Skip to content

Claude Fable 5.2 arrives on Vercel AI Gateway

Anthropic's new model is callable through Vercel's unified API, but only with data retention switched on and fallbacks configured for classifier declines.

Draftship Editorial3 min read

Claude Fable 5.2, an Anthropic model, is now callable through Vercel's AI Gateway, according to Vercel's changelog entry dated 9 November 2026. For builders, the Claude Fable 5.2 AI Gateway listing means one more model reachable through the same request formats Vercel already supports, addressed as anthropic/claude-fable-5.2.

What Vercel says the model does

Vercel describes Fable 5.2 as built for open-ended technical work, improving on earlier Fable models in coding, reasoning and vision. The changelog says it can run long, ambiguous tasks unattended. It can also act as an orchestrator that delegates simpler work to smaller models, or as an advisor that other Claude models consult on harder problems. Vercel adds that it communicates more clearly and produces output faster than earlier Fable models. An effort setting runs from low to max, with high as the API default.

You can reach it through the AI SDK, the Chat Completions, Responses and Anthropic Messages APIs, or a coding agent connected to the gateway. For Claude Code, Codex, Cursor and similar tools, Vercel points to its CLI: install the latest version and run npx vercel ai-gateway setup. There is also a model playground for trying prompts.

Fallback chains when a classifier declines

Two safety classifiers ship enabled: cybersecurity and biology. If the cybersecurity classifier declines a request, that request does not reach the model unless you have a fallback chain in place. Vercel's mechanism is a models array inside providerOptions.gateway, which the gateway works through in order. Vercel's sample sends anthropic/claude-opus-5.5 first, then anthropic/claude-sonnet-5.5. The changelog says the option behaves the same way across every AI Gateway API format.

The practical detail here is that a declined request is not always a malicious request. Classifiers can misfire, so an unscoped fallback chain is the difference between a slower, weaker answer and a visible error in production.

The retention condition

Access to Fable 5.2 requires data retention, and Anthropic does not support Zero Data Retention for this model. Prompts and completions are kept for 30 days and, per the changelog, are not used to train Claude. If a team already has ZDR turned on, AI Gateway returns an error rather than forwarding the request onward.

That is a harder constraint than a pricing change. Teams with contractual, regulatory or internal rules about prompt storage cannot simply switch this model on; they would have to change their retention posture for those calls or route the work to a different model.

Gateway mechanics and cost

Vercel says AI Gateway mirrors provider pricing with no markup and charges no platform fee on inference, including Bring Your Own Key requests. The gateway also tracks usage and cost, and supports retries, failover, custom reporting, per-key budgets and routing rules.

Limits and open questions

The changelog gives no price for Fable 5.2 itself, no context window, no latency figures and no benchmark numbers behind the claim of improvement over earlier Fable models. It does not say which regions serve the model, or whether availability exists outside AI Gateway. The fallback example only covers the cybersecurity classifier, so what happens when the biology classifier declines a request is unstated. Nor is it clear whether retention can be scoped per request rather than per project.

What it means for building apps with AI

For anyone shipping an app with model calls inside it, the pattern Vercel is pushing matters more than the model itself: name a primary model, declare an ordered fallback list, and treat refusals as a normal runtime case rather than an exception. If you are assembling an app with AI features and want to see how model choice and prompt design fit into the build, Draftship's guide to adding AI to your app covers that layer.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.