Anthropic's Claude Fable 5.1 targets long agentic work, lands on Vercel and Google platforms

Anthropic has released Claude Fable 5.1, with announcements landing on September 1 from two platforms now serving the model: Vercel's AI Gateway and Google's Agent Platform. According to Vercel's changelog, the gains over previous Claude models are concentrated in drawn-out, multi-step tasks — coding carried out through agents, knowledge work, and research that needs repeated rounds of searching and follow-up — rather than single-shot requests.

A Google AI post by Jen Harvey, published on dev.to, confirms the model is available in the Model Garden on Agent Platform, where it is positioned for extended, high-stakes work spanning code, documents and research.

Tuned for agentic coding and research

Both announcements frame Fable 5.1 around sustained, tool-using workflows. That is a deliberate positioning choice: rather than competing purely on conversation quality or benchmark scores, the model is being offered on its ability to hold together multi-step processes, such as writing and revising code through an agent or running iterative research loops where each result informs the next query. Vercel notes the improvements apply to knowledge work generally, not only programming.

Safety classifiers may refuse routine requests

Vercel's changelog highlights an operational quirk that developers should plan for: Anthropic ships Fable 5.1 with cybersecurity and biology safety classifiers enabled. Hunting for security flaws in code is allowed, but some ordinary coding and debugging requests may still be refused. For teams running the model inside an automated agent, an unexpected refusal partway through a task is a genuine failure mode, and Vercel recommends configuring model fallbacks so requests still get serviced when a classifier triggers.

Fallback chains on AI Gateway

The mechanism works through a models array added to providerOptions.gateway, which lists candidate models in priority order. AI Gateway sends the request to Fable 5.1 first; if Anthropic declines it, the gateway works down the array and returns the response from the first model that succeeds. In Vercel's example, the chain falls back to Opus 5 and then Sonnet 5. The same option is supported across every AI Gateway API format, including Chat Completions, Messages and OpenAI Responses.

The model can also be tried in Vercel's model playground. For coding agents, Vercel provides a setup command, vercel ai-gateway coding-agents setup, which connects tools such as Claude Code, Codex, OpenCode, Cursor and Pi; the agent then selects anthropic/claude-fable-5.1.

Data retention terms

One limitation worth flagging: according to Vercel, Anthropic does not support Zero Data Retention for Fable 5.1. Prompts and completions are kept for 30 days, though they are not used to train Claude. Organisations with strict data-handling rules will need to weigh that before routing production traffic to the model.

Why it matters

Fable 5.1 appearing on Vercel AI Gateway and Google's Agent Platform on the same day illustrates how distribution has become part of a model launch. Reaching developers both through an aggregation layer like AI Gateway and an enterprise platform like Google's Model Garden lets Anthropic land the model inside existing workflows rather than waiting for teams to migrate.

The safety classifier behaviour is the more consequential detail. Permitting vulnerability research while occasionally refusing ordinary debugging inverts what many developers would expect, and it turns fallback configuration from an optional resilience feature into a production requirement. Combined with the 30-day retention policy, the launch is a reminder that adopting a frontier model now involves routing, compliance and failure-handling design, not just swapping an API key.