Claude Fable 5.2 API

Zurelay charges $3.20 per 1M input tokens and $16.00 per 1M output tokens, with cached input at $0.20. Anthropic's list price is $8.00 input and $40.00 output, so you save 60%. The rate is the same at every prompt length, up to the full 1M-token window.

Yes. Zurelay serves claude-fable-5-2 at 60% below Anthropic's list price for input and output tokens, paid from prepaid credit with no subscription. It is the same model, not a smaller substitute. You keep your code and change the base URL and key.

Create a Zurelay API key, set the base URL to https://api.zurelay.com/v1 and set the model to claude-fable-5-2. Chat completions, streaming, tool calls, images and PDFs use the request shapes you know from OpenAI. Temperature and top_p have no effect, so Zurelay leaves them out. Anthropic's API refuses a forced tool_choice on this model, so when you force one (required, or a named function), Zurelay sends it as auto and tells the model which tool to call.

Yes. Set the Anthropic SDK's base URL, or ANTHROPIC_BASE_URL for Claude Code, to https://api.zurelay.com with no /v1, because the client adds /v1/messages itself. Authenticate with your Zurelay key and choose claude-fable-5-2 as the model. Anthropic notes that Claude Code runs Fable 5.2 at medium effort by default.

Claude Fable 5.2 has a 1M-token context window, roughly 555k English words on Anthropic's current tokenizer, and returns up to 128K output tokens per request. Thinking counts toward the output limit, so set a large max_tokens at high effort and stream long responses.

Anthropic built it for demanding reasoning and long-horizon agentic work: agent sessions that run for hours, multistep deep research, and analysis carried through to a finished document, spreadsheet or deck. Anthropic points to coding, scientific research and finance. For routine traffic, Claude Opus 5.5 or Claude Sonnet 5.5 usually costs less per task.

The context window and output limit are the same, the list price per token is 20% lower, and Anthropic reports gains in coding, reasoning and vision with output up to 30% faster. Forced tool choice is still refused (Zurelay sends it as auto, with the tool named). Two changes can break code that passes thinking blocks back: editing earlier turns now invalidates them on every account, and only Fable 5.2 itself reads its thinking blocks. A prompt line asking for less thinking also no longer stops the model from thinking at medium effort and above.

Start at high, the default. Anthropic says most of the gain in capability is already present there. Step up to xhigh for your hardest long-running coding and agentic work where you've measured a gain, and keep max for the rare task where time and tokens don't matter. For easy turns, low or medium costs less and answers sooner.

Yes. Streaming works on chat completions and on the Messages API; on chat completions, set stream_options.include_usage to get token counts at the end of the stream. Hard tasks at high effort can run long, so stream those requests. Each key can have its own monthly budget and requests-per-minute cap, and failed requests are retried on another route and never billed.

Yes. Your requests run on Claude Fable 5.2 itself. Zurelay doesn't swap in a different or smaller model. Only smart routing, which you turn on yourself, can answer with a close alternative when the model is down, and the x-zurelay-model header names the model that answered. Output is sampled, so exact wording varies between calls on any API. Zurelay is an independent service and is not affiliated with or endorsed by Anthropic.