Claude Fable 5.2 Launches on AI Gateway: Major Upgrades in Coding, Reasoning, and Vision

Anthropic's Claude Fable 5.2 is now available for invocation through AI Gateway. Designed for open-ended technical tasks, this model shows clear improvements in coding, reasoning, and vision capabilities over the previous Fable generation, and can handle ambiguous, complex long-running tasks unattended for extended periods.

Model Positioning: Both Executor and Orchestrator

Fable 5.2 is more flexible in usage than a single conversational model. It can directly run long-running tasks with unclear objectives, and can also play two roles within a system: as an orchestrator that dispatches simple subtasks to smaller models, and as an advisor that other Claude models can consult when they encounter difficult problems. This bidirectional collaboration capability makes it better suited for embedding in multi-model workflows rather than being used in isolation.

At the interaction level, it is officially reported to express itself more clearly than earlier Fable models and to generate output faster. For scenarios requiring repeated coding iteration and debugging, improvements in response speed and expression quality directly affect developers' working rhythm.

Effort Tiers and API Defaults

Fable 5.2 offers a reasoning effort tier ranging from low to max, with the API defaulting to high. This design allows callers to adjust the model's thinking depth according to task difficulty—using lower tiers for simple requests to control costs, and switching to higher tiers for complex tasks.

When calling through AI SDK, you can specify the model and reasoning tier as follows:

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'],
    },
  },
});

The model name is anthropic/claude-fable-5.2, callable through AI SDK, Chat Completions, Responses, Anthropic Messages, and other APIs, and can also be integrated with coding agents connected to AI Gateway.

Safety Classifiers and Data Retention Requirements

Fable 5.2 has cybersecurity and biosecurity classifiers enabled by default. Paired with this is an explicit data retention requirement: Anthropic does not support Zero Data Retention (ZDR) for this model—prompts and completions are retained for 30 days but are not used to train Claude. If a team has ZDR enabled, AI Gateway will return an error directly rather than forwarding the request to Fable 5.2.

This has significant implications for teams with strict data compliance requirements; before integration, teams need to confirm their ZDR status and assess whether they accept the 30-day retention policy.

Preserving Requests with Model Fallback

When a safety classifier rejects a request, the model fallback mechanism can prevent the request from failing entirely. By adding a models array in providerOptions.gateway, AI Gateway will try each model in order. For example, after the safety classifier rejects a request, the request will first fall back to Opus 5.5, then to Sonnet 5.5. This option behaves consistently across all AI Gateway API formats.

The code example above demonstrates exactly this usage: the models array lists backup models in sequence, automatically switching when the primary model is rejected.

Zero Data Retention (ZDR) is a contractual and engineering commitment on the vendor side: request prompts and completions are discarded after being processed in memory, without being written to any persistent logs or storage media, leaving only metadata for audit trails. Enterprises typically enable ZDR when handling regulated data (such as medical records, legal documents, and personally identifiable information) to meet data minimization and storage limitation requirements under compliance frameworks like GDPR and HIPAA. When model vendors need to review inputs and outputs for safety monitoring or abuse detection, they often cannot accept ZDR, creating a direct conflict between feature availability and compliance constraints. AI Gateway chooses to return an error directly here rather than silently forwarding, in order to prevent callers from unknowingly sending data to backends that lack ZDR guarantees.

Safety classifiers are discriminative models that run independently of the primary model, responsible for detecting specific risk categories (here, cybersecurity and biosecurity) when requests come in or responses go out, blocking on detection. Since classifiers make probabilistic judgments, false positives are possible—legitimate penetration testing code and vulnerability analysis requests may be flagged. The model fallback mechanism is precisely the engineering remedy for such false positives: it does not change the classifier's decision but instead, after rejection, hands the same request to backup models that either don't have the same classifiers enabled or have different decision policies, thereby reducing overall request failure rates while maintaining security boundaries. Note that each model in the fallback chain has its own capability limits and pricing, and cascading calls amplify latency and cost, so backup models should be chosen with comparable capability and acceptable pricing.

Integration and Gateway Capabilities

To use it in tools like Claude Code, Codex, and Cursor, install the latest Vercel CLI and run setup:

npx vercel ai-gateway setup

AI Gateway itself provides a unified API for invoking models, tracking usage and costs, and configuring retries, failover, and performance optimization to support availability beyond that of a single provider. It also includes custom reporting, ZDR support, API key budgets, routing rules, and more. On pricing, AI Gateway matches provider prices with no markup and no platform fees for inference, including Bring Your Own Key (BYOK) requests.

AI Gateway is a unified proxy layer between applications and model providers, exposing a single API contract externally while shielding internal differences among providers in authentication, request formats, error codes, and rate-limiting policies. Its core value is not just forwarding: usage and cost tracking rely on the gateway's unified metering of each call; retries and failover rely on the gateway holding multiple upstream credentials and being able to switch at runtime; API key budgets and routing rules move access control up from the code layer to the infrastructure layer. BYOK (Bring Your Own Key) means enterprises continue using their own contracts and keys with providers, with the gateway only passing through without handling billing—which is why it can charge no extra fees for BYOK requests.

A Few Observations

From a product design perspective, Fable 5.2's orchestrator and advisor positioning reflects that model capability tiering is becoming part of engineering practice: large models handle the hard parts, small models take on decomposable subtasks, and fallback mechanisms provide a safety net for security blocks. For developers, the effort tier and model fallback configurations are two settings worth prioritizing in call strategy design during integration. One thing to watch is the data retention limitation—it may become the primary barrier to adoption for some enterprises.

One-line Summary

Claude Fable 5.2 is available through AI Gateway, supporting effort tiers, orchestrator and advisor dual roles, and model fallback, but not ZDR.

Paragraph Summary

Claude Fable 5.2 is now available through AI Gateway. Designed for open-ended technical tasks, the model shows clear improvements in coding, reasoning, and vision over the previous Fable generation, and can handle ambiguous, complex long-running tasks unattended for extended periods. It can serve as an orchestrator dispatching subtasks to smaller models, or as an advisor for other Claude models. The API offers effort reasoning tiers from low to max, defaulting to high. The model has cybersecurity and biosecurity classifiers enabled by default and does not support Zero Data Retention—prompts and completions are retained for 30 days but not used for training, and teams with ZDR enabled will receive an error directly. When classifiers reject requests, fallback can be configured via the models array, cascading to Opus 5.5 and Sonnet 5.5 in order.