Claude Fable 5.2 is now available in Vercel's AI Gateway to orchestrate models at lower cost

What Fable 5.2 is and why it matters for anyone orchestrating models

Claude Fable 5.2, from Anthropic, is available in Vercel's AI Gateway, according to the product's official changelog. The central point isn't simply one more model on the list: it's the piece that Vercel describes as part of a multi-model workflow, designed to split tasks between AIs of different sizes and costs within the same application.

According to Vercel, Fable 5.2 was built for open-ended technical work and improves on previous versions of the Fable family in coding, reasoning, and vision. It can run long, ambiguous tasks without constant human supervision, and generates responses faster than the Fable models that came before it.

The effort layer ranges from low to max, with high as the default in the API, controlled by the reasoning parameter in the official example. For anyone already dealing with the trade-off between latency and response quality, this effort control works, in practice, as another cost dial: more effort tends to mean more tokens processed and more response time.

Orchestrator or advisor: two roles, two ways to spend less

The most interesting point of the announcement for readers who already orchestrate multiple models in production is the dual role Vercel assigns to Fable 5.2: it can act as an orchestrator, delegating simpler tasks to smaller models, or as an advisor, consulted by other Claude models when a problem is too difficult for them to solve on their own.

This opens up two different architecture patterns:

• Fable as conductor: the application calls Fable 5.2 to understand the problem, break it down into subtasks, and route each part to the cheapest model capable of solving it, reserving expensive capacity only for what truly needs it.
• Fable as advisor: smaller models (or other Claude models) stay on the application's critical path, and only escalate to Fable 5.2 when they get stuck, like a senior colleague called in on demand.

In both cases, the gain is the same: paying the price of the more expensive model only for the fraction of calls that actually need it. Since the AI Gateway, according to Vercel itself, reflects the provider's price with no markup and no platform fee on inference (including on BYOK requests), this routing becomes real savings on the bill, not just a theoretical architecture exercise.

The price of fallback: safety classifiers and data retention

Not everything about Fable 5.2 is a gain without a trade-off. Vercel documents that the model has active cybersecurity and biology safety classifiers, which means certain requests can be refused before they even reach response generation. For anyone building code analysis or offensive security tools, this is something to test early, not discover in production.

Vercel's recommended way out for this scenario is model fallback: an array of models inside providerOptions.gateway that the AI Gateway runs through in order when the cybersecurity classifier refuses the request. The changelog's official example shows exactly this use, with streamText from the AI SDK:

import { streamText } from 'ai';

const result = streamText({
  model: 'anthropic/claude-fable-5.2',
  reasoning: 'high',
  prompt: 'Review this login handler for security bugs.',
  providerOptions: {
    gateway: {
      models: ['anthropic/claude-opus-5.5', 'anthropic/claude-sonnet-5.5'],
    },
  },
});

In this example, if the classifier refuses to analyze the login handler, the AI Gateway tries claude-opus-5.5 and then claude-sonnet-5.5, in the declared order. It's the same fallback mechanism available in any AI Gateway API format, not an AI SDK exclusive.

The other point that deserves extra attention is data retention. Anthropic doesn't offer Zero Data Retention for Fable 5.2: prompts and completions are retained for 30 days, not used to train Claude models according to Vercel, and this is a condition of access, not an optional setting. Teams that already have ZDR enabled in the AI Gateway simply get an error instead of seeing the request go through to Fable 5.2.

In short: before putting Fable 5.2 into any flow that touches sensitive customer data, it's worth checking the retention requirement with whoever handles compliance at the company. It's not a footnote detail: it's a binary block on access to the model.

How to test Fable 5.2 in practice

To call the model, the name to use in any API is anthropic/claude-fable-5.2. Vercel lists four integration paths:

1. AI SDK (the example with streamText above)
2. Chat Completions API
3. Responses API
4. Anthropic Messages API

In addition, the model is also accessible via coding agents connected to the AI Gateway, such as Claude Code, Codex, and Cursor. To enable this, the changelog recommends installing the latest version of the Vercel CLI and running npx vercel ai-gateway setup.

Anyone who just wants to get a feel for the model's behavior before touching any code has the AI Gateway's model playground as a shortcut, also mentioned in the official announcement.

When Fable 5.2 doesn't solve the problem

For simple applications, with a single call and no orchestration between models, Fable 5.2 is probably overkill: the gain from using an expensive orchestrator to break down a task that's already simple doesn't make up for the cost or the extra routing latency. The model was designed for multi-step scenarios, not for a standalone request.

Teams with a contractual or regulatory requirement for Zero Data Retention are also left out for now: the lack of ZDR support is a hard restriction from Anthropic on Fable 5.2, not a setting that can be worked around on Vercel's side. Products that deal with content sensitive to security false positives, such as legitimate offensive security research code, need to design the fallback carefully, since the cybersecurity classifier can refuse legitimate requests.

It remains an open question, because the changelog doesn't provide this data, how much Fable 5.2 costs per token relative to Opus 5.5 or Sonnet 5.5, cited in the fallback example itself. That calculation, which the AI Gateway makes available in its own usage and cost dashboard, is what will actually decide whether the orchestrator architecture is worth it for each team's use case.