What Happened
The llm-anthropic 0.31 release added support for Anthropic’s Fable 5.2 and Mythos 5.2. Fable 5.2 is available through the llm CLI as claude-fable-5.2. [1]
The accompanying test highlights a deployment risk: when asked to generate an SVG of a pelican riding a bicycle, Fable 5.2 produced its best result at the xhigh reasoning setting. At max, it consumed 128,000 thinking tokens without producing an image. The report describes similar behavior with Opus 5.5, but not Fable 5 or 5.1. [1]
More reasoning did not guarantee a usable deliverable. This is a reported example, not a benchmark establishing overall model reliability. It nevertheless illustrates why production systems need explicit completion criteria and bounded resource consumption.
Why It Matters to Businesses
A model response is not necessarily a completed business task. For artifact-generating applications, success means delivering something valid and usable—not merely consuming tokens or returning a successful API response.
- Cost: Reasoning that produces no usable artifact can increase cost per successful task. Exact charges depend on provider pricing and billing rules.
- Latency: Long-running attempts can exhaust application deadlines and delay downstream workflows.
- Release risk: New model support in a client library does not establish production readiness for an existing workload.
The reported differences across models and reasoning settings justify workload-specific regression tests rather than assuming that a newer model or higher setting is better. [1]
Kimbodo Engineering Perspective
We would treat model choice, reasoning effort and output validation as one deployment configuration. Evaluating them independently misses the interaction that matters: whether a particular configuration completes the task within its quality, latency and cost constraints.
A CLI integration is useful for experimentation. A production application additionally needs orchestration around deadlines, validation, retries and observability. The reported SVG failure makes that distinction concrete. [1]
Our default would be the lowest reasoning setting that meets a representative quality threshold. Higher effort should be an evaluated escalation path, not a universal default. A single visual example is insufficient to select that threshold.
How We Would Implement It
Define and test the completion contract
For an SVG workflow, require a nonempty artifact, successful XML parsing, an SVG root element and successful rendering in an isolated environment. Keep structural validity separate from semantic quality: a valid SVG can still depict the wrong subject.
Build a regression suite covering representative requests and known failures. Compare model-and-effort combinations using completion rate, task quality, latency and cost per accepted artifact. Pin evaluated configurations and canary changes before broader rollout.
Put bounded orchestration around inference
- Request path: Application API, policy-controlled inference worker, model provider, artifact validator, then artifact storage.
- Budget controls: Enforce application deadlines, concurrency limits and supported token or reasoning controls. Verify what the provider actually allows the application to cap.
- Recovery: On missing or invalid output, allow a bounded retry or an independently evaluated fallback configuration. Do not automatically repeat the same maximum-effort request.
- Long-running work: Use a queue-backed worker when generation exceeds interactive latency targets, with explicit job states and expiry.
- Telemetry: Record model identifier, client version, effort setting, reported usage, elapsed time, validation outcome and fallback activity.
Risks, Costs and Security
Fallbacks and retries can improve completion rates while multiplying spend. Enforce a total task budget across attempts, and monitor cost per accepted result rather than token consumption alone. A client timeout also should not be assumed to cancel provider-side computation or billing.
Treat generated SVG as untrusted content. Reject scripts and external references, apply a restrictive sanitization policy, and render it in an isolated environment rather than inserting raw output into an authenticated application page.
Keep provider credentials server-side, minimize sensitive content in logs, and review retention and data-handling terms before transmitting business data. These controls remain necessary regardless of which model performs best in testing.
Where Kimbodo Comes In
Kimbodo builds and operates this in production for businesses — see our AI Infrastructure & MLOps practice, or Estimate My Infrastructure.
