, matching what Anthropic’s own API does. That last one is a compatibility fix that prevents a class of silent mismatches between the SDK and the server.

The client fixes are quieter but matter for anyone running long-lived services. API errors can now be unpickled and copied, which is what you need if you are passing exceptions across process boundaries or caching them in a task queue. And ANTHROPIC_CUSTOM_HEADERS is now read on the Bedrock, Bedrock Mantle, and Vertex clients, not just the direct API client. If your organization routes traffic through a gateway that requires a custom header, that fix is the difference between working and not.

What it means for builders

The pattern across v2.7.0 is consistent. Anthropic is hardening the edges that enterprise and multi-cloud deployments touch: cost reporting, org hierarchies, sandbox boundaries, custom headers, and error serialization. The model additions are the flashy part, but the admin and infrastructure fixes are what make the platform usable inside a company that has more than one team and more than one cloud.

The open question is the two new models. A changelog line is not a launch, and Tessera could not find pricing, context windows, or availability details for claude-fable-5-2 or claude-mythos-5-2. Until Anthropic says more, the safest read is that the SDK is ready before the announcement is. That has happened before, and it usually means the announcement is close.

For now, the upgrade is worth taking if you use Bedrock or Vertex, if you route through a gateway, or if you manage more than one organization. The sandbox_egress_restricted error is the one to watch. It is a small line in a long changelog, and it tells you more about where the platform is going than either of the new model names does.

The model additions are the flashy part, but the admin and infrastructure fixes are what make the platform usable inside a company that has more than one team and more than one cloud.

The release is dated November 9, tagged v2.7.0, and shipped by the anthropic-sdks-releaser account, which is the automation Anthropic uses for its client libraries. That automation cadence is itself a signal: the SDKs move on their own schedule, decoupled from model launches, which is why a version bump can carry a model name before any marketing page does. Developers who watch the changelog will keep seeing the future a few days early.