Claude Fable 5 deep dive: Anthropic’s Mythos-class flagship

What it is

Claude Fable 5 is a Mythos-class model from Anthropic, built for autonomous knowledge work and coding. It offers a 1M-token context window, accepts text, image, and file inputs with text output, and natively supports reasoning, tool calling, and structured outputs. On AI Hippo’s unified leaderboard it sits in the frontier multimodal tier.

Spec sheet at a glance

By the numbers: a 1M-token context window; input modalities text, image, and file; output text only. It exposes 10 API parameters—reasoning and include_reasoning, tools and tool_choice, structured_outputs, response_format, verbosity, max_tokens/max_completion_tokens, and stop—so it drops into agent and function-calling stacks without glue code. Knowledge cutoff is not disclosed in this snapshot, and no batch or cached-input tier is exposed on the verified hosts. Vendor: Anthropic.

Pricing and where to run it

Snapshot pricing is $10 input and $50 output per 1M tokens, verified consistently across the OpenRouter primary listing, anthropic-direct, and AWS Bedrock. Because output costs 5× input, long generations and agent loops inflate cost sharply—estimate with your real input/output ratio, not the input price alone.

How it compares on price

Fable 5’s blended cost is about $30 per 1M tokens (the $10/$50 mean), making it the priciest frontier option in the current snapshot. For context: GPT-5.4 blends to ~$8.75 (1.1M ctx), Google Gemini 3.1 Pro Preview ~$7 (1.0M), xAI Grok 4.20 ~$1.88 (2.0M), and Meta Llama 4 Scout just $0.20 (10M). That is also why Fable 5 does not top AI Hippo’s composite board: that score is a context-per-dollar proxy—not a capability benchmark—so cheap long-context models like Llama 4 Scout lead it. Judge Fable 5 on capability fit, not on that proxy rank.

Cost in practice: two scenarios

Model the real bill with your input/output split. An agentic coding turn of 100K input + 20K output costs about $2.00 on Fable 5 ($1.00 input + $1.00 output) versus roughly $0.55 on GPT-5.4—about 3.6× more. A long-document pass of 500K input + 5K output runs about $5.25 on Fable 5 versus about $1.33 on GPT-5.4. The takeaway: Fable 5’s premium bites hardest on output-heavy, high-frequency loops; reserve it for tasks where answer quality clearly outweighs token cost.

When to choose it—and when not to

When the task needs long context (whole repos, long documents), multimodal inputs (screenshots, PDFs, images), and reliable tool use with structured outputs, Fable 5 is a solid pick. For simple high-volume tasks with short inputs and outputs, the $50 output price rarely pays off—favor cheaper models and decide with AI Hippo’s cost presets and comparison pages.

Sources

• Claude Fable 5 on AI Hippo
• OpenRouter models API
• Anthropic pricing