How to Use Claude Fable 5.2: API Setup, Effort Levels and Cost

Claude Fable 5.2 is Anthropic's successor to Fable 5.1 and, at the moment of writing, the top model on Artificial Analysis Intelligence Index v5.0.1. On OrcaRouter it is anthropic/claude-fable-5.2 at $8.00 per million input tokens and $40.00 per million output, served through the same key as every other model; the Claude Fable 5.2 API page carries the live numbers. The setup is short. The parts that surprise people are thinking behaviour, the effort dial, and the cache economics — so this guide spends its length there.

What the model accepts, and what it gives back

A 1M-token context window, where the maximum is also the default rather than something you have to opt into. Up to 128K output tokens in a single response. Input is text, images and documents; output is text. Native tool use and structured JSON outputs are supported. Hard tasks can run for several minutes inside a single turn, which is a feature for multi-step agents and a trap for anything with a synchronous request timeout.

Thinking is always on. The model reasons adaptively on every request and you control the depth with an effort level from low through max. The raw chain of thought is never returned — you get a summary, or an empty thinking field. Plan your observability around summaries, not transcripts.

Endpoint one: the OpenAI-compatible path

Same three lines as any other model on the gateway: base URL https://api.orcarouter.ai/v1, your key from the environment, and the model id. If you already have a working client for another vendor, this is a one-line change and the rest of the guide is about tuning it.

Endpoint two: the Anthropic Messages API

Fable 5.2 also answers natively on /v1/messages, and here is the detail people get wrong: the base URL for the native path is https://api.orcarouter.ai with no /v1 suffix. The OpenAI-compatible path needs /v1 because that is the convention Chat Completions follows; the Messages path does not. Copying a base URL between the two samples is the most common cause of a 404 that looks like a routing outage.

Effort levels, in practice

The supported parameter list is include_reasoning, max_completion_tokens, max_tokens, reasoning, reasoning_effort, response_format, stop, structured_outputs, tools and verbosity. Two entries do the real work here: reasoning_effort sets how deep the model thinks, and verbosity sets how much it says when it answers. They are independent, and conflating them is why some teams conclude the model is slow.

A practical starting policy: low effort for classification, extraction and routing, where the answer is short and the input carries the difficulty; medium for code review and document analysis; high or max only for the tasks where you have measured that a wrong answer costs more than the latency. Because thinking is on every request, effort is the only dial you have over the time before the first token — set it per route, not globally.

What a request costs

Input is $8.00 per million and output $40.00 per million, with cache reads at $0.200 and cache writes at $10.00 per million. A 40,000-token document plus a 2,000-token answer bills $0.32 of input and $0.08 of output: forty cents a request, before any caching.

Caching is where the arithmetic gets interesting, and it is worth doing properly because the write is charged. A 40,000-token prefix costs $0.40 to write into the cache and $0.008 every time it is read afterwards. One uncached pass costs $0.32. So a cached prefix pays for itself on the second read and is dramatically cheaper from then on — but a corpus you only read once should not be cached at all.

For scale, Fable 5.1 is $10.00 in and $50.00 out, and Opus 4.7 is $5.00 in and $25.00 out. Fable 5.2 being 20 percent cheaper than its predecessor on both lines is the whole reason to migrate an existing prompt set rather than keep the older id.

Latency and failure rate

On OrcaRouter's live traffic over the last seven days: p50 time to first token 8.09 seconds, p95 10.00 seconds, output speed 101 tokens per second, over 35.6 million tokens served. The error rate was 10.8 percent of requests returning non-2xx across all upstream providers, which is worth knowing before you put it behind a synchronous UI with no fallback.

The 8.09 seconds is not the same measurement as the 267 seconds Artificial Analysis reports for time to first token. AA's harness lets the model think at maximum effort on nine hard benchmarks; the gateway figure is the median of real requests at whatever effort the caller set. If you run Fable 5.2 at max effort on a hard prompt, you will see something much closer to AA's number than to the p50.

Where to use it, and where not to

Use it where a wrong answer is expensive and the call can be asynchronous: multi-step agents, deep code work, long-horizon planning, anything where you would otherwise run three cheaper models and reconcile. At $40 per million output tokens with thinking on every request, the cost of a misfire is real, so the effort level is a budget control.

Do not use it as a default routing target for every prompt. The Claude Fable 5.2 vs Claude Opus 5.5 comparison is the one to read if you are deciding whether the top of the line is worth it for your workload, and Claude Fable 5.2 vs Claude Fable 5.1 covers the migration if you already run the previous generation.

If you are building a multi-model stack, the LLM models list has the current ids and prices, and the Gemini 4 Carbon API is the natural cheap tier alongside it for the 90 percent of calls that do not need this much model.

The checklist

Set the base URL according to the endpoint you picked — /v1 for Chat Completions, no suffix for Messages. Set effort per route instead of globally, and separate effort from verbosity in your mind. Cache prefixes you read more than twice and skip caching for one-shot documents. Budget 128K output and 1M input. Expect eight seconds to first token at moderate effort, minutes at max, and one in ten calls to fail. Then measure your own distribution: the model is the most capable one available, and whether that is worth $40 per million output is entirely a question about your prompts.