Claude Fable 5.1: the 75% cache-read cut, from a developer's seat

What it is

Anthropic's latest flagship pair. Claude Fable 5.1 and Claude Mythos 5.1 are the same model, but with different levels of safeguards. Fable 5.1 is generally available, while Mythos 5.1 is available only through our trusted access programs. The developer-relevant change this week isn't the benchmark chart — it's the pricing on cached input tokens, which reshapes any workload that replays the same system prompt, tool schema, or transcript prefix on every turn.

The next-work-session test

Concrete scenario: you maintain a coding agent with a ~40k-token system prompt (tool definitions, repo map, style rules) and a harness that runs it against a 200-case eval suite, five turns per case. On Fable 5, cache reads billed at $1.00/MTok; on 5.1 they bill at $0.25/MTok. Same code, no prompt changes — the eval run gets meaningfully cheaper the moment you point the SDK at claude-fable-5-1. Anthropic frames the impact as reducing the cost of typical workloads by an estimated 25% and highly agentic workloads by up to approximately 45%, which is a vendor claim, not a measurement — but the underlying per-token math is real and verifiable on your own bill.

Other things you'll notice inside the loop: Fable 5.1's default behavior is to write fewer user-facing updates during long tool-calling turns than Claude Fable 5 does. This becomes more pronounced at higher effort and in longer tool chains. If your UI streams "thinking" text between tool calls, expect quieter turns; there's a new beta display: "updates" knob to opt back in.

Pricing

Verified against Anthropic's own product page and secondary reporting:

• Input: $10 / MTok. Output: $50 / MTok. Both unchanged from Fable 5. Claude Fable 5.1 is priced at $10 per million input tokens and $50 per million output tokens (VentureBeat's breakdown matches).
• Cache reads: Cache reads now cost $0.25 per million tokens, 75% less than Fable 5.
• VentureBeat notes the multiplier oddity: Anthropic has cut a Fable 5.1 cache hit to just $0.25 on input, down from $1.00 for Fable 5. That's also just 2.5% of Fable's normal input-token price of $10, rather than the 10% multiplier used by most other Claude models.

Cache write pricing and 5-minute vs 1-hour TTL surcharges are not covered in the launch post — check the current pricing page before you re-architect a harness around long-lived caches.

What we'd actually use it for

Three narrow, honest wins:

Limits

• Output tokens still cost $50/MTok. If your agent is verbose (long plans, long diffs), the cache-read cut barely moves the needle — output dominates.
• Mythos 5.1 is gated. If your product needs the cybersecurity/bio safeguards profile, you're in a trusted-access queue, not the public API.
• The quieter-by-default streaming behavior will break UIs that assume steady inter-tool chatter. Budget a small migration.
• Vendor's "45% cheaper on agentic workloads" number is a modeled estimate, not your bill. Measure it.
• Cache hits require the prefix to be byte-identical. Any nondeterministic system-prompt template (timestamps, per-user IDs early in the prompt) silently kills the discount.

Try it if

• You run evals or agent loops with a large, stable prefix and were already using prompt caching on Claude.
• You have a long-horizon coding agent stuck on Fable 5 and want the price drop without a model swap risk.
• You're building against the SDK and can A/B claude-fable-5 vs claude-fable-5-1 on the same harness this week.

Skip it if

• Your workloads are short, output-heavy, or don't reuse prompt prefixes — the cache-read cut won't reach you.
• You need Mythos-tier safeguards and aren't in the trusted-access program.
• You're happy on Opus 5 and don't want to re-run your eval suite; Anthropic itself suggests For most workloads, start with Claude Opus 5 and reserve Fable 5.1 for the demanding end.

Deep links worth bookmarking: the Fable 5.1 platform docs, the what's-new page, and the prompting guide — read the last one before you ship, because the streaming-behavior change will bite otherwise.