Claude Fable 5.5 surfaces in LiteLLM’s Azure model list

An October 13 commit by BerriAI to LiteLLM lists a model called azure_ai/claude-fable-5-5 at $10 per million input tokens and $50 per million output tokens. The entry also gives it a 1,000,000-token input window and up to 128,000 output tokens. Anthropic has not announced a Fable 5.5.
LiteLLM is the open-source gateway many teams use to call several model providers through one interface, and its JSON file of prices and capabilities is what its cost tracking reads. The new entry’s source field points to Microsoft’s Azure Cognitive Services model listing for the eastus2 region. That suggests the figures were taken from Azure’s own catalogue and not estimated. The entry also already carries a deprecation date, December 5, 2027, roughly 14 months after the commit.
Caching is where the price sheet gets more detailed. Writing to the cache costs $12.50 per million tokens, and writing to the longer one-hour cache costs $20. Reading from the cache costs $0.25 per million. The minimum cacheable prompt is 512 tokens. Web search is billed at $0.01 per query, the same at low, medium and high search-context sizes.
For comparison, Opus 5.5 launched in September at $4 input and $20 output with cache reads at $0.20. Anthropic described it as matching Fable 5.1 on most work. Going by this entry, Fable 5.5 charges two and a half times Opus 5.5’s input and output rates, but its cache reads cost only a nickel more per million. A cached token on Fable 5.5 would cost one fortieth of a fresh input token, while Opus 5.5’s cached tokens cost one twentieth of its input rate. Anthropic said in September that cache reads make up the majority of agentic and coding costs. So on long agent loops the real gap between the two models would likely be much smaller than the headline rates suggest, although output tokens would still cost 2.5 times as much.
The capability flags show a model that always reasons before it answers. thinking_always_on is true, adaptive thinking is supported, and the reasoning-effort settings go up to both max and xhigh. A few common controls are switched off. The entry marks sampling parameters such as temperature, assistant prefill and forced tool use as unsupported, while tool_choice in general is supported. If the entry holds, code that prefills the assistant’s reply or sets temperature would need to change before moving to this model. Other flags are marked supported: system messages in the middle of a conversation, computer use, vision, PDF input, function calling and native structured output.
TestingCatalog, which pointed to the commit, wrote that if the model dropped that day “it would be a new SOTA across many areas.” The commit lists prices and feature flags only. It contains no benchmark scores.
On Anthropic’s system-cards page, the Fable line runs from Fable 5 and Mythos 5 in June to Fable 5.1 and Mythos 5.1 in September. The same page shows Opus 5.5 and Sonnet 5.5 in September and Haiku 5.5 in October. Fable 5.1 is still the newest Fable it lists.
Up next
Cursor’s new plans bill Claude from the very first request
Sign up for Cursor today at $30 a month and Grok comes out of an included weekly pool, while every Claude request is billed on top at API list prices. On that plan, would you stick with Grok or keep paying extra for Claude?
Related stories
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

