Anthropic announced Claude Fable 5 and Mythos 5 for hard knowledge work and coding, with pricing at $10 input and $50 output per million tokens.
What Fable 5 and Mythos 5 Are Built For
Anthropic’s Claude Fable 5 and Mythos 5 target hard knowledge work and coding—tasks where shallow autocomplete is not enough. That usually means multi-file changes, careful reasoning about edge cases, long context that must stay consistent, and writing that has to be accurate under review. Treat them as tools for work you would otherwise do slowly yourself: designing an approach, tracing a bug through layers of code, or turning a messy requirement into a clear implementation plan.
Use them when the cost of a wrong turn is high. For trivial renames or one-line fixes, a lighter model or local tooling is often enough. Reserve Fable 5 and Mythos 5 for problems where you need sustained attention: migrations, unfamiliar codebases, security-sensitive paths, and documents that must hold up to scrutiny.
How to Work With Them on Coding Tasks
Give structure before you ask for code. State the goal, constraints, and what “done” looks like. Point at the files or modules involved, name the runtime and test setup if you know them, and say what must not change. Models do better when the boundary of the problem is explicit than when they must guess from a vague request.
Prefer iterative loops over one giant prompt. Ask for a plan first, review it, then request changes in small slices. After each change, run tests or a build and feed failures back with the exact error text. When the model proposes an API or library call you do not recognize, verify it against your project rather than pasting it in. Keep ownership of architecture: the model can draft; you decide what ships.
- Start with a short problem statement and success criteria.
- Share only the relevant context—not the entire repo by default.
- Request diffs or file-by-file edits so reviews stay focused.
- Validate behavior with tests, types, or a minimal repro before merging.
Using Them for Hard Knowledge Work
Outside pure coding, the same models help with research synthesis, technical writing, and decision docs. Feed primary material when you have it: RFCs, error logs, design notes, or ticket threads. Ask for structured outputs—comparison tables, risk lists, open questions—rather than open-ended essays. That makes review faster and reduces the chance that confident prose hides a gap.
For long documents, work section by section. Approve the outline, then draft each part against sources you can check. When the topic is regulatory, financial, or safety-related, treat every claim as provisional until you confirm it. The model is a drafting partner; citation and sign-off stay with you.
Pricing, Cost Control, and Practical Tradeoffs
Pricing is listed at $10 per million input tokens and $50 per million output tokens. Output is the expensive side, so design prompts to reduce unnecessary generation. Ask for concise answers first; expand only when you need depth. Prefer short, targeted context over dumping large files that will never be used. Cache stable system instructions and shared project rules where your stack supports it so you do not resend the same preamble every turn.
Budget by task type. Exploratory brainstorming can stay brief. Implementation and long-form drafts consume more output and should be scoped tightly. Track rough token use on a few real sessions so you know whether a given workflow is sustainable for your team. When cost matters more than peak capability, route simple tasks elsewhere and keep Fable 5 and Mythos 5 for the hard path—where extra reasoning and coding strength justify the spend.